onlinelibrary.wiley.com
Open in
urlscan Pro
162.159.130.87
Public Scan
URL:
https://onlinelibrary.wiley.com/doi/10.1002/cjce.25525
Submission: On November 07 via manual from ES — Scanned from ES
Submission: On November 07 via manual from ES — Scanned from ES
Form analysis 8 forms found in the DOM
Name: thisJournalQuickSearch — GET /action/doSearch
<form action="/action/doSearch" name="thisJournalQuickSearch" method="get" title="Quick Search" role="search">
<div class="input-group option-0"><label for="searchField0" class="hiddenLabel">Search term</label><input type="search" aria-label="search text" role="combobox" aria-autocomplete="list" aria-controls="ui-id-1" aria-expanded="false" name="AllField"
id="searchField0" placeholder="Search" onfocus="this.value = this.value;" data-auto-complete-max-words="7" data-auto-complete-max-chars="32" data-contributors-conf="3" data-topics-conf="3" data-publication-titles-conf="3"
data-history-items-conf="3" value="" tabindex="9" class="autocomplete actualQSInput quickSearchFilter ui-autocomplete-input" autocomplete="off"><span role="status" aria-live="polite" class="ui-helper-hidden-accessible"></span><input
type="hidden" name="SeriesKey" value="1939019x">
<div class="search-options-wrapper quickSearchFilter">
<a href="https://onlinelibrary.wiley.com/search/advanced?publication=1939019x" title="" tabindex="12" class="advanced"> Advanced Search</a><a href="https://onlinelibrary.wiley.com/search/advanced?publication=1939019x#citation" title="" tabindex="12" class="citation"> Citation Search</a>
</div>
</div><button type="submit" title="Search" tabindex="11" aria-label="Submit Search" class="btn quick-search__button icon-search"></button>
</form>Name: defaultQuickSearch — GET https://onlinelibrary.wiley.com/action/doSearch
<form action="https://onlinelibrary.wiley.com/action/doSearch" name="defaultQuickSearch" method="get" title="Quick Search" role="search">
<div class="input-group option-1"><label for="searchField1" class="hiddenLabel">Search term</label><input type="search" aria-label="search text" role="combobox" aria-autocomplete="list" aria-controls="ui-id-1" aria-expanded="false" name="AllField"
id="searchField1" placeholder="Search" onfocus="this.value = this.value;" data-auto-complete-max-words="7" data-auto-complete-max-chars="32" data-contributors-conf="3" data-topics-conf="3" data-publication-titles-conf="3"
data-history-items-conf="3" value="" tabindex="9" class="autocomplete actualQSInput quickSearchFilter ui-autocomplete-input" autocomplete="off"><span role="status" aria-live="polite" class="ui-helper-hidden-accessible"></span>
<div class="search-options-wrapper quickSearchFilter">
<a href="https://onlinelibrary.wiley.com/search/advanced" title="" tabindex="12" class="advanced"> Advanced Search</a><a href="https://onlinelibrary.wiley.com/search/advanced#citation" title="" tabindex="12" class="citation"> Citation Search</a>
</div>
</div><button type="submit" title="Search" tabindex="11" aria-label="Submit Search" class="btn quick-search__button icon-search"></button>
</form>POST
<form method="post">
<fieldset>
<legend>Please review our <a href="https://onlinelibrary.wiley.com/termsAndConditions" target="_blank" class="terms-and-conditions__cycleElement">Terms and Conditions of Use</a> and check box below to share full-text version of article.</legend>
<div class="input-group"><label for="terms-and-conditions" class="checkbox--primary"><input id="terms-and-conditions" type="checkbox" value="yes" required="" name="terms-and-conditions"
data-ajax-link="/action/generateShareUrl?doi=10.1002/cjce.25525&shareType=P2P&format=PDF" data-shareable-link=""><span class="label-txt terms-and-conditions__cycleElement">I have read and accept the Wiley Online Library Terms and
Conditions of Use</span></label></div>
</fieldset>
<hr class="separator">
<div class="shareable"><label>Shareable Link</label>
<p>Use the link below to share a full-text version of this article with your friends and colleagues.
<a href="https://onlinelibrary.wiley.com/researchers/tools-resources/sharing" target="_blank" class="emphasis more-link__cycleElement">Learn more.</a></p>
<div class="shareable__box">
<div class="shareable__text">
<div class="shareable__field"><span id="shareable__text"></span><textarea tabindex="-1" class="shareable__text-area"></textarea></div>
</div><button type="submit" disabled="" class="btn shareable__btn shareable-btn__cycleElement">Copy URL</button>
</div>
<div class="error shareable__error hidden"></div>
</div>
</form>POST /action/doLogin?societyURLCode=
<form action="/action/doLogin?societyURLCode=" class="bordered" method="post"><input type="hidden" name="id" value="67065c09-4a88-49cd-934c-ac707951d35c">
<input type="hidden" name="popup" value="true">
<input type="hidden" name="loginUri" value="/doi/10.1002/cjce.25525">
<input type="hidden" name="remoteLoginUri" value="">
<input type="hidden" name="redirectUri" value="/doi/10.1002/cjce.25525">
<div class="input-group">
<div class="label">
<label for="username">Email or Customer ID</label>
</div>
<input id="username" class="login" type="text" name="login" value="" size="15" placeholder="Enter your email" autocorrect="off" spellcheck="false" autocapitalize="off" required="true">
<div class="actions">
</div>
</div>
<div class="input-group">
<div class="label">
<label for="password">Password</label>
</div>
<input id="password" class="password" type="password" name="password" value="" autocomplete="off" placeholder="Enter your password" autocorrect="off" spellcheck="false" autocapitalize="off" required="true">
<span class="password-eye-icon icon-eye hidden" role="button" tabindex="0" aria-label="Password visibility" aria-pressed="false"></span>
</div>
<div class="actions">
<a href="/action/requestResetPassword" class="link show-request-reset-password">
Forgot password?
</a>
</div>
<div class="loginExtraBeans-dropZone" data-pb-dropzone="loginExtraBeans">
</div>
<div class="align-end">
<span class="submit " disabled="disabled">
<input class="button btn submit primary no-margin-bottom accessSubmit" type="submit" name="submitButton" value="Log In" disabled="disabled">
</span>
</div>
</form>POST /action/changePassword
<form action="/action/changePassword" method="post">
<div class="message error"></div>
<input type="hidden" name="submit" value="submit">
<div class="input-group">
<div class="label">
<label for="a589574e-bb98-4c6e-8fed-67365ff05357-old">Old Password</label>
</div>
<input id="a589574e-bb98-4c6e-8fed-67365ff05357-old" class="old" type="password" name="old" value="" autocomplete="off">
<span class="password-eye-icon icon-eye hidden"></span>
</div>
<div class="input-group">
<div class="label">
<label for="a589574e-bb98-4c6e-8fed-67365ff05357-new">New Password</label>
</div>
<input id="a589574e-bb98-4c6e-8fed-67365ff05357-new" class="pass-hint new" type="password" name="new" value="" autocomplete="off">
<span class="password-eye-icon icon-eye hidden"></span>
<div class="password-strength-indicator" data-min="10" data-max="32" data-strength="4">
<span class="text too-short">Too Short</span>
<span class="text weak">Weak</span>
<span class="text medium">Medium</span>
<span class="text strong">Strong</span>
<span class="text very-strong">Very Strong</span>
<span class="text too-long">Too Long</span>
</div>
<div id="pswd_info" class="pass-strength-popup js__pswd_info" style="display: none;">
<h4 id="length"> Your password must have 10 characters or more: </h4>
<ul>
<li id="letter" class="invalid">
<span>a lower case character, </span>
</li>
<li id="capital" class="invalid">
<span>an upper case character, </span>
</li>
<li id="special" class="invalid">
<span>a special character </span>
</li>
<li id="number" class="invalid">
<span>or a digit</span>
</li>
</ul>
<span class="strength">Too Short</span>
</div>
</div>
<input class="button primary submit" type="submit" value="Submit" disabled="disabled">
</form>POST /action/registration
<form action="/action/registration" class="registration-form" method="post"><input type="hidden" name="redirectUri" value="/doi/10.1002/cjce.25525">
<div class="input-group">
<div class="label">
<label for="4e647394-f751-4441-baa4-df426bca4b6e.email">Email</label>
</div>
<input id="4e647394-f751-4441-baa4-df426bca4b6e.email" class="email" type="text" name="email" value="" size="15">
</div>
<div class="submit">
<input class="button submit primary" type="submit" value="Register" disabled="disabled">
</div>
</form>POST /action/requestResetPassword
<form action="/action/requestResetPassword" class="request-reset-password-form" method="post"><input type="hidden" name="requestResetPassword" value="true">
<div class="input-group">
<div class="input-group">
<div class="label">
<label for="email">Email</label>
</div>
<input id="email" class="email" type="text" name="email" value="" size="15" placeholder="Enter your email" autocorrect="off" spellcheck="false" autocapitalize="off">
</div>
</div>
<div class="password-recaptcha-ajax"></div>
<div class="message error"></div>
<div class="form-btn">
<input class="button btn primary submit" type="submit" name="submit" value="RESET PASSWORD" disabled="disabled">
</div>
</form>POST /action/requestUsername
<form action="/action/requestUsername" method="post"><input type="hidden" name="requestUsername" value="requestUsername">
<div class="input-group">
<div class="label">
<label for="ac834f24-aa07-4ad2-9d13-f77c843f21cb.email">Email</label>
</div>
<input id="ac834f24-aa07-4ad2-9d13-f77c843f21cb.email" class="email" type="text" name="email" value="" size="15">
</div>
<div class="username-recaptcha-ajax">
</div>
<input class="button primary submit" type="submit" name="submit" value="Submit" disabled="disabled">
<div class="center">
<a href="#" class="cancel">Close</a>
</div>
</form>Text Content
* Skip to Article Content * Skip to Article Information Search withinThis JournalAnywhere * Search term Advanced Search Citation Search * Search term Advanced Search Citation Search Login / Register * Individual login * Institutional login * REGISTER The Canadian Journal of Chemical Engineering Early View SPECIAL ISSUE ARTICLE Open Access ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING AT VARIOUS STAGES AND SCALES OF PROCESS SYSTEMS ENGINEERING Karthik Srinivasan, Karthik Srinivasan Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Anjana Puliyanda, Anjana Puliyanda Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Writing - original draft, Conceptualization, Investigation Search for more papers by this author Devavrat Thosar, Devavrat Thosar Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Abhijit Bhakte, Abhijit Bhakte Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai, India Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Kuldeep Singh, Kuldeep Singh Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Prince Addo, Prince Addo Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Rajagopalan Srinivasan, Rajagopalan Srinivasan * orcid.org/0000-0002-8790-4349 Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai, India American Express Lab for Data Analytics, Risk & Technology, Indian Institute of Technology Madras, Chennai, India Contribution: Conceptualization, Investigation, Writing - original draft, Writing - review & editing, Funding acquisition Search for more papers by this author Vinay Prasad, Corresponding Author Vinay Prasad * vprasad@ualberta.ca Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Correspondence Vinay Prasad, Department of Chemical and Materials Engineering, University of Alberta, Edmonton, AB, T6G 1H9, Canada. Email: vprasad@ualberta.ca Contribution: Conceptualization, Investigation, Writing - original draft, Writing - review & editing, Funding acquisition Search for more papers by this author Karthik Srinivasan, Karthik Srinivasan Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Anjana Puliyanda, Anjana Puliyanda Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Writing - original draft, Conceptualization, Investigation Search for more papers by this author Devavrat Thosar, Devavrat Thosar Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Abhijit Bhakte, Abhijit Bhakte Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai, India Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Kuldeep Singh, Kuldeep Singh Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Prince Addo, Prince Addo Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Contribution: Conceptualization, Investigation, Writing - original draft Search for more papers by this author Rajagopalan Srinivasan, Rajagopalan Srinivasan * orcid.org/0000-0002-8790-4349 Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai, India American Express Lab for Data Analytics, Risk & Technology, Indian Institute of Technology Madras, Chennai, India Contribution: Conceptualization, Investigation, Writing - original draft, Writing - review & editing, Funding acquisition Search for more papers by this author Vinay Prasad, Corresponding Author Vinay Prasad * vprasad@ualberta.ca Department of Chemical and Materials Engineering, University of Alberta, Edmonton, Alberta, Canada Correspondence Vinay Prasad, Department of Chemical and Materials Engineering, University of Alberta, Edmonton, AB, T6G 1H9, Canada. Email: vprasad@ualberta.ca Contribution: Conceptualization, Investigation, Writing - original draft, Writing - review & editing, Funding acquisition Search for more papers by this author First published: 06 November 2024 https://doi.org/10.1002/cjce.25525 About * FIGURES * REFERENCES * RELATED * INFORMATION * PDF Sections * Abstract * 1 INTRODUCTION * 2 REPRESENTATIONS * 3 HYBRID AI * 4 HUMAN AND AI * 5 GENERATIVE AI * 6 TO AI OR NOT TO AI? * AUTHOR CONTRIBUTIONS * Open Research * REFERENCES PDF Tools * Request permission * Export citation * Add to favorites * Track citation ShareShare Give access Share full text access Close modal Share full-text access Please review our Terms and Conditions of Use and check box below to share full-text version of article. I have read and accept the Wiley Online Library Terms and Conditions of Use -------------------------------------------------------------------------------- Shareable Link Use the link below to share a full-text version of this article with your friends and colleagues. Learn more. Copy URL Share a link Share on * Email * Facebook * x * LinkedIn * Reddit * Wechat ABSTRACT We review the utility and application of artificial intelligence (AI) and machine learning (ML) at various process scales in this work, from molecules and reactions to materials to processes, plants, and supply chains; furthermore, we highlight whether the application is at the design or operational stage of the process. In particular, we focus on the distinct representational frameworks employed at the various scales and the physics (equivariance, additivity, injectivity, connectivity, hierarchy, and heterogeneity) they capture. We also review AI techniques and frameworks important in process systems, including hybrid AI modelling, human-AI collaborations, and generative AI techniques. In hybrid AI models, we emphasize the importance of hyperparameter tuning, especially in the case of physics-informed regularization. We highlight the importance of studying human-AI interactions, especially in the context of automation, and distinguish the features of human-complements-AI systems from those of AI-complements-human systems. Of particular importance in the AI-complements-human framework are model explanations, including rule-based explanation, explanation-by-example, explanation-by-simplification, visualization, and feature relevance. Generative AI methods are becoming increasingly relevant in process systems engineering, especially in contexts that do not belong to ‘big data’, primarily due to the lack of high quality labelled data. We highlight the use of generative AI methods including generative adversarial networks, graph neural networks, and large language models/transformers along with non-traditional process data (images, audio, and text). 1 INTRODUCTION Artificial intelligence (AI) and machine learning (ML) have found increasing usage in various domains, including many varieties of industrial practice, in the recent past, and process systems are no exception. If, as with other reviews,[1] we adopt the definition of Rich[2] for AI: ‘Artificial Intelligence is the study of how to make computers do things at which, at the moment, people are better’. The popularity of AI lies not just in its current successes, but in the promise that it can get much better and surpass humans at practically any endeavour. In this work, we review the use and promise of AI and ML in relation to process systems engineering. We do note that reviews in the recent past have addressed questions around the state of the art in great detail, drawing particular attention to the work of Venkatsubramanian,[1] where Venkatasubramanian provides a detailed history of the evolution of AI in the context of process systems engineering, and gives a perspective on future directions. He makes the point that there is both a ‘technology push’ and a ‘market pull’ for the current interest in AI, with greater computing power pushing and advanced algorithms (including model predictive control and reinforcement learning [RL]) pulling. Thon et al.[3] do have a brief discussion on the applications at various process scales, but their focus is primarily on particle technology. Both of these reviews, when they describe AI techniques, organize the presentation around methods (such as supervised and unsupervised learning, and describe methods in each area in greater detail) and applications (such as modelling, fault detection, optimization, and control). Dutta and Upreti[4] provide an overview of the application of AI and ML to process control in particular, and their framework of presentation is also organized around techniques (expert systems, fuzzy logic, and ANN-based controllers) and applications in various process contexts. What distinguishes our review from other notable ones in the recent past[1, 3, 4] is our focus on the various scales and stages of process engineering. Specifically, as shown in Figure 1, we analyze the application of AI at scales ranging from molecules and reactions to materials, and then to processes, plants, and supply chains. The other aspect we highlight is related to the stage of deployment. We broadly consider two stages (design and operation), with design also encompassing aspects of modelling and discovery (such as materials discovery) and operation including modelling, monitoring, control, and optimization. Both product and process design are considered in this categorization. In the figure, the solid lines with arrows indicate the scales of current application of the areas of generative AI, humans and AI, and hybrid AI, while the dotted lines represent opportunities to extend the application of these areas. Representations, of course, are required at all scales. We identify the most common sources of data at the various scales and in the various techniques: operational data includes all experimental investigations at the laboratory scale and process data, while databases (of publicly available data on molecules and reaction templates, for instance) encode some form of prior knowledge in their data. We note that hybrid AI includes not just data but also knowledge, typically encoding the known physics in the form of differential equations. Also, we provide a comprehensive look at representations, which is arguably the most important aspect of AI and ML in process systems engineering: a majority of the tasks involve regression and classification tasks, for which a variety of algorithms are available and can be tested, but as we shall see, the choice of data representation has a significant impact on the performance. Finally, we cover aspects of AI such as hybrid and generative AI, where a somewhat unique contribution is our evaluation of human-AI interactions in automation and other process engineering contexts. What is somewhat underappreciated in current practice is the importance of the human in the loop with AI. For instance, while AI algorithms may be deployed most of the time, there may be times when they are offline, and humans need to exercise judgement and make decisions. In instances where AI is used to complement humans, it is very important that the results of the AI are explainable to aid humans. In this context, we also present some thoughts on when not to use AI, emphasizing that not all process applications necessarily need AI, even if it is possible to develop AI algorithms for those cases, and that a cost–benefit analysis of AI should guide its deployment. FIGURE 1 Open in figure viewerPowerPoint An overview of artificial intelligence (AI) across the scales (molecules to supply chains) and stages (design and operation) of process engineering. In the rest of this paper, Section 2 focuses on representations, Section 3 focuses on hybrid AI modelling, Section 4 focuses on human-AI interactions, and Section 5 describes generative AI methods. We finish with Section 6, which describes the conditions that are favourable (or unfavourable) for the deployment of AI. 2 REPRESENTATIONS The adoption of ML in domain-specific areas is primarily to develop a suitable model of a process (i) to achieve a computational advantage over the traditionally used first-principles models when they are available, or (ii) to act as a surrogate when first-principles based models cannot be identified. The scale of chemical engineering spans the spectrum from molecular-level engineering to global supply chain optimization. Incorporation of ML strategies at any level of scale requires an effective machine-readable representation of entities at the considered scale. The class of ML models to be utilized is also intricately connected with the representation chosen. In this section, we present a discussion on some representational schemes that are widely employed at different scales of engineering in accordance with the taxonomy presented in Figure 2A. Note that the choice of representation falls into what is conventionally regarded as data preparation but also impacts feature engineering. Figure 2A provides a bird's eye view of the different representations detailed in this section. The ‘physics’ at the top of the figure describes the latent properties of the scale that are to be captured by the representations. Molecular representations, especially topological descriptors, are expected to be equivariant to translation and rotation. In general, molecular representations capture the permutation-invariant nature of the ordering of atoms in molecules. In the case of reactions, capturing the additive nature of the reaction mixture is deemed to be an important characteristic of the representation. Material representations that are produced through injective mappings are favoured in design and discovery due to ease of invertability. At the process level, string or graph-based representations aim to capture the connectivity between unit operations of the process and the inherent mass and energy flows. At the larger plant scale, representations that incorporate the hierarchy and relationship between different processing streams become more informative. Supply chain representations are typically constructed to encode the heterogeneity in the source and destination nodes and transportation routes. It is to be noted here that the qualities ascribed to representations are not limited to the specific scale but can be features of representations at other scales as well. For example, connectivity is still an important property of molecular representations as is heterogeneity. The more distinct feature of each scale has been denoted in Figure 2A. Figure 2B provides a classification of the embeddings mentioned in Section 2 under different scales and physics, and Figure 3 lists applications of the representations across scales. FIGURE 2 Open in figure viewerPowerPoint An overview of representations across different scales. (A) Schematic of representations across scales. (B) Classification of representations based on physics at different scales. DEXPI, Data Exchange in Process Industry. IUPAC, International Union of Pure and Applied Chemistry. HAZOP, hazard and operability. SFILES, simplified flowsheet-input line-entry system. SMILES, Simplified molecular-input line-entry system. SMARTS, SMILES arbitrary target specification; SMIRKS, SMILES reaktion specification. FIGURE 3 Open in figure viewerPowerPoint Application of representations across scales. HAZOP, hazard and operability. QSAR, quantitative structure–activity relationship. QSPR, quantitative structure–property relationship. LLM, large language models. Enforcing domain physics in ML models mitigates the need to train on large datasets and, most importantly, enhances the generalizability, transferability, and interpretability of these models, which is crucial for their adoption in mission-critical domains as chemical engineering.[5] Domain physics can be encoded in ML either by regularizing the structure of the models by way of model architectures and loss functions, or by suitably representing digitized data, which is typically heterogeneous, multi-scale, unstructured, and otherwise inaccessible for training ML models[6] (also see Section 3). Suitable data representations are a set of features/descriptors enriched with metadata from which the system physics emerges, and are task and scale-specific. Metadata is an avenue to incorporate knowledge representations at varying levels of system abstraction, ranging from differential equations, theoretical simulations, logic rules, and neural symbolic expressions to knowledge graphs, text embeddings, and ontologies for semantic relationships.[7, 8] In the absence of prior knowledge, an operator-theoretic approach is used to obtain global data representations. For instance, the Koopman operator uses infinite, intrinsic linear coordinates to decouple the underlying nonlinear dynamics by spectral decomposition of time series sensor data, a finite approximation of which is achieved by dynamic mode decomposition.[9] However, partial knowledge about a process like the conservation of physical quantities is incorporated by affine neural operators for spatial and temporal invariances via convolution and time recurrence, respectively.[10] Section 2 is intended to act as a primer to familiarize the readers with the more widely used representations at different levels of chemical engineering. In this section, we also identify representation schemes that incorporate metadata and operational information to inject domain knowledge of the process. Representations of molecules aim to capture the intrinsic topology and properties of constituent atoms to enable computer-aided modelling and exploration of the chemical space. While we acknowledge that other representations of molecules (such as Molecular Design Limited (MDL)) Molfiles, Wiswesser line notation) exist, we focus this section on the more commonly used representations of molecules for ML applications.[11-13] String-based representations depict molecular structure as a combination of numerals, alphabets and certain special characters and follow a set of semantics typically unique to the type of representation. Simplified molecular-input line-entry system (SMILES)[14] is a widely used string-based notation that encodes molecular connectivity. Each atom is represented by a unique sets of alphabets (same as their symbols in the periodic table). Numbers and parentheses are employed to indicate rings and branches. Special characters such as (=, #, @, +, etc.) indicate the type of bond, stereochemistry, and charge. The SMILES representation of a molecule is non-unique and the enumerative expanse of equivalent SMILES structures leads to issues with its application. Attempts to develop canonical variations of SMILES strings have been noted in the literature. Several modifications to the base SMILES notation such as DeepSMILES[15] and SELFIES[16] have been proposed to overcome specific drawbacks. International Union of Pure and Applied Chemistry (IUPAC) nomenclature describes functional groups in a molecule in words rather than symbols. Similar to SMILES, the mapping between a molecule and its IUPAC name is non-unique (due to the use of retained names) though the reverse mapping is unique. The International Chemical Identifier (InChI) is a canonical, open source molecular representation developed by IUPAC that represents a molecule in a hierarchical text notation. The inherent sequential nature of string representations lends itself to be readily adapted into ML models incorporating recurrence and convolutional operations. Reaction and product distribution predictions using recurrent neural networks (RNNs) and transformers have been carried out on SMILES and InChI strings.[17-21] Neural machine translation tasks have traditionally used SMILES as input to translate to other molecular representations.[22, 23] IUPAC names and InChI identifiers have been used in machine translation as well[24-28] but with a lower degree of success due to large token space and rigorous grammatical rules. Generative algorithms such as variational autoencoders (VAE), generative adversarial networks (GANs), and transformers employ SMILES (and its variants) as the decoding from a latent vector space.[29-32] SMILES also form the input representation for many molecular property prediction techniques.[33, 34] The strength of InChI strings comes to the forefront as an effective library management representation scheme due to the uniqueness of the identifier.[35-37] Molecules are typically depicted as molecular graphs with atoms forming the nodes and bonds forming the edges. Graphical representations of molecules highlight their topological connectivity. A featurized version of these graphs, that is, a graph where each node and edge has an associated feature vector, is used in performing computations on molecules. A straightforward application of the molecular graph is the application of rule-based reaction templates to molecules. The templates describe the edges to be formed or removed between the molecular graphs of the substrates in a reaction to form the graph of the product molecule. Graphical representations also allow for substructure matching used by reaction templates to identify reaction centers. The change in connectivity of the molecular graph is achieved by performing modifications to the adjacency matrix.[36, 38, 39] Molecular fingerprints such as the Extended Connectivity Fingerprint (ECFP) take advantage of the connectivity information encoded in molecular graphs to generate molecular fingerprints. Similarly, the graphical information is taken into account for molecular property prediction (quantitative structure–property relationship [QSPR]/quantitative structure–activity relationship [QSAR] studies) through the repeated application of convolution operation to the molecular graph. Convolution-based approaches have also been employed in automated reaction prediction to identify possible products from a set of reactants.[40-42] Generative models have been developed recently that aim to generate molecules as molecular graphs using GAN.[43, 44] The order-invariant nature of the adjacency matrix lends to computational complexity in one-shot generation approaches. Nonetheless, the methodology has been found to be effective for smaller molecules. RNN and VAE-based methodologies have also been used for sequential graph generation in node-wise and fragment-wise fashions.[29-32, 45, 46] As an extension of molecular graphs, reactions have also been encoded in graphical format, thus allowing access to graph theory-based approaches in reaction outcome prediction and identification of pathways for reactions. Reaction hypergraphs have been used in reaction classification.[47] Molecular fingerprints form a class of representations where information about a molecule is encoded in a vector representation. These may include bit vectors such as Molecular ACCess System (MACCS) keys, which indicate the presence or absence of a particular sub-structure or fingerprints based on connectivity, such as Morgan fingerprints or topological fingerprints. First generated to aid isomeric structure screening, fingerprints have found a variety of applications. The primary use still remains database screening and substructure matching. Fingerprints are used in tandem with a distance metric such as the Tanimoto similarity or Dice coefficient to identify the degree of closeness between the query and the database.[48-50] The compactness of fingerprints allows for faster screening of databases. As an extension, reaction fingerprints have also been developed to find reactions similar to each other in the chemical space.[51, 52] The ability to generate fingerprints focused on certain aspects of the molecule allows for contextual extraction of information for specific tasks. This has led to widespread usage of molecular fingerprints as inputs for molecular property prediction, especially for drug-like molecules.[53-55] In a similar vein, localized reactivity information of molecules can be extracted through fingerprints to aid in product prediction.[52] Computer-aided representations capture latent abstractions of the molecule through nonlinear transformations. These embeddings represent the molecule in a continuous real-valued space and are usually extracted for a specific purpose.[23, 28] Invertible mappings of molecules allow for complex mathematical operations to be performed on molecules in an efficient manner. For example, convolution operations on molecular graphs followed by pooling of the node features generates a vectoral representation of the molecule, which is further used downstream for prediction of partition coefficients.[23] Generative models such as GANs and VAEs take advantage of this representation to generate embeddings of new molecules, which are then decoded to generate a SMILES representation.[29-32] Typically extracted using auto-encoders, the embeddings learnt depend on the type of information present in the input.[56-58] The dimensionality reduction achievable through the use of latent space embeddings is exploited in reconstruction of potential energy surfaces and to develop functionals for DFT calculations.[12, 59, 60] A wide variety of molecular representations have been discussed to encode spatial geometry and connectivity among atoms. However, descriptors that capture topological equivariance are most sought after to avoid confounding among molecular conformers.[61, 62] However, graph-based geometric and topological descriptors of molecules are ill-posed to describe multi-component systems and may even misrepresent emergent properties of component interactions by merely concatenating descriptors of individual molecules. Hence, molecular images of chemical structures are widely used in training ML models to predict solubility[63] and kinetics[64] of molecular mixtures. Coming to reactions, graphical and string-based representations stem from their applicability towards molecules. Reactions are encoded as modifications to molecular SMILES through atom-atom mapping or as graph edits to the molecular graphs. The bond-based representations of several molecules are widely supplemented with metadata about molecular charge densities and reaction energies from quantum calculations, in order to capture pairwise interactions and the algebraic additive nature of reaction properties uniquely.[65] The aforementioned representation would suffice if the task at hand was to train generative ML models that enumerate candidate products from a set of reactants, but it would be context-limited in tasks requiring translation from computational retrosynthesis to an experimental protocol. Context-free grammar-based ontological representations of molecules have been used in a neural seq-to-seq machine translation framework to capture hierarchical information of reactant molecules in predicting the products from a reaction, although the context of experimental conditions is still absent.[66] However, text-based reaction representations of a sequence of experimental steps have been mined from patents to train context-enriched seq-to-seq transformer language model (Smiles2Actions) to predict the experimental protocol for batch organic synthesis.[67] Materials representation by descriptors of molecular structure and composition, that respect symmetry, similarity, density, continuity, locality, and additivity to ensure that distinct materials do not have identical descriptors is vital to develop ML surrogates linking molecular-level material descriptors to its macroscopic properties by way of capturing QSARs/QSPRs.[68] Kernels to capture symmetry and continuity, distance metrics to capture similarity, summation of atom-centred symmetry functions (ACSFs) to capture density, connectivity matrices to capture interactions (e.g., Coulomb) or compositional variations, spectral decompositions of connectivity matrices to capture invariance, and persistence of topological homologues over radius to capture locality are among the phylogenetic tree of material descriptors that successfully embed the additive decomposition of material properties implicit in the functional forms of inter-atomic potentials from theoretical calculations.[69] Applying these meaningful transformations to metadata from high throughput theoretical calculations helps develop reliable, faster, and generalizable ML surrogates (QSAR/QSPR models) that have ushered in a new era of discovery in computational materials science via robotic self-driving labs (SDLs) for material synthesis.[70] The development of over 40 novel featurization approaches from topological representations of crystalline materials as molecular building blocks has advanced the field of digital reticular chemistry for materials discovery, with the promise of being transferable to other classes of materials.[71] The multi-scale material descriptors for QSAR/QSPR models are prediction task-specific and range from atom-centred descriptors for local properties, building unit-centred descriptors for shape and flexibility properties, and finally, more coarse-grained volume elements in the material space, like voxels or point cloud representations for global properties. Imaging data from scanning tunnelling and transmission electron microscopy, from which periodic features are extracted using VAE, is found to relate to material properties like polarizability and strain gradients,[72] as do neural network-extracted features from hyperspectral images of functional materials relate to their macroscopic thermochemical properties.[73] One may also use the operator-theoretic approach of using a hierarchy of Hamiltonian matrices to devise global material descriptors that are robust to structural confounders and guarantee the injectivity of ML-based QSAR/QSPR model to encourage transferable learning.[74] The Hamiltonian is the energy operator for the wave function in the Schrodinger's equations, whose eigenvalues are representative of the total energy of a quantum mechanical system. The use of context-aware text-embeddings like BERT and ELMo, as opposed to Word2vec and GloVe to capture collective associations of materials and molecules to target properties from unstructured knowledge in literature, enables the development of trustworthy QSAR/QSPR models for materials discovery.[75] Note that only a small portion of the associations are digitized into structured property databases. This also limits the chances of making future discoveries that were already reported in past publications. Very recently, the claims made by a self-driven autonomous lab (A-lab) of having discovered 41 novel compounds in 17 days were challenged by researchers who contested that some of the materials were misidentified because compositional order invariant descriptors for inorganic materials were not considered, while some of the other discovered materials, had already been reported in the literature.[76] Processes generally comprise several unit operations connected by material and energy flows and are adequately represented by flowsheets. The topology of flowsheets, in turn, has been represented in varied forms to aid exploration and generation. Representing chemical process flowsheets as strings facilitates the use of machine-learning algorithms applied to sequences, such as RNNs and transformers.[77] Generative models employing SFILES notation have been developed using large language models (LLMs). Pattern recognition and feature extraction tasks have been carried out through the usage of string notation.[78-81] Using images of process flow diagrams (PFDs) opens up the usage of convolutional operations for feature extraction. Pattern recognition in images of PFDs has been performed through incorporation of spatial relationships between process units.[82, 83] The sequential nature of PFDs lends itself to a graphical representation such as the P-graph, S-graph, SSR (state space representation), and so on. Generative models can be used for auto-completion of PFDs by addition of operational modules as nodes to the graph of the process. Canonical descriptions of P&ID diagrams such as the Data Exchange in Process Industry (DEXPI) Standard allow for vendor-independent exchange of information. DEXPI standard has been used as an input to generate graphical representations of P&ID diagrams to aid in sequential node/edge predictions using RNNs and GNNs.[84] Graphical representations also lend themselves to aid the superstructure optimization problem, attaching attributes corresponding to dynamics of material and energy flows of each node and edge and optimizing over a combination of these units subject to certain production and physical constraints provided as algebraic and differential equations.[85, 86] Time series records of process variables from measurement sensors have long been used to construct representations of process history,[87] and even to infer trends of the unmeasured process variables from hierarchical representations via more sophisticated soft-sensor modelling.[88] The modularity of chemical processes supports their digitization using knowledge graphs to represent semantic relations among heterogeneous entities, that is, inputs, outputs, and unit operations are designated as the nodes, while the material and energy flows among them are designated as the edges to improve production lines by efficiently assessing downstream environmental and economic impacts.[89] Graphical representation of ontologies results in knowledge graphs. Ontologies are a way of representing hierarchical knowledge by using logical theory to formalize conceptual semantics among entities and comprise: (i) a set of classes/unique concepts/entities, (ii) attributes of these classes, (iii) instances of classes, (iv) relations among classes, and (v) axioms or logic rules that are restrictions defined on the classes that regulate their properties.[90] Resource description framework (RDF) triples of the subject-predicate-object format are used for knowledge graphs parsed from ontologies.[91] Building ontologies and thereafter knowledge graphs is a very time-consuming and non-standardized practice, but it encourages interoperability among disparate process entities and decentralized databases by facilitating recommender systems, identifying inductive relations, and entity recognition.[92] To encourage the transferability and reusability of ontologies across domains,[93] only generic modelling knowledge is incorporated and very few relational restrictions are imposed to create a meta-ontology, with a provision of adding domain-specific knowledge separately, as with OntoCAPE, a large-scale ontology for computer-aided process engineering.[94] Querying knowledge graphs of sensor data represented as class instances has been used for process anomaly detection.[95] Process safety by risk analysis is achieved by querying knowledge graphs where equipment, chemicals, and flows are treated as entities whose attributes and relations are ascribed from text mining hazard and operability (HAZOP) reports,[96] while hazardous chemical management is achieved by named entity recognition using knowledge graphs of incident records.[97] Knowledge graphs constructed from databases or text mining efforts have been used to mitigate risk, enhance process safety, and even recommend green manufacturing alternatives across a variety of industries from petrochemical to pharmaceutical.[98-100] Plants constitute an aggregate of processes, but it is not trivial to represent them by linking process knowledge graphs at common pinch points, owing to the siloed and static nature of knowledge graphs, unless the processes are represented by meta-ontologies that can easily be reused. A distributed plant ontology, OntoSafe, adds to the capabilities of OntoCAPE and benefits from using semantics for plant safety supervision, based on process changes.[101] Task-specific representations at the plant level primarily cater to safety and decision-making involving several stakeholders. Unstructured data from incident records and HAZOP reports are mined to extract text-embeddings to classify hazards and thereby calibrate consequence severity.[102] Industrial safety knowledge (ISK) from HAZOP reports are represented as knowledge graphs,[103] followed by using description logic to build heavy-weight plant ontologies from knowledge-bases to diagnose faults and characterize hazards.[104] Once the elements of an ontology have been defined/standardized, the process of constructing knowledge graphs has been automated by mining unstructured text using language transformer models for entity recognition and relation discovery,[105] so that the entire product and process development lifecycles can be informed by considering not just material safety datasheets, mass and energy balances, and various process technologies, but also manufacturing guidelines set by stakeholders.[106] Supply chains encompass resources (goods, money) that move across enterprises, whose semantic knowledge is represented using modular supply chain ontologies (SCONTO) for interoperability in decision-making.[107] Several other ontologies have been used to describe the relations and restrictions on resource movement among entities that include suppliers, manufacturers, distributors, logistics, and customers, but are still far from catering to interoperability because they do not adequately represent the reality,[108] are highly non-standardized, and do not integrate information from ever-evolving supply chains.[109] This is proposed to be tackled by the industrial ontology foundry (IOF) that provides an open-source, standardized, collaborative, and modular approach to build enterprise ontologies.[110] Further, enterprises across several geographical locations are best represented as geospatial knowledge graphs that are constructed from embeddings of heterogeneous geographic information systems (GIS) in the form of images (satellite images, street views, aerial photos), text (tags, reviews, social media posts), and numeric data (weather, traffic).[111] Geospatial knowledge graphs enable decision-making using supply chain ontologies to be resilient to entity disruptions.[112] 3 HYBRID AI Building predictive models for physical processes is important for achieving the objectives of process systems engineering (PSE), which involves improving decision-making in the production of chemical products.[113] The development of digital models that represent a physical process is the key component in PSE.[114] Therefore, most efforts in chemical engineering have been to develop ‘digital twins’, which are virtual representations of the processes. Digital twins include a predictive component that is accurate, robust, and fast enough to be used for optimization, control, fault detection, and diagnosis.[115] Predictive models can be 1. Mechanistic: derived using physics-based knowledge about the process in the form of transport equations (mass, momentum, and energy balance), physical phenomena (absorption, adsorption, crystallization, etc.), chemical kinetics, initial and boundary conditions, or 2. Data-driven: developed by fitting the process data to one or more mathematical functions. Developing mechanistic models requires a thorough understanding of physical principles and mechanisms involved in the process and a computational capability for simulations. The parameters used in mechanistic models usually have a physical meaning. On the other hand, data-driven models are developed purely based on correlations between input–output data, and the parameters typically have no physical meaning, making the models less transparent in predicting the behaviour of the process variables. These models, however, are easier to develop than physics-based mechanistic models. Examples of data-driven models used in PSE are the traditional time series models such as ARX, NARX, and so forth, and the more recent ML models (Gaussian processes, neural networks, etc.). Recently, the abundance of data and enhanced computational capabilities have led to an increase in the use of ML. Such approaches have been used extensively for various PSE applications such as process monitoring, fault detection, control, optimization, and so forth.[1, 116] Usually, these are black-box models, having no knowledge about the underlying physics of the process. Although such models are easy to develop, they are highly system-specific, lack extrapolation capability and interpretability, and may produce physically inconsistent results. Hybrid models, also called as grey-box models, offer a promising solution to tackle some of these problems, with other advantages of robustness and efficient model development for complex processes.[5] Hybrid models in PSE first appeared in works of Psichogios and Ungar,[117] Kramer et al.,[118] Johansen and Foss,[119] Mavrovouniotis and Chang,[120] and Su et al.,[121] and have seen increasing use since then. A variety of methods for incorporating mechanistic knowledge into a ML framework can be observed right at the start of the hybrid modelling paradigm in chemical engineering, which has led multiple researchers to propose taxonomical guidelines for the use of hybrid AI in PSE. Note that while taxonomical guidelines exist for the use and implementation of hybrid AI in PSE, the applications have focused almost exclusively on the process scale. This reflects the fact that the modelling focus and first principles knowledge of many process systems engineers is at the process scale, but it also speaks to this being an opportunity to extend the scope of hybrid AI in PSE. Sansana et al.[122] provide a review on hybrid modelling and classified approaches into series, parallel, and surrogate models. However, it can be argued that surrogate models may not necessarily belong to the category of hybrid models. Bradley et al.,[123] reviewed approaches for hybrid modelling for cases where mechanistic models are used in some way during the training process of the data-driven model. Sharma and Liu[124] describe a comprehensive classification of hybrid modelling techniques, dividing the approaches into ML assisting science and science assisting ML. Rajulapati et al.[125] give a perspective on integrating physics with ML and classify hybrid models into residual models, first-principles constrained models, and first-principles initialization models. Gallup et al.[126] demonstrated the benefits of three different classes of hybrid models using a CSTR case study. Physics-guided architecture, physics-guided loss, and physics-guided initialization are used as classes for distinguishing different hybrid methods. Narayanan et al.[127] outlined a step-wise procedure to obtain different types of hybrid models in bio-pharmaceutical processes. The most relevant classes of hybrid models for chemical engineering are the semi-parametric approach (series–parallel architectures), physics-informed regularization, and physics-guided architectures. Most works on hybrid AI can be classified into these three types. 3.1 SEMI-PARAMETRIC APPROACH This form of hybrid modelling is the oldest and most popular approach. Semi-parametric or conjunction approaches can be further divided into series, parallel, and combined architectures. The term ‘semi-parametric’ derives from the fact that a part of the model is described using physics-derived equations having physically meaningful parameters, while the data-driven part is considered as non-parametric. Agarwal[128] described a framework for combining prior process knowledge with data-driven models in the context of state estimation and control using a semi-parametric approach. von Stosch et al.[129] reviewed applications of series–parallel hybrid model structures in PSE. McBride et al.[130] reviewed applications of semi-parametric hybrid modelling to separation processes, while Zendehboudi et al.[131] reviewed applications of semi-parametric strategies in chemical, petrochemical, and energy industries. Also, Schweidtmann et al.[132] reviewed the current state, challenges, and future directions of semi-parametric hybrid models. The serial architecture can be further divided into two types. For the first approach, outputs from one model are used as inputs to the subsequent model. Usually (see Figure 4A), a black box model is used to estimate the part of the physics-based model that is too complex to model. The output of this model is used as input to the mechanistic model. For example, in the work of Psichogios and Ungar,[117] the authors proposed a serial architecture where the kinetic parameters of a fed-batch bioreactor were predicted using a neural network and fed to a first principles model for simulation. Such models are useful when rich process data is available, but interpretation of mechanistic parameters is difficult. In the work of von Stosch et al.,[133] the authors use a nonlinear partial least squares method to approximate cellular system dynamics in a mass balance equation, thereby avoiding the unrealistic estimation of metabolic fluxes from concentration measurements. Recently, Nielsen et al.[134] used such a model in particle analysis in crystallization and flocculation applications. FIGURE 4 Open in figure viewerPowerPoint (A) Serial architecture: Output from data-driven model used as inputs for mechanistic model. (B) Serial architecture: Output from mechanistic model used as inputs for data-driven model. (C) Parallel architecture. As seen in Figure 4B, output from the first-principles model are used as inputs to the data-driven model. This type of model employs feature engineering, where some knowledge about the process is used to create custom inputs to be fed to the data-driven model. Although, this type of model is less used, there have been some reports in literature.[135-137] Recently, Yan et al.[138] applied serial architecture to predict gasification products using a neural network, in a process optimization framework during design stage. Knowledge about the thermodynamic equilibrium is used to generate feasible gas temperature inputs to a neural network that predicts product composition. The parallel architecture is given in Figure 4C. Such types of models aim to capture the unmodelled part of the physical system or the model-plant-mismatch by modelling the residuals using a data-driven approach. The outputs from mechanistic and data-driven models are combined to get a pooled prediction of outputs, resembling an ensemble-based ML algorithm. The first works in application of the parallel approach can be found in the works of Su et al.[121] and Mark et al.[118] Ghosh et al.[139] demonstrated a parallel structure for modelling residuals of system identification models in a control framework. Bikmukhametov and Jaschke[140] developed different types of parallel models using measurement and simulation data for virtual flow metering. There have been multiple works using a parallel architecture, especially in control and state estimation, as reviewed in the literature.[129, 132] Most works on semi-parametric approaches report increased efficiency in training the models and improved accuracy. The data requirement is also reduced for achieving an equivalent performance level. One major advantage of developing hybrid model architectures is an increasing capability to extrapolate in the unexplored input space. For instance, van Can et al.[141] demonstrate the extrapolation capability of series hybrid models for prediction of the pH effect on the conversion of Penicillin G using enzymes. Fairly accurate predictions were obtained for input values outside the range of experimental training data. Also, in the work of Narayanan et al.,[142] an improvement in extrapolation capability with increasing knowledge about the process is demonstrated. This shows that combining mechanistic and data-driven approaches enhances the performance of either and can complement each other in terms of model fitting. Depending on the application, availability of physics-based knowledge, and amount and quality of data, different semi-parametric architectures can be created to increase model performance. Though most reported works use either serial or parallel, a combined approach can also be used to model a complex system. 3.2 PHYSICS-INFORMED REGULARIZATION The idea for incorporating physics-based terms in the overall loss function of a neural network was proposed by Raissi et al.[143] and was demonstrated using PDE equations derived from physics. It was subsequently successfully adapted into several domains, as reviewed by Cuomo et al.,[144] by crafting the loss functions according to the application. Naturally, it was adapted to chemical engineering applications to model complex processes described using PDEs that capture spatiotemporal variation. In any black-box model, the parameters are obtained by minimizing the errors between predictions and true values of the labelled data. In physics-informed regularization, an additional loss function is included that incorporates the available knowledge about the process, for example, partial differential algebraic equations (PDAEs), thereby introducing a ‘learning bias’ during training of the data-driven model.[145] This nudges the ML model in the direction to learn the underlying physics of the problem, thereby increasing its extrapolation or generalization capability. As shown in Figure 5, the prediction loss is first generated using labelled data, and the predictions from the data-driven model are used to generate the regularization loss using physics-based model equations or constraints. One key difference from semi-parametric approaches is that the inputs for enforcing physics-based losses can be sampled independently of labelled data. The model can therefore be trained to extrapolate beyond the range of the input space. FIGURE 5 Open in figure viewerPowerPoint Physics-informed regularization. PDAE, partial differential algebraic equations. Physics-informed neural networks (PINNs) are applied across a wide range of engineering applications. After their introduction, there has been a notable increase in their applications to chemical engineering. Because of the extensive use of PDEs in the form of Navier–Stokes equations, a wide range of articles can be found in fluid mechanics and heat transfer that use PINNs. Multiple reports about both physics-guided architectures (which is covered subsequently) and physics-informed regularization in the field of fluid mechanics have been reviewed by Sharma et al.[146] Wang and Ren[147] use PINN to predict temperatures in a heat conduction process. Patel et al.[148] use a PINN to predict temperature profile in a PFR using limited measurements. PINNs have also been used to model adsorption-based processes such as chromatography[149, 150] and pressure swing adsorption.[151] These approaches are used in several other chemical engineering applications to predict process states. Chen et al.[152] develop a PINN for the initial estimation of phase-equilibrium calculations in shale reservoir models. In this work, a PINN is used to reduce the time required to predict equilibrium ratios. Chen et al.[153] provide a comparative evaluation of PINN models developed for various 1D, 2D, and 3D simulations in voltammetry and suggest best practices to be followed for developing PINN models in the field of electrochemical analysis. Takehara et al.[154] used a PINN as a surrogate model for predicting fluid and thermal fields in the growth of single bulk crystals and reported a significant reduction in computation time for predictions. Merdasi et al.[155] used PINNs to predict zeta potential in a mixing process and demonstrated the efficacy of PINN in comparison to finite volume method (FVM) simulations. Bibeau et al.[156] trained and used PINNs to predict the kinetics of biodiesel to show the regularization obtained by PINNs. Liu et al.[157] used PINNs to predict turbulent combustion fields using sparse data. Zhang and Li[158] developed a PINN to predict haemoglobin response in an Erythropoietin treatment. Ren et al.[159] developed PINNs for predicting the product composition of a biomass gasification process. In the work of Ryu et al.,[160] a PINN was developed to predict variables in a polymer reactor. Asrav et al.[161] apply a PINN for modelling an industrial wastewater treatment unit. Most of the models developed based on PINNs are used to predict process variables, which describe the state of the system, and are often modelled using PDEs. This enables researchers to develop digital twins for processes by creating spatiotemporal profiles using limited measurements. The advantages of PINNs are similar to those of semi-parametric approaches. Since there is no requirement of labelled data to enforce physics-based soft constraints, the desired input space can be easily explored and incorporated into the model. A physics-informed neural ODE is introduced in Sorourifar et al.[162] to model multiphase chemical reaction systems. The extrapolation capability of the model after adding information about the kinetics is demonstrated on a real-world experimental data set. The extrapolation capability and robustness of a PINN can then be utilized in process monitoring and control. For instance, Zheng and Wu[163] introduce a PIRNN in an MPC framework to control nonlinear processes and emphasize the utility of physics-based regularization on a process with uncertain parameters. Also, Franklin et al.[164] use PINN as a soft sensor to monitor flow rates in an oil well system. Wu et al.[165] develop a PIRNN-based MPC for controlling crystal growth and nucleation in a batch crystallization process. PINNs are also used in inverse problems or parameter estimation. Rogers et al.[166] use PINNs for identifying a model with time-varying parameters. Lu et al.[167] use PINNs to identify model parameters for an optimal finned heat sink system. Selvarajan et al.[168] introduce differential flatness in neural ODEs for parameter identification. Tappe et al.[169] use PINN-based design of experiments for parameter identification. There can be other forms of adding regularization to train a ML model. For instance, prior knowledge about the process was used to incorporate monotonicity constraints for training the neural network in the work of Muralidhar et al.[170] for predicting oxygen solubility in water. The improvement in performance over conventional neural network models for training with limited data and noisy measurements is demonstrated. However, at this time, PINN-style approaches dominate the literature. PINNs have shown a promise in increasing interpretability and generalizability of conventional feed-forward neural networks. Since the information about the physics of the process is added as a soft constraint in loss formulation, an issue of convergence to a sub-optimal solution may arise. Cuomo et al.[144] provide a comprehensive review for different types of PINNs developed in literature and discuss challenges related to convergence and error analysis and future research directions related to PINNs. The weighting parameter on residuals is a critical factor in determining convergence. Currently, most methods use a trial and error approach to obtain an appropriate value of this hyperparameter. There is a need to develop a generalized framework to address this issue. One way to fix this is by adding a hard constraint in the NN formulation. Asrav and Aydin[171] used a physics-informed recurrent neural network (PIRNN), trained using a genetic algorithm (GA) with regularization loss as a hard constraint to model a dynamic process, and it was reported that the testing accuracy for PIRNN increased despite a reduction in training accuracy as compared to conventional neural networks. GAs, however, impose a huge computational burden for optimizing model parameters. Nonlinear programming (NLP) algorithms such as interior point methods may be used as an alternative to GA. However, there is no guarantee of convergence of NLP algorithms. 3.3 PHYSICS-GUIDED ARCHITECTURE This type of model makes use of physics-based knowledge about the process in designing a neural network architecture. The neurons and hidden layers are arranged such that the parameters and connections represent physical knowledge about the process. A physics-guided hierarchical neural network architecture was proposed by Mavrovouniotis and Chang[120] for monitoring a complex distillation column. Instead of using a fully connected architecture, different sub-nets were defined at each hierarchical level to monitor specific aspects of the process. This modular design results in a sparser network than a fully-connected network and more interpretability of model outputs. Russell and Baker[172, 173] applied the subnet neural network modelling strategy, using prior knowledge, to model falling-film evaporator. Each sub-network represented a specific sub-system with a localized set of inputs. This resulted in a sparse and a relatively interpretable neural network. Munõz-Ibañez et al.[174] developed a material-properties that influenced the design of a hierarchical neural network to model material properties in a die casting process of aluminium alloys. Reductions in computational times have been achieved using a modular design of a low-fidelity neural network, representing a simplified physics-based model, followed by a high-fidelity neural network model, for predicting magnetic interactions between particles in close proximity of each other in colloids.[175] The prior knowledge about material composition and properties is used in designing the graph structure in a neural network. Another notable work in chemical engineering from this category would be the chemical reaction neural network (CRNN) by Ji and Deng.[176] The architecture of the neural network is designed in a way that the weights and biases of the model represent kinetic parameters, and inputs and outputs are the concentrations of reactants and products. This enables the model to learn the reaction parameters inherently using measurement data. Bangi et al.[177] develop a hybrid model called universal differential equation by modelling unknown process dynamics using a neural ODE[178] and combining it with known differential equations for a lab-scale batch fermentation process. Puliyanda et al.[179] use a neural ODE to obtain reaction mechanism-constrained kinetic models from spectroscopic data. Additionally, Muralidhar et al.[180, 181] used a physics-guided architecture to predict drag force on a particle. Machalek et al.[182] used knowledge about energy balance to isolate and predict individual phenomena occurring within each boiler of heat absorption units in a thermal power plant. Gallup et al.[126] developed a physics-guided neural network design such that each layer represents a component of the CSTR system. Physics-based feature engineering could also be considered a sub-category of this class. For example, Yang et al.[183] used physical information to create features that improved the extrapolation capabilities of the model in predicting wall shear stress in large eddy simulations. Prior knowledge based on physics about a CO injection process is used to craft features to conduct a risk assessment in the work of Yamada et al.[184] Most works also include physics-guided initialization as a separate class, but we argue that this can be considered as a subset of the physics-guided architecture. Carranza-Abaid and Jakobsen[185] have developed a neural network programming paradigm to build hybrid neural networks. Alhajeri et al.[186] develop a RNN structure inspired by physics to control a noisy process. 3.4 PERSPECTIVE All the effort in modelling a chemical process is usually to accurately and quickly simulate the process and predict values of state variables in a continuum of time and space. All effort is therefore to create a digital twin of the process that can virtually represent an actual physical process. The predictors in a digital twin completely lack interpretability, especially if neural networks are used for modelling. Hybrid models may overcome this problem using physics-based knowledge along with experimental data. Sitapure and Kwon[187] introduced a transformer-based approach to develop a hybrid model of time-series data with semi-parametric approaches to improve the interpretability of neural network-based digital twins. Uncertainty in measurements and processes can be handled effectively using hybrid models. Although any data-driven model may be used for developing hybrid models, artificial neural networks (ANNs) have gained the most popularity due to their universal approximation ability. All issues arising for developing ANNs are consequently inherited by hybrid models. Hyperparameter tuning is a major hurdle in developing better models and is an active area of research. This is mostly carried out using trial and error methods. Using an elaborate optimization scheme may be developed but it is computationally heavy. In a PINN, the weights of the loss function are the most critical parameters that require tuning. Since the regularization acts as a soft constraint, the value determines the local minima to which the solution converges. If an incorrect value is chosen, the accuracy reduces drastically, and adding physics information becomes counterproductive. There have been some works for adaptively generating an optimal value of the weights on loss function components. For instance, McClenny and Braga-Neto[188] developed a self-adaptive PINN by setting the weights as trainable parameters for every collocation point. These methods, effective for simpler equations in the physics domain as demonstrated in the paper, may be far more complex for chemical processes with a high number of states and parameters. One reason why PINNs have become popular is the use of automatic differentiation (AD) in Python packages for developing neural network models. AD stores the gradient of the function in the forward pass during training and accesses it while calculating physics-based constraints, thereby reducing computational effort. Calculation of gradients for predicted variables becomes much easier. Hybrid models are currently focused almost exclusively at the process scale and with operational data. There is an opportunity to extend them to other scales, however, it will require capturing/encoding knowledge or data in other forms (such as from databases) into the hybrid framework, and this may involve looking at different formats for PINNs other than including differential equations into the loss function. 4 HUMAN AND AI The integration of AI alongside human knowledge can enhance (and has enhanced) the efficiency and effectiveness of chemical engineering processes. Human-AI collaboration is a transformative partnership where each entity complements the strengths of the other, leading to enhanced decision-making, problem-solving, and innovation. In the context of chemical engineering, the collaboration between humans and AI can be categorized into two main types, depending on who ultimately benefits (see Figure 6): (1) Human complements AI and (2) AI complements human. In these frameworks, the human-AI interactions are mostly focused at the process and plant scales in the context of automation and automated systems, but they also span the scales of molecules, reactions, and materials in the context of active learning systems. FIGURE 6 Open in figure viewerPowerPoint Overview of human-artificial intelligence (AI) collaborations. 4.1 HUMAN COMPLEMENTS AI The human complements AI framework involves integrating human knowledge into AI systems to enhance their capabilities, particularly in real-world scenarios where data may be limited, expensive, or of poor quality.[170] It is important to highlight that when referring to ‘human’, we are specifically addressing domain experts such as plant operators, plant managers, or R&D scientists, depending on the nature of the problem. The aim is to build a robust AI model by incorporating human knowledge by following the workflow, which includes feature engineering, model design, model training, and model validation. While this workflow is familiar to data scientists, augmenting human expert knowledge can enhance model performance beyond purely data-driven approaches. The human knowledge may take various forms, including relationships between features, physical models, constraints, or scientific laws. This approach not only addresses challenges related to data availability but also promotes adaptability and trustworthiness in AI systems by end-users, ultimately leading to more robust and effective solutions. Next, we delve into various techniques through which humans complement AI. 4.1.1 FEATURE ENGINEERING The performance of the AI model largely depends on the representation of the feature vector, necessitating significant effort in designing preprocessing pipelines and data transformations during algorithm deployment.[189] Human experts play a crucial role in selecting relevant features for AI models, identifying key variables that contribute to model performance and enhancing efficiency. Feature engineering is the process of creating new features that capture the most important information from existing data for AI modelling. These new features might be ratios, differences, or other mathematical transformations of existing features. Domain expertise can help to design and extract features that are relevant, meaningful, and useful for the specific problem or goal. Their profound understanding of the domain enables them to identify key variables, relationships, and patterns that matter most in the specific context. Examples include using the log mean temperature difference in modelling heat exchangers and the use of dimensionless numbers such as the Reynolds number in flow systems. While considering feature engineering, it is crucial to recognize that it is an important yet labour-intensive aspect of ML applications.[189] Sometimes, different types of engineered features yield varied performances across models, such as the count features proving effective for DNNs and SVMs but not for tree-based methods such as random forests in the work of Heaton.[190] 4.1.2 MODEL DESIGN Model developers play a crucial role in designing AI models by determining architectures, selecting hyper-parameters, and designing loss functions to enhance performance and address specific tasks. Domain experts are vital in this step as well, by leveraging their expertise to select the most suitable model, considering domain knowledge, assumptions, and constraints. For instance, consider the scenario of a pressurized reactor vessel where the reaction rate increases with increasing pressure. An AI model controls the reactor pressure with the goal of maximizing production. However, if the model continuously raises the pressure without considering safety limits, it risks surpassing the design pressure, potentially leading to a catastrophic accident. This reflects the importance of domain knowledge during model design. Incorporating domain knowledge into AI models can be achieved through various means, including adjusting model initialization, modifying architecture, or modifying the loss function.[126] Model initialization refers to the process of setting initial values for the parameters (such as weights and biases of a neural network) before training begins. This initialization step is crucial as it can significantly impact the convergence and performance of the model during training. Model initialization involves a migration strategy wherein a new model is initialized by copying weights and biases from a pre-trained model on similar systems. This approach leverages existing models and data from the new process to build the model when there is limited data availability or to reduce training time. Model migration can be seen in the work of Lu and Gao[191] in building a soft-sensor model for online prediction of melt-flow length in injection moulding. To achieve this, existing mould geometries were used, reducing the need for extensive experimentation. The models corresponding to each mould geometry were aggregated and trained with new data, resulting in an effective model compared to the base case. Alternatively, in situations where similar systems are not available, model learning can be employed as an alternative. This technique entails initial pre-training of the model using simulated data generated by a rough first principles model. A notable example of this is in the work of Brand Rihm et al.,[192] wherein the lack of real plant batch distillation data is addressed by leveraging expert knowledge to create a preliminary simulated dataset. Subsequently, refining the model through training with actual plant data significantly enhances its model performance. Modification in architecture refers to a model structure that incorporates principles and insights from physics to enhance its performance, interpretability, and generalization capabilities. Such architectures typically involve various adjustments to the network's structure, connections, neuron functionalities, or combinations of first principle models to better align with the underlying physical processes governing the system being modelled. Daw et al.[193] present a physics-guided architecture employing LSTM models to simulate lake temperature dynamics. Leveraging the known monotonic relationship between water density and depth, the authors introduce physical intermediate variables into the model architecture, thereby ensuring the preservation of monotonicity within the LSTM framework. Alternatively, one can adopt a hybrid approach by combining AI with first principle models. This can be achieved through a series combination, where outputs from each model feed into the other, or by working in parallel to enhance overall performance. For more details, refer to Section 3 on hybrid AI models. As mentioned in Section 3, modification of the loss function is a technique used to enforce faithfulness to physical laws by penalizing model outputs that violate these laws. It is typically implemented by augmenting the standard loss function with an additional term based on the deviation from physical constraints. The purpose is to guide the model towards solutions that not only fit the data but also conform to the underlying physical principles, such as in PINNs. A notable example is PANACHE,[151] which utilizes PINNs to accurately simulate cyclic adsorption processes. It achieves this by integrating a physics-constrained loss function, enabling the model to learn underlying partial differential equations without the need for system-specific inputs like isotherm parameters. This approach enhances computational efficiency and reliability while ensuring accurate representation of adsorption phenomena. 4.1.3 MODEL TRAINING Human expertise can complement AI during model training by providing domain knowledge, context, and nuanced understanding that may not be readily captured by algorithms alone. Through a human-in-the-loop mechanism, experts can curate and label data, offer insights on ambiguous cases, and guide the training process towards more relevant and meaningful outcomes. One such approach is active learning. Active learning is a type of semi-supervised learning with a query strategy for selecting specific instances from which it wants to learn. This method, when coupled with human expertise within a human-in-the-loop framework for labelling selected instances, enhances the refinement process, leading to more meaningful outcomes in a much faster timeline. Moreover, active learning operates efficiently with smaller labelled datasets, thereby reducing manual annotation expenses while maintaining elevated accuracy levels. This symbiosis of human insight and ML prowess sets the stage for advancements in the realm of high throughput experimentation (HTE). HTE is a technique facilitating the simultaneous execution of large numbers of experiments to be conducted in parallel, offering increased efficiency compared to traditional experimental approaches.[194] In the R&D labs of the chemical industry, HTE is primarily utilized for the rapid determination of solvents, reagents, and catalysts by examining a wide array of reactions. Notably, the study of heterogeneous catalyst performance involves intricate interactions among various catalyst attributes and operational conditions, resulting in a multidimensional catalyst discovery and optimization space. In general, HTE studies are capable of screening large amounts of catalysts efficiently. However, their efficacy in discovery depends on encountering a high-performing catalyst within the screened parameter space. Therefore, the integration of HTE techniques with predictive algorithms is advantageous to guide experiments effectively.[195] The integration of ML and HTE into integrated workflows has given rise to SDLs, leveraging the complementary strength of both approaches. The ML component of the workflow aids in predictions and experiment design, while the HTE component executes the suggested experiments, allowing the results to inform ML model updates.[196] The synergy between ML and HTE becomes much more powerful when combined with a fully automated system in a closed-loop fashion to autonomously perform the experiments selected based on the ML model and decision-making algorithm.[197] In the past decade, SDL applications have shown promise in diverse areas such as the discovery/development of complex organic compounds,[198-200] nanomaterials,[201-206] thin-film materials,[207, 208] and carbon nanotubes.[209] Despite these achievements, the full potential of SDLs in chemical and materials sciences has been hindered by challenges such as the lack of standardized hardware, accessible software, and user-friendly operational guidelines, as well as the inability to incorporate physics-based models easily. Several open-access SDL software packages have emerged to facilitate autonomous experimentation in chemical and materials sciences, including ChemOS[210] and ARES OS.[211] These software platforms incorporate distinct experiment planning algorithms such as Phoenics,[212] Gryffin,[213] and Golem.[214] Phoenics, a Bayesian global optimization algorithm based on kernel density estimation, proposes new experimental conditions by leveraging prior results to minimize redundant evaluations.[212] Gryffin, a general-purpose Bayesian optimization framework for categorical variables, relies on kernel density estimation and utilizes descriptors to approximate categorical variables in a continuous space.[213] Golem is an algorithm applicable to any experimental planning strategy, accounting for input uncertainties through probability distributions to locate optima robust to variations arising from uncertainties in experimental conditions or instrument imprecision.[214] A comprehensive description of the ML algorithms utilized in SDLs can be found in the cited literature.[215-218] While SDLs hold tremendous promise, several challenges persist, including the automatic dispensing of solids and handling heterogeneous mixtures. Dispensing solids, especially those with varying properties and in small quantities, remains a significant challenge. Additionally, handling heterogeneous mixtures poses difficulties in systems designed for liquid transfer due to the risk of damage or malfunction. Addressing these challenges is crucial for unlocking the full potential of SDLs in chemical and materials sciences.[196] 4.1.4 MODEL VALIDATION In model validation, human expertise is a crucial asset. As AI algorithms process vast amounts of data and identify patterns, human experts bring contextual understanding and detailed judgement to the validation process. Human experts possess the ability to interpret results in the broader context of their field, discerning between genuine insights and spurious correlations that might mislead an AI system. Moreover, they can identify biases or limitations within the data or the model itself, ensuring that the validation process is comprehensive and rigorous. One such approach is adversarial AI (discussed next). Since the 1960s, cybersecurity in the process industry, particularly in systems like SCADA, has evolved from mainframe to distributed architectures with LAN introduced in the 1980s.[219] Presently, these systems are complex and highly interconnected, increasing vulnerability to cyber-threats.[220] For example, in 2010, attackers used the Stuxnet worm to infect the Natanz nuclear plant's network via USB. It manipulated centrifuge speeds, causing damage and halting uranium enrichment for a week, leading to major economic losses.[221] Cyber-attacks can have significant consequences even without targeting process equipment. A cyber-attack on an operator's computer leaked confidential data from multiple Japanese nuclear plants onto the internet, including inspection forms and reports from 2003 to 2005.[221] The past two decades have seen a surge of AI and ML techniques for different applications, such as process monitoring and predictive maintenance in chemical plants. However, recent research has uncovered a series of security vulnerabilities within ML models.[222] These vulnerabilities not only pose economic and reputational risks but also have the potential to trigger catastrophic events such as hazardous material releases, fires, and explosions, with severe consequences for workers, the population, and the environment, just as with cyber-attacks on conventional systems. The necessity for attention in this direction is underscored by a survey conducted by Microsoft.[223] In this survey, the authors interviewed 28 ‘security-sensitive’ organizations in fields such as finance, cybersecurity, and food processing, aiming to comprehensively elucidate the tactical and strategic methodologies employed to safeguard ML systems against potential attacks. The study's findings highlight a prevalent deficiency: a significant portion of ML engineers and incident responders lack the necessary proficiency to fortify industry-grade ML systems against adversarial threats. Additionally, the research revealed that 25 out of the 28 organizations surveyed acknowledged the absence of appropriate tools essential for effectively securing their ML systems. To address these vulnerabilities, there has been significant interest in adversarial ML (AML). AML a mixture of cybersecurity and ML, is most commonly defined as the design of ML algorithms that can resist sophisticated attacks and the study of the capabilities and limitations of attackers.[224] Attacks on ML models can target various components, including training data, model parameters, and desired outputs. AML attacks are classified based on the attacker's knowledge of the ML model's inner workings into three categories: white-box, grey-box, and black-box.[225] White-box attacks occur when attackers possess complete knowledge of the ML model, similar to the model's operator or developer. For instance, FGSM,[226] C&W,[227] JSMA,[228] and Deep-Fool[229] are examples of white box attacks. Grey-box attacks happen when attackers have partial knowledge of the model, potentially causing model failure. Black-box attacks involve attackers operating blindly without any knowledge of the model. In a black-box attack scenario, various methods like decision-based attacks,[230] alternative model attacks,[231] one-pixel attacks,[232] and others are employed. Another way to categorize AML involves assessing the attack's capability and its impact on the ML model. This classification encompasses poisoning, evasion, and oracle attacks.[233] Poisoning attacks disrupt the model training process by corrupting either the training data or the model's logic itself, sometimes targeting the input parameters of ML models. Evasion attacks manipulate the ML model's decisions to induce misclassifications but don't affect the training process. Oracle attacks occur when an attacker crafts a malicious substitute for the original ML model by accessing its application programming interface (API), a trend amplified by the proliferation of commercial cloud-based computing services. Oracle attacks can be further subdivided into three categories: (i) Extraction attacks, which glean architectural details by observing the model's predictions and probabilities; (ii) Inversion attacks, which involve reconstructing the training data; and (iii) Inference attacks, facilitating the identification of specific data points within the distribution of the training dataset. In response to such attacks, researchers have also proposed and adapted numerous defence methods such as GAN-based defence,[234] adversarial training,[235] defence distillation,[236] and adversarial example detection.[237] Early approaches, such as information hiding and randomization, aimed to increase model robustness.[238] More recent studies have categorized defensive techniques into reactive and proactive approaches.[239] Reactive approaches detect adversarial attacks post-model deployment, while proactive approaches focus on designing models inherently resistant to attacks. Although significant advancements have been achieved in AML, challenges persist within the chemical process industry. Adversarial examples are predominantly utilized in image classification, whereas chemical process data is often presented in multivariate time-series format. Consequently, there is a pressing need for dedicated research in AML, particularly in domains like chemical engineering where safety is of paramount importance. Many defensive techniques are specialized to counter specific attacks and may be susceptible to other forms of attacks. Additionally, these defence mechanisms are closely tied to specific ML models and network architectures. Thus, there is a demand for more generalized defences and ML designs that inherently possess resilience against diverse attacks. Predicting the exact attacks to train against, as well as those to disregard, presents further challenges.[240] Ultimately, attempting to train models against every existing ML attack poses a significant challenge within the chemical process industry. In conclusion, while AML has made significant progress in addressing security concerns in other domains, ongoing research and development are crucial in the chemical industry to enable the widespread deployment of AI in process industries. To our knowledge, there has been limited work done in this area, indicating future opportunities for the development of robust AI and ML systems resilient to adversarial attacks. 4.2 AI COMPLEMENTS HUMAN AI complements humans involves leveraging AI to enhance human capabilities, particularly in tasks where human expertise may be limited, time-consuming to deploy, or prone to errors. Contrary to the concept of human complements AI, where human knowledge is integrated into AI systems, AI complements human focuses on how AI technologies can support and empower human experts. This entails utilizing AI algorithms for tasks such as data analysis, pattern recognition, and decision support. Rather than replacing human expertise, AI serves as a tool to augment and amplify human capabilities. By harnessing the strengths of both human intelligence (HI) and AI, organizations can achieve greater productivity, make better-informed decisions, and solve complex problems more effectively. Moreover, AI complementing humans fosters collaboration and mutual learning, leading to continuous improvement and adaptation in dynamic environments. The various techniques through which AI complements humans are presented next. 4.2.1 MODEL EXPLANATIONS In the chemical industry, processes are developed by process engineers and executed by process operators based on their knowledge in chemical engineering and practical field experience. Therefore, domain experts in chemical engineering may find it hard to trust the decisions and recommendations made by AI-based models, limiting their practical adoption. This challenge is especially prominent in cases that demand a high degree of reliability, a common requisite in the chemical industry. To overcome this challenge, recent years have witnessed the development of distinct Explainable Artificial Intelligence (XAI) techniques to enhance the interpretability of AI models. The most common ways humans interpret any system or process is by explanation-by-example, feature relevance explanation, visual explanation, rule-based explanation, or explanation-by-simplification,[241] as shown in Figure 7. FIGURE 7 Open in figure viewerPowerPoint Overview of explanation types. Explanation-by-example considers the extraction of data examples that relate to the result generated by a certain model, enabling a better understanding of the model itself, similar to how humans behave when attempting to explain a given process. These techniques include: (1) Prototype explanations, which entail the identification and presentation of typical examples that best represent a specific prediction. Such explanations play a crucial role in identifying typical process conditions that lead to desired outcomes, aiding in the optimization of chemical processes. (2) Adversarial explanations (in contrast) serve the purpose of revealing potential weaknesses and vulnerabilities within the model's decision-making process. These adversarial examples are crucial for identifying potential vulnerabilities in safety-critical systems, ensuring the reliability and security of chemical processes. (3) Counterfactual explanations involve the generation of examples or scenarios that are similar to the input but would result in a distinct prediction from the model. These counterfactual examples assist operators in understanding how minor adjustments to operational parameters can rectify the abnormality, thereby facilitating process control and troubleshooting. Manca et al.[242] present the XAI dashboard for the process industry, which includes a statistical module for data visualization and an XAI module for exploring counterfactual explanations at varying levels of abstraction. An illustrative application in batch process control showcases the utilization of counterfactual variable values that must be altered to attain a target outcome. Concurrently, Harinarayan and Shalinie[243] established a process monitoring framework with dual objectives: the provision of explanations for AI predictions and recommendations to restore processes from abnormal to normal states. Employing the TreeSHAP method, the study achieves explanatory insights into fault identification in the Tennessee Eastman process. Furthermore, it employs the diverse counterfactual explanations method to recommend corrective actions in response to identified faults. Feature relevance explanations aim to describe the functioning of complex models by measuring the importance of individual features in the model prediction. In the process industry, these methods play a crucial role in identifying the key factors, such as temperature, pressure, and reactant concentrations, that most significantly influence the outcomes of a chemical process, thus aiding in optimizing and controlling processes efficiently. In chemical engineering, the Shapley additive explanation (SHAP) is a widely used feature relevance explanation method for interpreting FDI models in process systems[244, 245]; However, it is important to note that the computational complexity of SHAP increases with an increase in number of features. To mitigate this challenge, Bhakte et al. have proposed the integrated gradient (IG) method, which effectively utilizes gradient calculations to operationalize SHAP.[246] Additionally, Agarwal et al. introduced the layer wise relevance propagation method to explain FDI results and employed it to prune irrelevant input variables, thereby enhancing the accuracy of test classifications.[247] Sivaram and Venkatasubramanian presented one such work that generates a textual explanation. This work pioneers the development of an XAI framework capable of generating mechanistic and causal explanations by integrating symbolic AI with numeric AI.[248] In contrast, Gandhi and White employed the LIME methodology to acquire explanations based on molecular descriptors.[249] These explanations are subsequently converted into textual form by leveraging the GPT-3 text-davinci-001 model. These explanations are further used to interpret the molecular structure property predictions. Visual explanation refers to techniques and tools that use visualizations to make AI model predictions or decisions more understandable and interpretable for humans. Visual explanations often include visual representations such as heatmaps, feature importance plots, variable attentions, and other graphical elements that highlight the most relevant input features, regions, or patterns contributing to a model's output. In chemical engineering, visualization is important in tasks like quality assurance and process monitoring. For example, Sun et al.[250] used a class activation map (CAM) to generate heatmaps that characterize the machine's status using real-time vibration footage, depicting the normal or fault state of the cantilever beam and water pump. A similar application was also seen in process monitoring, where Bhakte et al.[251] employed the Grad-CAM technique to identify the input data feature responsible for the occurrence of a fault in a chemical process. Danesh et al.[252] used partial dependence plots, individual conditional expectation, and accumulated local effects plots to understand the AI model used in a case study of a combined cycle power plant. Wu et al.[253] leveraged self-attention weights to interpret FDI results, while Aouichaoui et al.[254] utilized attention weights as explanations, representing the molecular components contributing to property predictions. Additionally, Schwaller et al.[255] applied attention mechanisms to interpret the atom-mapping information between products and reactants. Explanation-by-simplification encompasses techniques wherein an entirely new system is constructed based on a complex model to be explained. This simplified model typically attempts to optimize its likeness to the complex model while keeping a similar performance score. The simplification of the complex model makes it more interpretable, leading to more transparent process control strategies. One such method is limit-based explanations for monitoring (LEMON) that build the local linear model in the vicinity of input samples to explain the FDI results generated by AI models.[256] LEMON uses the alarm limits, which makes the explanations more operator-friendly and easier to understand due to model simplification. Rule-based-explanations employ a rule-based approach to elucidate the outcomes of AI models utilized in various processes. This method entails the generation of interpretable rules tailored to explain decisions, aiding users in comprehending the intricacies of decision-making processes inherent in AI systems applied to chemical engineering. In chemical engineering applications, the rule-based approach allows engineers and stakeholders to not only trust the decisions made by AI models but also gain insights into the governing principles behind those decisions. For instance, in the optimization of chemical processes or the design of new materials, rule-based explanations enable professionals to comprehend the variables, conditions, and parameters that significantly influence AI-driven decisions. This transparency becomes crucial for ensuring the safety, efficiency, and reliability of chemical engineering systems, aligning technology with human expertise in a synergistic manner. The selection of the aforementioned techniques allows end-users to choose the most suitable approach that aligns with their domain knowledge and serves specific purposes, providing diverse insights into black-box models. In the above classification of XAI methods, we acknowledge the potential for overlap among the categories. Notably, some methods may exhibit dual characteristics, such as feature relevance and visual explanation. An important consideration arises regarding when to leverage XAI and when not to. XAI proves particularly valuable during high-stakes decision-making and model development. In the chemical industry, where decisions impact operational safety and environmental concerns, XAI provides transparency, enabling engineers to grasp the rationale behind decisions and adhere strictly to safety protocols to prevent accidents and mitigate risks to human health and the environment. Similarly, during model development, XAI aids in understanding factors influencing predictions and identifying and rectifying biases, thus facilitating model debugging. Conversely, it is essential to outline situations where XAI may not be necessary: first, in low-stakes applications where errors have minimal consequences, such as in niche product R&D or experimental trials with negligible economic impact. The second instance is in situations with minimal interpretability needs, where decisions are straightforward and well-understood, and the demand for interpretability may be marginal. Third, in cases where there is a trade-off with performance; incorporating XAI may lead to compromises in performance, particularly when performance is paramount. Finally, a self-interpretable model (decision tree) can be preferred, especially when the accuracy difference between white box and black box models is negligible. 4.2.2 OPERATOR MONITORING In modern process industries, despite advanced automation and safety protocols, over 70% of accidents stem from human errors.[257] Addressing this challenge necessitates methodologies for assessing operators' competence. A notable contribution in this domain is the AI-powered eye-tracking system developed by Shajahan et al.[258] This system, known as Dhrushti-AI, establishes a robust framework for capturing operator cognitive behaviour within process plant environments through multi-screen-multi-user eye tracking capabilities. By integrating facial recognition algorithms and image processing techniques, such systems offer valuable insights into operators' cognitive capabilities, thereby supplementing human judgement and bolstering safety across industrial settings. Within this framework, various image processing methods, such as thresholding, closing, and contour extraction, are employed to detect pupils in captured images. The eye tracker accurately estimates gaze direction by leveraging information such as head orientation, pupil centre locations, and personalized calibrated gaze models. It is widely acknowledged that understanding cognitive behaviour is pivotal in addressing the root causes of abnormal situations and accidents. 4.3 HUMAN-AI CONFLICT Incorporating AI-based decision-making in process industries is quickly gaining traction, and it is but natural that dissonance can occur between the operator's decision and AI's. Arunthavanathan et al.[259] classify these conflicts into (i) Observation conflicts, (ii) Interpretation conflicts, and (iii) Action conflicts. Situational awareness demonstrated by human operators affects human decision-making to be an emotional and intellectual process, allowing for creative problem solving, which is found to be lacking in AI-based decision-making. Furthermore, AI-based decision-making is closely linked to its ability to contextual-learn these decisions from available data. With projected trends in AI's learning capacity, solely AI-based decision-making in process industries can be a realizable future, but at present, the general consensus is that AI needs to be overseen by HI owing to safety concerns.[260] 5 GENERATIVE AI Process engineering concepts enable chemical transformation at the highest scale determined by macroeconomic trends. The inherent challenges involve new product discovery, design and upscaling, unit operations development, and integration of multifunctional processes while meeting quality and safety standards and reducing costs and complexity with increased business flexibility. This underscores the significant potential for generative AI applications in various aspects of process engineering because of its multi-model and multi-scale big data processing and generative capabilities, along with natural language augmentation. In this section, we highlight the key generative AI work done across process system engineering, that is, in process design and development, process modelling and control, and process diagnosis and prognosis. 5.1 PROCESS DESIGN A significant fraction of generative AI applications are focused on process design, including molecular and materials design and process development. 5.1.1 PRODUCT DEVELOPMENT The process development cycle involves novel compound generation, property prediction, and optimization. Properties like synthesizability, solubility, affinity, and drug-like characteristics such as ADMET (adsorption, distribution, metabolism, elimination, and toxicity), and so forth are specifically targeted in the pharmaceutical and chemical industry. AI has been used for targeted product development without synthesizing materials with significantly lower resources. In the molecular discovery space, RNNs, graph neural networks (GNNs), autoencoders, and specifically VAEs, GANs, and most recently, transformer-based architectures have shown potential in generating novel products with desired properties. RNNs were the earliest architecture that utilized generative capabilities in molecular design applications, property optimization, and inverse design settings.[17, 261-264] Notably, Segler et al. applied long short-term memory (LSTM)-based models along with RL to fine-tune the molecular property for ab initio drug design.[262] GNNs are especially well-suited for molecular design due to their ability to model non-Euclidean data and representations similar to the chemical structures.[265-267] Various authors used variants like the message passing neural network (MPNN),[268-270] Schnet,[271] and graph convolutional networks (GCNs)[272] to predict properties and obtained better performance than any previous model. Although GNNs show good accuracy, they require large amounts of data to learn the mappings. Autoencoders (AEs) were employed next, but had limited success due to the sparse representation by simple AEs. However, the probabilistic version of AEs, known as VAEs, proved effective in modelling the latent distributions of molecular structure to map and extrapolate the properties. In 2018, Gomez-Bombarelli et al. first mapped molecular properties and designed new molecules using VAE.[23] Further, Flam-Shepherd et al. added a MPNN to the VAE architecture to enhance the performance of the model.[273] Moreover, variants like junction tree VAE (JTVAE) have shown immense potential in molecule generation.[274] The conditional version of VAEs usually constrains the representation towards a smoother distribution surface that is optimized for desired goals such as novelty, validity, and functionality.[275-278] The generative nature of adversarial networks has also been utilized for molecular design to map and generate the molecular structure distribution. The competing nature of the generator and discriminator were utilized to generate novel molecules with desired characteristics. Cao et al. and others applied GANs with RL-based optimization to achieve the required property for chemical molecule generation.[43, 279, 280] The adoption of transformers for de-novo design of molecules is favoured by the shared similarity of the synergistic combination of building blocks between natural language and molecular representations.[281, 282] These models have shown state-of-the-art performance compared to their counterparts.[283, 284] Transformer models utilize various architectures depending on the end application, such as encoder-decoder architectures (mapping and translation), encoder-only models (representation learning like property prediction[285-287]), and decoder-only architectures (primarily generative applications like generating novel valid molecules[288]). Honda et al. reported the use of transformers for the prediction of molecular properties in 2019.[289] The conditional variants of transformers such as MolGPT[288] constrain the input with desired objectives such as properties or scaffold information to guide molecule generation. The notable applications of the transformer model for molecular representation and property prediction include MolBERT[290] and ChemBERTA.[291] The outcome of these models can be further optimized towards certain objectives using techniques like Bayesian optimization[23, 292] and RL. Although the majority of generative AI efforts have focused on the design of drug-like molecules, yet we observe the emerging trend across other chemical sectors such VAE-based solid-state materials design,[293] porous crystalline material discovery,[294, 295] and GAN-based design of metal organic frameworks.[296] 5.1.2 PROCESS DEVELOPMENT Conventional methods like full factorial designs,[297] fractional factorial designs,[298] central composite designs,[299] designs for mixtures,[300] and optimal design methods such as definitive screening designs[301] have traditionally been used for process development.[302-304] However, the uncertainty in design space, intractability, laborious rule-based compilation, and scalability issues limit their application for optimal process development,[305] and the recent use cases of generative AI in forward synthetic, retrosynthetic, and condition recommendation are promising in this direction. The molecular transformer model of Schwaller et al.[306] can predict reaction outcomes based on available reactants and reagents and is considered a landmark in this direction. It utilizes the transformer architecture to achieve over 90% top-1 accuracy. Inspired by this, some retrosynthetic models based on the same architecture have been proposed.[20, 307-309] Further, Wang et al. proposed a single-step template-free transformer-based method to improve the chemical validity and diversity of the previous work.[310] Lin et al. improved the model accuracy further in their work.[19] Other notable variants include the graph-enhanced transformer and hybrid models.[311, 312] Moreover, generative AI has also shown the potential to infer the appropriate reaction conditions, but limited to a single reaction class.[313, 314] The lack of high-quality data makes it difficult to develop a model to extract reaction kinetics. However, Angello et al. reported a closed-loop workflow to discover general reaction conditions.[315] Schwaller et al., on the other hand, applied the transformer model to predict reaction yields.[17, 287, 316] Similarly, Sato et al. utilized a MPNN model for the prediction of yield.[317] Likewise, Sandfort et al. have utilized multiple representations to improve yield prediction.[318] Thus, generative AI has shown promising results, but the lack of multi-step synthetic capabilities and process condition augmentation, such as catalyst, solvent condition, and so forth, requires further improvements. Another noteworthy trend is the introduction of automated flowsheet generation, where generative transformer models have taken centre stage. Examples such as ‘Learning from Flowsheets’[80] and ‘Flowsheet Synthesis through Hierarchical RL and Graph Neural Networks’[319] highlight advancements in implementing the representation and enhancement of flowsheets. These methods include the representation of process topology using simplified flowsheet-input line-entry system (SFILES)[77] and mapping of PFDs and P&IDs for generative purposes[78, 80, 81, 84] as also highlighted in Section 2. Approaches like RL using deep Q-networks[320] and hierarchical RL[319] introduce a structured and automated approach to flowsheet synthesis. The hierarchical framework systematically breaks down the generation process, improving efficiency and the quality of flowsheet representations. Further, Oeing et al. utilized the RNN to generate the subsequent processing stage[84] and Vogel et al. utilized the transformer model for flowsheet completion.[80] Further, they utilized the transformer model for decentralized control generation for the given PFDs with top-5 accuracy above 85%. Along with control structure design, the method showed promise in piping and instrumentation diagrams (P&ID) generation as well.[81] 5.2 PREDICTIVE MODELLING AND CONTROL The next aspect of process engineering includes modelling process behaviour for process monitoring, control, and optimization.[321-323] Although the literature highlights the success of generative AI in multivariate time series forecasting, the full process scale application for the chemical industry is yet to be seen.[324, 325] In this direction, fully automated chemical synthesis using robots is a future trend. The advantage not only is the faster and more efficient synthesis of even hazardous material but also in achieving expert-level yields and purity. The conceptual and physical realization of this concept using the AI robots that perform experimentation, synthesis, and analysis at a laboratory scale has been shown in the cited literature.[198, 208, 326] Moreover, the emergence of visual synthesis through generative models has made considerable progress in capturing and reproducing complex visual patterns within process systems, leading to better monitoring and control. Specifically, works like ‘Bubble Generative Adversarial Networks for Synthesizing Realistic Bubbly Flow Images’[327] and ‘Generative Principal Component Thermography for Enhanced Defect Detection and Analysis’[328] highlight the versatility of GANs in depicting various processes, whether capturing bubbly flow dynamics or improving defect detection in manufacturing. These models demonstrate a robust capability to improve decision-making in complex and dynamic systems. 5.3 PROCESS DIAGNOSIS AND PROGNOSIS In process engineering, process diagnosis and prognosis involve identification of the root cause of underlying faults and, subsequently, time-to-failure (TTF) estimation by extrapolating the underlying root cause. These tools provide a foundation for condition-based predictive maintenance while mitigating the risk of unplanned downtime. Traditionally, the method involves techniques like PCA-based process diagnosis, transfer entropy,[329] time delay analysis,[330] Granger causality,[331] and process causal maps[332, 333] to identify the root cause of the anomaly. However, these methods have been shown to struggle for processes with a very large number of variables that are highly correlated, leading to a smearing effect.[334] Therefore, generative AI is particularly suited to model the vast topological space of multivariate time series. Some work for anomaly detection has been reported in the literature,[311, 335, 336] and has the potential to be applied directly to process level data. Figure 8 provides an overview of generative AI methods in process systems engineering with the type of data and end-use application. FIGURE 8 Open in figure viewerPowerPoint The application of generative artificial intelligence (AI) methods in process systems engineering. GAN, generative adversarial networks. VAE, variational autoencoders. FBM, fractional Brownian motion. LLM, large language model. EBM, energy-based model. 5.4 FUTURE The innovations in generative AI have already transformed its capabilities from unimodal applications such as molecule to molecule[306, 308, 337] to multimodal applications such as spectral to compound,[338] natural language (query-based molecule generation[339]), inferring experimental procedures,[67, 340] and so forth. The continually expanding range of capabilities exhibited by large language models like GPT and BERT[341] serve as motivation for researchers to leverage these capabilities to effectively address unique challenges faced by generative AI in the process industry, such as (1) scarcity of process level quality data and experimental uncertainty[342]; (2) suboptimal representation learning leading to generative gaps[343, 344]; (3) lack of process-specific benchmarks[342, 345]; (4) process level integration issues; and (5) interpretability and ethical concerns associated due to the black box nature of generative AI, hindering its ability to overcome the confidence barrier. Use cases like Ligppt[346] that incorporate GPT-style architectures for generative applications are promising in this direction. The low data issue necessitates conducting high-quality high-throughput experimentation or meta-learning coupled with data mining to collect the reported data. Moreover, the use of active learning,[347, 348] transfer learning,[349] GANs, and hybrid models[131] are shown to be robust in such scenarios. RL[350-353] or Bayesian-based goal directed generative AI[255] have shown superior performance in limited data regimes. Further, benchmarking platforms like GuacaMol[354] that give different performance criteria to evaluate the models are available but limited to product development, and such platforms are required at process scale benchmarking. Recent innovations in generative AI have shown emerging capabilities such as chain of thought (CoT) reasoning,[355] instruction following,[356] and few-shot[357] or zero-shot[358] generalization with their sizes. Further, innovation in quantum computing has the potential to reduce the computational load and improve the model performance and size of generative AI.[359, 360] These new capabilities have the potential to benefit process development, monitoring, and control, along with better interpretability and interactivity. The development of an AI-based centralized process operator to monitor every process operation, control in real-time, and optimize conflicting objectives such as reaction yield, cost, purity, emission, and so forth, is an aspirational target. In this direction, agents powered by large language models[361] along with chain-of-thought capabilities represent a promising application in generative AI with the capacity for reasoning, executing tasks in an active learning style, and validating results through collaboration with other computational tools. Similarly, highly automated robotic factories have the potential to transform process engineering towards miniature factories capable of being installed in laboratories, houses, and so forth for better efficiency and resource management.[199, 362] 6 TO AI OR NOT TO AI? The promise of AI, as reviewed in the preceding sections, is now being tested in many real-life applications. While there are many success stories, there have also been a large number of cases where they have been found wanting. One survey indicated that up to 70% of AI initiatives may yield no or minimal impact.[363] Another survey by SAS, Accenture Applied Intelligence, Intel, and Forbes Insights revealed the need for human oversight, especially when AI is used for automating critical operational decisions. Specifically, it found that one in four enterprises had to reassess, redesign, or override AI-based systems due to unsatisfactory results. Reasons for AI system failures include deviations from intended use (48%), inconsistent outputs (38%), and ethical concerns (34%).[363] These highlight the need for an understanding of the factors that contribute to the success or failure of AI projects. Five key factors can be distinguished as elaborated next. Data quality is a multifaceted concept encompassing attributes such as accuracy, completeness, consistency, relevance, and timeliness.[364, 365] Inaccurate data (for instance, arising from faulty sensor readings) can introduce biases or errors. Incomplete data (such as missing entries or fragmented records) can hinder the model's ability to determine meaningful patterns. Inconsistencies in data (different sampling frequencies or data integration from different sources) can confound AI algorithms, leading to inaccuracies. Irrelevant data (e.g., for a fault diagnosis application, extensive data from normal operations in lieu of data from abnormal conditions) may introduce biases. Additionally, data that does not account for system changes (such as equipment degradation or catalyst deactivation) or changes in environmental conditions can compromise long-term performance in the dynamic settings that are common in the chemical industry. Data quantity: The development of AI models relies on adequate data. Data adequacy must be evaluated with respect to the complexity of the AI model (both structural and parametric). Limited data (faulty data for fault diagnosis or reaction data in retrosynthesis), in general, leads to poorer model performance. To mitigate data limitations, various strategies can be adopted, such as augmenting the data using process simulations, leveraging existing datasets from similar processes, or incorporating insights from domain experts, as discussed in Section 4.1. If, even with such augmentation, the quantity of data is inadequate, then alternate non-AI approaches should be considered. Usability refers to the usage of the AI system's functionalities and interfaces by the end-user, easily and without errors. The complexity of AI models, along with their black-box nature, can result in (often unstated) expectations being placed on the end-user related to the proper usage. If an end-user, such as a control room operator, is not well-versed with the intricacies of the AI model but is responsible for the outcome arising from its usage (overall responsibility for ensuring stable operations), there is a potential for mal-usage. The extent of usage of AI models and the model's complexity may need to be constrained in order to ensure compatibility with the user's capabilities. Also related is the need for regulatory compliance, such as in pharmaceutical manufacturing, of the data used for developing and maintaining AI models.[366] Model explanations discussed in Section 4.2.1 offer one means to address or ameliorate such limitations. The paradox of automation, especially the long-term effects of replacing human decision-making with an AI system, must also be considered when evaluating the benefits. As the AI model gets better, the frequency at which the human end-user would be called upon will decrease. Excessive reliance on AI would, over time, diminish human end-users skills.[367] However, the out-of-practice human end-user would still be required to step in during the rarest of the rare but most complex situations where the AI is itself likelier to fail (sparsity of rare event data in training). A careful evaluation of these considerations is crucial when deciding if high-level AI should not be employed, especially in human-in-the-loop systems. AI economics: Any large-scale project would require economic justification and a positive return on investment (ROI). The development, deployment, and long-term maintenance of AI systems often involve unforeseen expenses associated with additional data collection (see data quality and quantity issues above) and human resources. Training end-users to become savvy users of AI could be essential for project success but can be complicated due to the usability challenges. Further, new personnel responsible for model maintenance may be needed since automated model updates are still in their infancy. These requirements may not be fully evident at the AI project's inception or evolve over its lifecycle and significantly affect the AI project's ROI.[368] Hence, lifecycle cost analysis, accounting for uncertainty and risks, should be used to decide on the attractiveness of AI projects. AUTHOR CONTRIBUTIONS Karthik Srinivasan: Conceptualization; investigation; writing – original draft. Anjana Puliyanda: Writing – original draft; conceptualization; investigation. Devavrat Thosar: Conceptualization; investigation; writing – original draft. Abhijit Bhakte: Conceptualization; investigation; writing – original draft. Kuldeep Singh: Conceptualization; investigation; writing – original draft. Prince Addo: Conceptualization; investigation; writing – original draft. Rajagopalan Srinivasan: Conceptualization; investigation; writing – original draft; writing – review and editing; funding acquisition. Vinay Prasad: Conceptualization; investigation; writing – original draft; writing – review and editing; funding acquisition. OPEN RESEARCH DATA AVAILABILITY STATEMENT Data sharing not applicable to this article as no datasets were generated or analyzed during the current study. REFERENCES * 1V. Venkatasubramanian, AIChE J. 2019, 65, 466. 10.1002/aic.16489 CASWeb of Science®Google Scholar * 2E. Rich, Artificial intelligence, McGraw-Hill, Inc., New York 1983. Google Scholar * 3C. Thon, B. Finke, A. Kwade, C. Schilde, Advanced Intelligent Systems 2021, 3, 2000261. 10.1002/aisy.202000261 Web of Science®Google Scholar * 4D. Dutta, S. R. Upreti, Can. J. Chem. Eng. 2021, 99, 2467. 10.1002/cjce.24246 CASWeb of Science®Google Scholar * 5A. M. Schweidtmann, E. Esche, A. Fischer, M. Kloft, J. U. Repke, S. Sager, A. Mitsos, Chem. Ing. Tech. 2021, 93, 2029. 10.1002/cite.202100083 CASWeb of Science®Google Scholar * 6J. M. Weber, Z. Guo, C. Zhang, A. M. Schweidtmann, A. A. Lapkin, Chem. Soc. Rev. 2021, 50, 12013. 10.1039/D1CS00477H CASPubMedWeb of Science®Google Scholar * 7L. von Rueden, S. Mayer, K. Beckh, B. Georgiev, S. Giesselbach, R. Heese, B. Kirsch, J. Pfrommer, A. Pick, R. Ramamurthy, M. Walczak, J. Garcke, C. Bauckhage, J. Schuecker, IEEE Transactions on Knowledge and Data Engineering 2023, 35, 614. Web of Science®Google Scholar * 8L. S. Keren, A. Liberzon, T. Lazebnik, Sci. Rep. 2023, 13, 1249. 10.1038/s41598-023-28328-2 CASPubMedGoogle Scholar * 9S. L. Brunton, B. W. Brunton, J. L. Proctor, J. N. Kutz, PLoS One 2016, 11, e0150171. 10.1371/journal.pone.0150171 PubMedWeb of Science®Google Scholar * 10J. Willard, X. Jia, S. Xu, M. Steinbach, V. Kumar, ACM Computing Surveys 2022, 55, 1. 10.1145/3514228 Web of Science®Google Scholar * 11D. S. Wigh, J. M. Goodman, A. A. Lapkin, Wiley Interdiscip. Rev: Comput. Mol. Sci. 2022, 12, e1603. 10.1002/wcms.1603 Web of Science®Google Scholar * 12L. David, A. Thakkar, R. Mercado, O. Engkvist, J. Cheminf. 2020, 12, 1. 10.1186/s13321-020-00460-5 PubMedGoogle Scholar * 13K. C. Leonard, F. Hasan, H. F. Sneddon, F. You, ACS Sustainable Chem. Eng. 2021, 9, 6126. 10.1021/acssuschemeng.1c02741 CASWeb of Science®Google Scholar * 14D. Weininger, J. Chem. Inf. Comput. Sci. 1988, 28, 31. 10.1021/ci00057a005 CASWeb of Science®Google Scholar * 15N. M. O'boyle, A. Dalke, 2018, https://doi.org/10.26434/chemrxiv.7097960.v1 10.26434/chemrxiv.7097960.v1 Google Scholar * 16M. Krenn, F. Häse, A. K. Nigam, P. Friederich, A. Aspuru-Guzik, Machine Learning: Science and Technology 2020, 1, 045024. 10.1088/2632-2153/aba947 Google Scholar * 17P. Schwaller, T. Gaudin, D. Lányi, C. Bekas, T. Laino, Chem. Sci. 2018, 9, 6091. 10.1039/C8SC02339E CASPubMedWeb of Science®Google Scholar * 18W. Bort, I. I. Baskin, T. Gimadiev, A. Mukanov, R. Nugmanov, P. Sidorov, G. Marcou, D. Horvath, O. Klimchuk, T. Madzhidov, A. Varnek, Sci. Rep. 2021, 11, 3178. 10.1038/s41598-021-81889-y CASPubMedWeb of Science®Google Scholar * 19K. Lin, Y. Xu, J. Pei, L. Lai, Chem. Sci. 2020, 11, 3355. 10.1039/C9SC03666K CASPubMedWeb of Science®Google Scholar * 20A. A. Lee, Q. Yang, V. Sresht, P. Bolgar, X. Hou, J. L. Klug-Mcleod, C. R. Butler, Chem. Commun. 2019, 55, 12152. 10.1039/C9CC05122H CASPubMedWeb of Science®Google Scholar * 21N. Pillai, A. Dasgupta, S. Sudsakorn, J. Fretland, P. D. Mavroudis, Drug Discovery Today 2022, 27, 2209. 10.1016/j.drudis.2022.03.017 CASPubMedWeb of Science®Google Scholar * 22R. Winter, F. Montanari, F. Noá, D. A. Clevert, Chem. Sci. 2019, 10, 1692. 10.1039/C8SC04175J CASPubMedWeb of Science®Google Scholar * 23R. Gómez-Bombarelli, J. N. Wei, D. Duvenaud, J. M. Hernández-Lobato, B. Sánchez-Lengeling, D. Sheberla, J. Aguilera-Iparraguirre, T. D. Hirzel, R. P. Adams, A. Aspuru-Guzik, ACS Cent. Sci. 2018, 4, 268. 10.1021/acscentsci.7b00572 CASPubMedWeb of Science®Google Scholar * 24L. Krasnov, I. Khokhlov, M. V. Fedorov, S. Sosnin, Sci. Rep. 2021, 11, 1. 10.1038/s41598-021-94082-y PubMedGoogle Scholar * 25K. Rajan, A. Zielesny, C. Steinbeck, J. Cheminf. 2021, 13, 1. 10.1186/s13321-020-00477-w PubMedGoogle Scholar * 26J. M. Goodman, I. Pletnev, P. Thiessen, E. Bolton, S. R. Heller, J. Cheminf. 2021, 13, 1. 10.1186/s13321-021-00517-z PubMedGoogle Scholar * 27H. Kim, J. Na, W. B. Lee, J. Chem. Inf. Model. 2021, 61, 5804. 10.1021/acs.jcim.1c01289 CASPubMedWeb of Science®Google Scholar * 28J. Handsel, B. Matthews, N. J. Knight, S. J. Coles, J. Cheminf. 2021, 13, 1. 10.1186/s13321-021-00535-x PubMedGoogle Scholar * 29L. Yao, M. Yang, J. Song, Z. Yang, H. Sun, H. Shi, X. Liu, X. Ji, Y. Deng, X. Wang, Anal. Chem. 2023, 95, 5393. 10.1021/acs.analchem.2c05817 CASPubMedGoogle Scholar * 30O. Mández-Lucio, B. Baillif, D. A. Clevert, D. Rouquiá, J. Wichard, Nat. Commun. 2020, 11, 1. 10.1038/s41467-019-13993-7 PubMedGoogle Scholar * 31M. Simonovsky, N. Komodakis, in Artificial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neural Networks, Proc., Part I 27, Springer, Berlin 2018, pp. 412–422. Google Scholar * 32O. Prykhodko, S. V. Johansson, P. C. Kotsias, J. Arús-Pous, E. J. Bjerrum, O. Engkvist, H. Chen, J. Cheminf. 2019, 11, 1. 10.1186/s13321-019-0397-9 PubMedGoogle Scholar * 33S. Wang, Y. Guo, Y. Wang, H. Sun, J. Huang, in Proc. of the 10th ACM Int. Conf on Bioinformatics, Computational Biology and Health Informatics, Association for Computing Machinery, Niagra Falls, NY 2019, pp. 429–436. Google Scholar * 34G. A. Pinheiro, J. Mucelini, M. D. Soares, R. C. Prati, J. L. D. Silva, M. G. Quiles, J. Phys. Chem. A 2020, 124, 9854. 10.1021/acs.jpca.0c05969 CASPubMedWeb of Science®Google Scholar * 35G. Grethe, G. Blanke, H. Kraut, J. M. Goodman, J. Cheminf. 2018, 10, 1. 10.1186/s13321-018-0277-8 PubMedGoogle Scholar * 36C. W. Gao, J. W. Allen, W. H. Green, R. H. West, Comput. Phys. Commun. 2016, 203, 212. 10.1016/j.cpc.2016.02.013 CASWeb of Science®Google Scholar * 37S. N. Elliott, K. B. M. Iii, A. V. Copan, M. Keçeli, C. Cavallotti, Y. Georgievskii, H. F. S. Iii, S. J. Klippenstein, Proc. Combust. Inst. 2020, 000, 1. Google Scholar * 38S. Rangarajan, A. Bhan, P. Daoutidis, Comput. Chem. Eng. 2012, 45, 114. 10.1016/j.compchemeng.2012.06.008 CASWeb of Science®Google Scholar * 39P. P. Plehiers, G. B. Marin, C. V. Stevens, K. M. V. Geem, J. Cheminf. 2018, 10, 1. 10.1186/s13321-018-0269-8 PubMedGoogle Scholar * 40C. W. Coley, W. Jin, L. Rogers, T. F. Jamison, T. S. Jaakkola, W. H. Green, R. Barzilay, K. F. Jensen, Chem. Sci. 2019, 10, 370. 10.1039/C8SC04228D CASPubMedWeb of Science®Google Scholar * 41M. H. Segler, M. P. Waller, Chem. – Eur. J. 2017, 23, 5966. 10.1002/chem.201605499 CASPubMedWeb of Science®Google Scholar * 42M. Sacha, M. BłaŻ, P. Byrski, P. Dąbrowski-Tumański, M. Chromiński, R. Loska, P. Włodarczyk-Pruszyński, S. Jastrzębski, J. Chem. Inf. Model. 2021, 61, 3273. 10.1021/acs.jcim.1c00537 CASPubMedWeb of Science®Google Scholar * 43N. De Cao, T. Kipf, ArXiv preprint 2022, arXiv:1805.11973, http://arxiv.org/abs/1805.11973 (accessed: November 2023). Google Scholar * 44A. E. Blanchard, C. Stanley, D. Bhowmik, J. Cheminf. 2021, 13, 1. 10.1186/s13321-021-00494-3 PubMedGoogle Scholar * 45S. H. Hong, S. Ryu, J. Lim, W. Y. Kim, J. Chem. Inf. Model. 2020, 60, 29. 10.1021/acs.jcim.9b00694 CASPubMedWeb of Science®Google Scholar * 46D. A. Clevert, T. Le, R. Winter, F. Montanari, Chem. Sci. 2021, 12, 14174. 10.1039/D1SC01839F CASPubMedWeb of Science®Google Scholar * 47V. Mann, V. Venkatasubramanian, React. Chem. Eng. 2023, 8, 619. 10.1039/D2RE00309K CASGoogle Scholar * 48L. Xue, J. Bajorath, Comb. Chem. High Throughput Screening 2012, 3, 363. 10.2174/1386207003331454 Google Scholar * 49A. Cereto-Massaguá, M. J. Ojeda, C. Valls, M. Mulero, S. Garcia-Vallvá, G. Pujadas, Methods 2015, 71, 58. 10.1016/j.ymeth.2014.08.005 PubMedGoogle Scholar * 50E. F.-D. Gortari, C. R. García-Jacas, K. Martinez-Mayorga, J. L. Medina-Franco, J. Cheminf. 2017, 9, 1. 10.1186/s13321-016-0187-6 PubMedWeb of Science®Google Scholar * 51N. Schneider, D. M. Lowe, R. A. Sayle, G. A. Landrum, J. Chem. Inf. Model. 2015, 55, 39. 10.1021/ci5006614 CASPubMedWeb of Science®Google Scholar * 52J. N. Wei, D. Duvenaud, A. Aspuru-Guzik, ACS Cent. Sci. 2016, 2, 725. 10.1021/acscentsci.6b00219 CASPubMedWeb of Science®Google Scholar * 53K. Z. Myint, L. Wang, Q. Tong, X. Q. Xie, Mol. Pharmaceutics 2012, 9, 2912. 10.1021/mp300237z CASPubMedWeb of Science®Google Scholar * 54A. U. Danishuddin, Drug Discovery Today 2016, 21, 1291. 10.1016/j.drudis.2016.06.013 CASPubMedWeb of Science®Google Scholar * 55Q. Zang, K. Mansouri, A. J. Williams, R. S. Judson, D. G. Allen, W. M. Casey, N. C. Kleinstreuer, J. Chem. Inf. Model. 2017, 57, 36. 10.1021/acs.jcim.6b00625 CASPubMedWeb of Science®Google Scholar * 56A. Ilnicka, G. Schneider, Mol. Inf. 2023, 42, 2300059. 10.1002/minf.202300059 CASWeb of Science®Google Scholar * 57M. Nasser, N. Salim, F. Saeed, S. Basurra, I. Rabiu, H. Hamza, M. A. Alsoufi, Biomolecules 2022, 12, 508. 10.3390/biom12040508 CASPubMedGoogle Scholar * 58E. J. Bjerrum, B. Sattarov, Biomolecules 2018, 8, 131. 10.3390/biom8040131 PubMedWeb of Science®Google Scholar * 59K. T. Schütt, M. Gastegger, A. Tkatchenko, K. R. Müller, R. J. Maurer, Nat. Commun. 2019, 10(1), 5024, https://doi.org/10.1038/s41467-019-12875-2 10.1038/s41467-019-12875-2 CASPubMedWeb of Science®Google Scholar * 60J. Behler, M. Parrinello, Phys. Rev. Lett. 2007, 98, 146401. 10.1103/PhysRevLett.98.146401 CASPubMedWeb of Science®Google Scholar * 61S. Liu, W. Du, Y. Li, Z. Li, Z. Zheng, C. Duan, Z. Ma, O. Yaghi, A. Anandkumar, C. Borgs, J. Chayes, H. Guo, J. Tang, ArXiv preprint 2023, arXiv:2306.09375, https://arxiv.org/abs/2306.09375 (accessed: May 2024). Google Scholar * 62A. D. Smith, P. Dłotko, V. M. Zavala, Comput. Chem. Eng. 2021, 146, 107202. 10.1016/j.compchemeng.2020.107202 CASGoogle Scholar * 63M. R. Wilkinson, U. Martinez-Hernandez, C. C. Wilson, B. Castro-Dominguez, J. Mater. Res. 2022, 37, 2293. 10.1557/s43578-022-00628-9 CASGoogle Scholar * 64S. Zhong, J. Hu, X. Yu, H. Zhang, Chem. Eng. J. 2021, 408, 127998. 10.1016/j.cej.2020.127998 CASGoogle Scholar * 65P. van Gerwen, A. Fabrizio, M. D. Wodrich, C. Corminboeuf, Machine Learning: Science and Technology 2022, 3, 045005. 10.1088/2632-2153/ac8f1a Google Scholar * 66V. Mann, V. Venkatasubramanian, AIChE J. 2021, 67, e17190. 10.1002/aic.17190 CASWeb of Science®Google Scholar * 67A. C. Vaucher, P. Schwaller, J. Geluykens, V. H. Nair, A. Iuliano, T. Laino, Nat. Commun. 2021, 12, 2573. 10.1038/s41467-021-22951-1 CASPubMedWeb of Science®Google Scholar * 68M. Haghighatlari, J. Li, F. Heidar-Zadeh, Y. Liu, X. Guan, T. Head-Gordon, Chem 2020, 6, 1527. 10.1016/j.chempr.2020.05.014 CASPubMedWeb of Science®Google Scholar * 69F. Musil, A. Grisafi, A. P. Bartók, C. Ortner, G. Csányi, M. Ceriotti, Chem. Rev. 2021, 121, 9759. 10.1021/acs.chemrev.1c00021 CASPubMedWeb of Science®Google Scholar * 70J. Damewood, J. Karaguesian, J. R. Lunger, A. R. Tan, M. Xie, J. Peng, R. Gómez-Bombarelli, Annu. Rev. Mater. Res. 2023, 53(1), 399. 10.1146/annurev-matsci-080921-085947 Google Scholar * 71K. M. Jablonka, A. S. Rosen, A. S. Krishnapriyan, B. Smit, ACS Cent. Sci. 2023, 9, 563. 10.1021/acscentsci.2c01177 CASPubMedGoogle Scholar * 72M. Ziatdinov, C. Y. T. Wong, S. V. Kalinin, Machine Learning: Science and Technology 2023, 4, 045033. 10.1088/2632-2153/ad073b Google Scholar * 73J. Rickman, T. Lookman, S. Kalinin, Acta Mater. 2019, 168, 473. 10.1016/j.actamat.2019.01.051 CASWeb of Science®Google Scholar * 74A. Fabrizio, K. R. Briling, C. Corminboeuf, Discovery 2022, 1, 286. CASGoogle Scholar * 75V. Tshitoyan, J. Dagdelen, L. Weston, A. Dunn, Z. Rong, O. Kononova, K. A. Persson, G. Ceder, A. Jain, Nature 2019, 571, 95. 10.1038/s41586-019-1335-8 CASPubMedWeb of Science®Google Scholar * 76N. J. Szymanski, B. Rendy, Y. Fei, R. E. Kumar, T. He, D. Milsted, M. J. McDermott, M. Gallant, E. D. Cubuk, A. Merchant, H. Kim, A. Jain, C. J. Bartel, K. Persson, Y. Zeng, G. Ceder, Nature 2023, 624, 86. 10.1038/s41586-023-06734-w CASPubMedWeb of Science®Google Scholar * 77G. Vogel, E. Hirtreiter, L. S. Balhorn, A. M. Schweidtmann, Optimization and Engineering 2023, 24, 2911. 10.1007/s11081-023-09798-9 Google Scholar * 78C. Zheng, X. Chen, T. Zhang, N. V. Sahinidis, J. J. Siirola, Comput. Chem. Eng. 2022, 159, 107676. 10.1016/j.compchemeng.2022.107676 CASGoogle Scholar * 79T. Zhang, N. V. Sahinidis, J. J. Siirola, AIChE J. 2019, 65, 592. 10.1002/aic.16443 CASWeb of Science®Google Scholar * 80G. Vogel, L. S. Balhorn, A. M. Schweidtmann, Comput. Chem. Eng. 2023, 171, 108162. 10.1016/j.compchemeng.2023.108162 CASGoogle Scholar * 81E. Hirtreiter, L. S. Balhorn, A. M. Schweidtmann, AIChE J. 2023, 70, e18259. 10.1002/aic.18259 Google Scholar * 82M. F. Theisen, K. N. Flores, L. S. Balhorn, A. M. Schweidtmann, Digital Chemical Engineering 2023, 6, 100072. 10.1016/j.dche.2022.100072 Google Scholar * 83E. S. Yu, J. M. Cha, T. Lee, J. Kim, D. Mun, Energies 2019, 12, 4425. 10.3390/en12234425 Google Scholar * 84J. Oeing, W. Welscher, N. Krink, L. Jansen, F. Henke, N. Kockmann, Digital Chemical Engineering 2022, 4, 100038. 10.1016/j.dche.2022.100038 Google Scholar * 85L. Mencarelli, Q. Chen, A. Pagot, I. E. Grossmann, Comput. Chem. Eng. 2020, 136, 106808. 10.1016/j.compchemeng.2020.106808 CASWeb of Science®Google Scholar * 86P. O. Ludl, R. Heese, J. Höller, N. Asprion, M. Bortz, Front. Chem. Sci. Eng. 2022, 16, 183. 10.1007/s11705-021-2073-7 Web of Science®Google Scholar * 87J.-Y. Cheung, G. Stephanopoulos, Comput. Chem. Eng. 1990, 14, 495. 10.1016/0098-1354(90)87023-I CASWeb of Science®Google Scholar * 88X. Yuan, J. Zhou, B. Huang, Y. Wang, C. Yang, W. Gui, IEEE Transactions on Industrial Informatics 2020, 16, 3721. 10.1109/TII.2019.2938890 Web of Science®Google Scholar * 89E. Martinez-Hernandez, Digital Chemical Engineering 2023, 6, 100075. 10.1016/j.dche.2022.100075 Google Scholar * 90S. C. Brandt, J. Morbach, M. Miatidis, M. Theißen, M. Jarke, W. Marquardt, Comput. Chem. Eng. 2008, 32, 320. 10.1016/j.compchemeng.2007.04.013 CASWeb of Science®Google Scholar * 91G. Buchgeher, D. Gabauer, J. Martinez-Gil, L. Ehrlinger, IEEE Access 2021, 9, 55537. 10.1109/ACCESS.2021.3070395 Google Scholar * 92A. Yang, W. Marquardt, Comput. Chem. Eng. 2009, 33, 822. 10.1016/j.compchemeng.2008.11.015 CASWeb of Science®Google Scholar * 93N. Trokanas, L. Koo, F. Cecelja, in 28th European Symposium on Computer Aided Process Engineering, Vol. 43 (Eds: A. Friedl, J. J. Klemeš, S. Radl, P. S. Varbanov, T. Wallek), Elsevier, Amsterdam, The Netherlands 2018, p. 471. 10.1016/B978-0-444-64235-6.50084-X Google Scholar * 94J. Morbach, A. Yang, W. Marquardt, Engineering Applications of Artificial Intelligence 2007, 20, 147. 10.1016/j.engappai.2006.06.010 Web of Science®Google Scholar * 95N. Hamedi, I. Mutlu, F. Rani, L. Urbas, in 33rd European Symposium on Computer Aided Process Engineering, Vol. 52 (Eds: A. C. Kokossis, M. C. Georgiadis, E. Pistikopoulos), Elsevier, Amsterdam, The Netherlands 2023, p. 1687. 10.1016/B978-0-443-15274-0.50268-7 Google Scholar * 96S. Mao, Y. Zhao, J. Chen, B. Wang, Y. Tang, Comput. Chem. Eng. 2020, 143, 107094. 10.1016/j.compchemeng.2020.107094 CASGoogle Scholar * 97X. Zheng, B. Wang, Y. Zhao, S. Mao, Y. Tang, Neurocomputing 2021, 430, 104. 10.1016/j.neucom.2020.10.095 Google Scholar * 98Y. Zhao, B. Zhang, D. Gao, J. Loss Prev. Process Ind. 2022, 76, 104736. 10.1016/j.jlp.2022.104736 CASGoogle Scholar * 99A. Menon, N. B. Krdzavac, M. Kraft, Curr. Opin. Chem. Eng. 2019, 26, 33. 10.1016/j.coche.2019.08.004 Web of Science®Google Scholar * 100A. Pavel, L. A. Saarimäki, L. Möbus, A. Federico, A. Serra, D. Greco, Biotechnol. J. 2022, 20, 4837. CASGoogle Scholar * 101S. Natarajan, K. Ghosh, R. Srinivasan, Comput. Chem. Eng. 2012, 46, 124. 10.1016/j.compchemeng.2012.06.009 CASWeb of Science®Google Scholar * 102X. Feng, Y. Dai, X. Ji, L. Zhou, Y. Dang, Process Saf. Environ. Prot. 2021, 155, 41. 10.1016/j.psep.2021.09.001 CASWeb of Science®Google Scholar * 103Z. Wang, B. Zhang, D. Gao, Computers in Industry 2022, 139, 103647. 10.1016/j.compind.2022.103647 Google Scholar * 104E. Musulin, F. Roda, M. Basualdo, Comput. Chem. Eng. 2013, 59, 164. 10.1016/j.compchemeng.2013.06.009 CASWeb of Science®Google Scholar * 105V. Mann, S. Viswanath, S. Vaidyaraman, J. Balakrishnan, V. Venkatasubramanian, Comput. Chem. Eng. 2023, 179, 108446. 10.1016/j.compchemeng.2023.108446 CASGoogle Scholar * 106V. Venkatasubramanian, C. Zhao, G. Joglekar, A. Jain, L. Hailemariam, P. Suresh, P. Akkisetty, K. Morris, G. Reklaitis, Comput. Chem. Eng. 2006, 30, 1482. 10.1016/j.compchemeng.2006.05.036 CASWeb of Science®Google Scholar * 107M. M. Vegetti, A. Böhm, H. P. Leone, G. P. Henning, Domain Ontologies for Research Data Management in Industry Commons of Materials and Manufacturing, https://epubs.stfc.ac.uk/manifestation/53374896/DL-CONF-2021-002.pdf#page=47 (accessed: January 2024). Google Scholar * 108T. Grubic, I.-S. Fan, Computers in Industry 2010, 61, 776. 10.1016/j.compind.2010.05.006 Web of Science®Google Scholar * 109N. E. Samaridi, N. N. Karanikolas, M. Papoutsidakis, E. C. Papakitsos, C. E. Papakitsos, International Journal of Production Management and Engineering 2023, 11, 89. 10.4995/ijpme.2023.18702 Google Scholar * 110F. Ameri, B. Kulvatunyou, International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, Vol. 59179, American Society of Mechanical Engineers, Anaheim, CA 2019, V001T02A052. Google Scholar * 111S. Dopler, J. Scholz, in Short Paper Proc. of the Spatial Data Science Symposium. 2021 pp. 1–7. Google Scholar * 112J. Rao, S. Gao, M. Miller, A. Morales, in Proc of the 1st ACM SIGSPATIAL International Workshop on Geospatial Knowledge Graphs, GeoKG 22, Association for Computing Machinery, New York, NY. 2022 pp. 17–25. Google Scholar * 113I. E. Grossmann, A. W. Westerberg, AIChE J. 2000, 46, 1700. 10.1002/aic.690460902 CASWeb of Science®Google Scholar * 114E. Örs, R. Schmidt, M. Mighani, M. Shalaby, in 2020 IEEE Int. Conf. on Engineering, Technology and Innovation (ICE/ITMC). 2020 pp. 1–8. Google Scholar * 115M. Sokolov, M. von Stosch, H. Narayanan, F. Feidl, A. Buttá, Curr. Opin. Chem. Eng. 2021, 34, 100715. 10.1016/j.coche.2021.100715 Web of Science®Google Scholar * 116J. H. Lee, J. Shin, M. J. Realff, Comput. Chem. Eng. 2018, 114, 111. 10.1016/j.compchemeng.2017.10.008 CASWeb of Science®Google Scholar * 117D. C. Psichogios, L. H. Ungar, AIChE J. 1992, 38, 1499. 10.1002/aic.690381003 CASWeb of Science®Google Scholar * 118M. A. Kramer, M. L. Thompson, P. M. Bhagat, in 1992 American Control Conf, IEEE. 1992 pp. 475–479. Google Scholar * 119T. A. Johansen, B. A. Foss, in American Control Conference, IEEE. 1992 pp. 3037–3043. Google Scholar * 120M. L. Mavrovouniotis, S. Chang, Comput. Chem. Eng. 1992, 16, 347. 10.1016/0098-1354(92)80053-C CASWeb of Science®Google Scholar * 121H.-T. Su, N. Bhat, P. Minderman, T. McAvoy, Distillation Columns and Batch Processes 1992, 25, 327. Google Scholar * 122J. Sansana, M. N. Joswiak, I. Castillo, Z. Wang, R. Rendall, L. H. Chiang, M. S. Reis, Comput. Chem. Eng. 2021, 151, 1073654. Google Scholar * 123W. Bradley, J. Kim, Z. Kilwein, L. Blakely, M. Eydenberg, J. Jalvin, C. Laird, F. Boukouvala, Comput. Chem. Eng. 2022, 166, 107898. 10.1016/j.compchemeng.2022.107898 CASWeb of Science®Google Scholar * 124N. Sharma, Y. A. Liu, AIChE J 2022, 68, e17609. 10.1002/aic.17609 CASWeb of Science®Google Scholar * 125L. Rajulapati, S. Chinta, B. Shyamala, R. Rengaswamy, AIChE J 2022, 6, e17715. 10.1002/aic.17715 Google Scholar * 126E. Gallup, T. Gallup, K. Powell, Comput. Chem. Eng. 2023, 170, 108111. 10.1016/j.compchemeng.2022.108111 Google Scholar * 127H. Narayanan, M. von Stosch, F. Feidl, M. Sokolov, M. Morbidelli, A. Buttá, Front. Chem. React. Eng. 2023, 5, 1157889. 10.3389/fceng.2023.1157889 Google Scholar * 128M. Agarwal, International Journal of Systems Science 1997, 28, 65. 10.1080/00207729708929364 Web of Science®Google Scholar * 129M. von Stosch, R. Oliveira, J. Peres, S. F. de Azevedo, Comput. Chem. Eng. 2014, 60, 86. 10.1016/j.compchemeng.2013.08.008 CASWeb of Science®Google Scholar * 130K. McBride, E. I. S. Medina, K. Sundmacher, Chem. Ing. Tech. 2020, 92, 842. 10.1002/cite.202000025 CASGoogle Scholar * 131S. Zendehboudi, N. Rezaei, A. Lohi, Appl. Energy 2018, 228, 2539. 10.1016/j.apenergy.2018.06.051 CASWeb of Science®Google Scholar * 132A. M. Schweidtmann, D. Zhang, M. von Stosch, Digital Chemical Engineering 2024, 10, 100136. 10.1016/j.dche.2023.100136 Google Scholar * 133M. von Stosch, R. Oliveira, J. Peres, S. Feyo de Azevedo, Expert Systems with Applications 2011, 38, 10862. 10.1016/j.eswa.2011.02.117 Web of Science®Google Scholar * 134R. F. Nielsen, N. Nazemzadeh, L. W. Sillesen, M. P. Andersson, K. V. Gernaey, S. S. Mansouri, Comput. Chem. Eng. 2020, 140, 106916. 10.1016/j.compchemeng.2020.106916 Google Scholar * 135L. Zhang, M. Pan, S. Quan, Q. Chen, Y. Shi, in 2006 6th World Congress on Intelligent Control and Automation, IEEE. pp. 8319–8323. Google Scholar * 136T.-M. Hwang, H. Oh, Y.-J. Choi, S.-H. Nam, S. Lee, Y.-K. Choung, Desalination 2009, 247, 210. 10.1016/j.desal.2008.12.025 CASWeb of Science®Google Scholar * 137V. Mahalec, Y. Sanchez, Comput. Chem. Eng. 2012, 45, 15. 10.1016/j.compchemeng.2012.05.012 CASWeb of Science®Google Scholar * 138B. Yan, S. Zhao, J. Li, G. Chen, J. Tao, Chem. Eng. J. 2022, 427, 130881. 10.1016/j.cej.2021.130881 Google Scholar * 139D. Ghosh, E. Hermonat, P. Mhaskar, S. Snowling, R. Goel, Ind. Eng. Chem. Res. 2019, 58, 13533. 10.1021/acs.iecr.9b00900 CASWeb of Science®Google Scholar * 140T. Bikmukhametov, J. Jäschke, Comput. Chem. Eng. 2019, 138, 106834. 10.1016/j.compchemeng.2020.106834 Google Scholar * 141H. J. L. van Can, H. te Braake, S. Dubbelman, C. Hellinga, K. Luyben, J. J. Heijnen, AIChE J. 1998, 44, 1071. 10.1002/aic.690440507 Web of Science®Google Scholar * 142H. Narayanan, M. Luna, M. Sokolov, A. Buttá, M. Morbidelli, Ind. Eng. Chem. Res. 2022, 61, 8658. 10.1021/acs.iecr.1c04507 CASWeb of Science®Google Scholar * 143M. Raissi, P. Perdikaris, G. E. Karniadakis, J. Comput. Phys. 2019, 378, 686. 10.1016/j.jcp.2018.10.045 Web of Science®Google Scholar * 144S. Cuomo, V. S. di Cola, F. Giampaolo, G. Rozza, M. Raissi, F. Piccialli, ArXiv preprint 2022 arXiv:2201.05624v4, http://arxiv.org/abs/2201.05624 (accessed: December 2023). Google Scholar * 145G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Nat. Rev. Phys. 2021, 3, 422. 10.1038/s42254-021-00314-5 Google Scholar * 146P. Sharma, W. T. Chung, B. Akoush, M. Ihme, Energies 2023, 16, 2343. 10.3390/en16052343 CASGoogle Scholar * 147Y. Wang, Q. Ren, Int. J. Heat Mass Transfer 2022, 186, 4. Google Scholar * 148R. Patel, S. Bhartiya, R. Gudi, J. Process Control 2023, 128, 103003. 10.1016/j.jprocont.2023.103003 CASWeb of Science®Google Scholar * 149V. V. Santana, M. S. Gama, J. M. Loureiro, A. E. Rodrigues, A. M. Ribeiro, F. W. Tavares, A. G. Barreto, I. B. Nogueira, ChemEngineering 2022, 6, 6020021. Google Scholar * 150S. G. Subraveti, Z. Li, V. Prasad, A. Rajendran, Ind. Eng. Chem. Res. 2023, 62, 5929. 10.1021/acs.iecr.2c04355 CASGoogle Scholar * 151S. G. Subraveti, Z. Li, V. Prasad, A. Rajendran, Ind. Eng. Chem. Res. 2022, 61, 4095. 10.1021/acs.iecr.1c04731 CASGoogle Scholar * 152F. Chen, S. Luo, S. Wang, H. Nasrabadi, Fluid Phase Equilib. 2022, 558, 113423. 10.1016/j.fluid.2022.113423 CASGoogle Scholar * 153H. Chen, C. Batchelor-McAuley, E. Kätelhön, J. Elliott, R. G. Compton, J. Electroanal. Chem. 2022, 925, 116918. 10.1016/j.jelechem.2022.116918 CASWeb of Science®Google Scholar * 154Y. Takehara, Y. Okano, S. Dost, J. Chem. Eng. Jpn. 2023, 56, 2236656. 10.1080/00219592.2023.2236656 Google Scholar * 155A. Merdasi, S. Ebrahimi, X. Yang, R. Kunz, Chem. Eng. Prog. 2023, 193, 109540. 10.1016/j.cep.2023.109540 CASGoogle Scholar * 156V. Bibeau, D. C. Boffito, B. Blais, Chem. Eng. Prog. 2024, 196, 109652. 10.1016/j.cep.2023.109652 Google Scholar * 157S. Liu, H. Wang, J. H. Chen, K. Luo, J. Fan, Combust. Flame. 2024, 260, 113275. 10.1016/j.combustflame.2023.113275 CASGoogle Scholar * 158Z. Zhang, Z. Li, Can. J. Chem. Eng. 2023, 101, 4307. 10.1002/cjce.24922 CASWeb of Science®Google Scholar * 159S. Ren, S. Wu, Q. Weng, Bioresour. Technol. 2023, 369, 128472. 10.1016/j.biortech.2022.128472 CASPubMedGoogle Scholar * 160Y. Ryu, S. Shin, J. J. Liu, W. Lee, J. Na, Computer Aided Chemical Engineering, Vol. 52, Elsevier, Athens, Greece 2023, p. 493. Google Scholar * 161T. Asrav, E. S. Koksal, E. E. Esenboga, A. Cosgun, G. Kusoglu, D. Aydin, E. Aydin, Computer Aided Chemical Engineering, Vol. 52, Elsevier, Athens, Greece, 2023, p. 227. Google Scholar * 162F. Sorourifar, Y. Peng, I. Castillo, L. Bui, J. Venegas, J. A. Paulson, Ind. Eng. Chem. Res. 2023, 62, 15563. 10.1021/acs.iecr.3c01471 CASGoogle Scholar * 163Y. Zheng, Z. Wu, Ind. Eng. Chem. Res. 2023, 62, 2804. 10.1021/acs.iecr.2c03691 CASWeb of Science®Google Scholar * 164T. S. Franklin, L. S. Souza, R. M. Fontes, M. A. Martins, Digital Chemical Engineering, 2022, 5, 100056. 10.1016/j.dche.2022.100056 Google Scholar * 165G. Wu, W. T. G. Yion, K. L. N. Q. Dang, Z. Wu, Chem. Eng. Res. Des. 2023, 192, 556. 10.1016/j.cherd.2023.02.048 CASWeb of Science®Google Scholar * 166A. W. Rogers, I. O. S. Cardenas, E. A. D. Rio-Chanona, D. Zhang, Computer Aided Chemical Engineering, Vol. 52, Elsevier B.V, Athens, Greece 2023, p. 83. Google Scholar * 167Z. Lu, Y. Li, C. He, J. Ren, H. Yu, B. Zhang, Q. Chen, Comput. Chem. Eng. 2024, 180, 108500. 10.1016/j.compchemeng.2023.108500 CASWeb of Science®Google Scholar * 168S. Selvarajan, A. A. Tappe, C. Heiduk, S. Scholl, R. Schenkendorf, Processes 2022, 10, 1764. 10.3390/pr10091764 Google Scholar * 169A. A. Tappe, S. Selvarajan, C. Heiduk, S. Scholl, R. Schenkendorf, Computer Aided Chemical Engineering, Vol. 52, Elsevier, Athens, Greece 2023, p. 837. Google Scholar * 170N. Muralidhar, M. R. Islam, M. Marwah, A. Karpatne, N. Ramakrishnan, in 2018 IEEE International Conference on Big Data (Big Data), IEEE, Seattle, WA 2018, p. 36. 10.1109/BigData.2018.8621955 Google Scholar * 171T. Asrav, E. Aydin, Comput. Chem. Eng. 2023, 173, 108195. 10.1016/j.compchemeng.2023.108195 Google Scholar * 172N. T. Russell, H. H. C. Bakker, Artificial Intelligence in Engineering 1997, 11, 347. 10.1016/S0954-1810(96)00053-2 Google Scholar * 173N. T. Russell, H. H. C. Bakker, R. I. Chaplin, Control Engineering Practice 2000, 8, 49. 10.1016/S0967-0661(99)00123-9 Web of Science®Google Scholar * 174C. Muñz-Ibañez, M. Alfaro-Ponce, I. Chairez, International Journal of Advanced Manufacturing Technology 2019, 104, 1541. 10.1007/s00170-019-04019-z Google Scholar * 175C. Pan, M. Mahmoudabadbozchelou, X. Duan, J. C. Benneyan, S. Jamali, R. M. Erb, J. Colloid Interface Sci. 2022, 611, 29. 10.1016/j.jcis.2021.11.195 CASPubMedGoogle Scholar * 176W. Ji, S. Deng, J. Phys. Chem. A 2021, 125, 1082. 10.1021/acs.jpca.0c09316 CASPubMedWeb of Science®Google Scholar * 177M. S. F. Bangi, K. Kao, J. S. I. Kwon, Chem. Eng. Res. Des. 2022, 179, 415. 10.1016/j.cherd.2022.01.041 CASWeb of Science®Google Scholar * 178R. T. Q. Chen, Y. Rubanova, J. Bettencourt, D. Duvenaud, ArXiv preprint 2018, arXiv:1806.07366v5, http://arxiv.org/abs/1806.07366 (accessed: March 2024). Google Scholar * 179A. Puliyanda, K. Srinivasan, Z. Li, V. Prasad, Engineering Applications of Artificial Intelligence 2023, 125, 106690. 10.1016/j.engappai.2023.106690 Google Scholar * 180N. Muralidhar, J. Bu, Z. Cao, L. He, N. Ramakrishnan, D. Tafti, A. Karpatne, ArXiv preprint 199, arXiv:1911.04240v1, http://arxiv.org/abs/1911.04240 (accessed: March 2024). Google Scholar * 181N. Muralidhar, J. Bu, Z. Cao, L. He, N. Ramakrishnan, D. Tafti, A. Karpatne, Big Data 2020, 8, 431. 10.1089/big.2020.0071 PubMedGoogle Scholar * 182D. Machalek, J. Tuttle, K. Andersson, K. M. Powell, Energy and AI 2022, 9, 100172. 10.1016/j.egyai.2022.100172 Google Scholar * 183X. I. Yang, S. Zafar, J. X. Wang, H. Xiao, Physical Review Fluids 2019, 4, 034602. https://doi.org/10.1103/PhysRevFluids.4.034602 10.1103/PhysRevFluids.4.034602 Google Scholar * 184K. Yamada, B. R. B. Fernandes, A. Kalamkar, J. Jeon, M. Delshad, R. Farajzadeh, K. Sepehrnoori, Fuel 2024, 357, 129670. 10.1016/j.fuel.2023.129670 Google Scholar * 185A. Carranza-Abaid, J. P. Jakobsen, Comput. Chem. Eng. 2022, 163, 107858. 10.1016/j.compchemeng.2022.107858 Google Scholar * 186M. S. Alhajeri, F. Abdullah, Z. Wu, P. D. Christofides, Chem. Eng. Res. Des. 2022, 186, 34. 10.1016/j.cherd.2022.07.035 CASWeb of Science®Google Scholar * 187N. Sitapure, J. S.-I. Kwon, Ind. Eng. Chem. Res. 2023, 62, 21278. 10.1021/acs.iecr.3c02624 CASGoogle Scholar * 188L. D. McClenny, U. M. Braga-Neto, J. Comput. Phys. 2023, 474, 111722. 10.1016/j.jcp.2022.111722 Google Scholar * 189Y. Bengio, A. Courville, P. Vincent, IEEE Transactions on Pattern Analysis and Machine Intelligence 2013, 35, 1798. 10.1109/TPAMI.2013.50 PubMedWeb of Science®Google Scholar * 190J. Heaton, in SoutheastCon 2016, IEEE. 2016 pp. 1–6. Google Scholar * 191J. Lu, F. Gao, Ind. Eng. Chem. Res. 2008, 47, 9508. 10.1021/ie800595a CASWeb of Science®Google Scholar * 192G. Brand Rihm, E. Esche, J.-U. Repke, M. Schueler, C. Nentwich, Chem. Ing. Tech. 2023, 95, 1125. 10.1002/cite.202200228 Google Scholar * 193A. Daw, R. Q. Thomas, C. C. Carey, J. S. Read, A. P. Appling, A. Karpatne, in Proc. of the 2020 SIAM Int. Conf. on Data Mining (SDM). pp. 532–540. Google Scholar * 194M. Shevlin, ACS Med. Chem. Lett. 2017, 8, 601. 10.1021/acsmedchemlett.7b00165 CASPubMedWeb of Science®Google Scholar * 195T. Williams, K. McCullough, J. A. Lauterbach, Chem. Mater. 2020, 32, 157. 10.1021/acs.chemmater.9b03043 CASWeb of Science®Google Scholar * 196M. Abolhasani, E. Kumacheva, Nature Synthesis 2023, 2, 483. 10.1038/s44160-022-00231-0 Google Scholar * 197N. S. Eyke, B. A. Koscher, K. F. Jensen, Trends Chem 2021, 3, 120. 10.1016/j.trechm.2020.12.001 CASGoogle Scholar * 198B. Burger, P. M. Maffettone, V. V. Gusev, C. M. Aitchison, Y. Bai, X. Wang, X. Li, B. M. Alston, B. Li, R. Clowes, N. Rankin, B. Harris, R. S. Sprick, A. I. Cooper, Nature 2020, 583, 237. 10.1038/s41586-020-2442-2 CASPubMedWeb of Science®Google Scholar * 199C. W. Coley, D. A. Thomas, J. A. M. Lummiss, J. N. Jaworski, C. P. Breen, V. Schultz, T. Hart, J. S. Fishman, L. Rogers, H. Gao, R. W. Hicklin, P. P. Plehiers, J. Byington, J. S. Piotti, W. H. Green, A. J. Hart, T. F. Jamison, K. F. Jensen, Science 2019, 365, eaax1566. 10.1126/science.aax1566 CASPubMedWeb of Science®Google Scholar * 200J. M. Granda, L. Donina, V. Dragone, D.-L. Long, L. Cronin, Nature 2018, 559, 377. 10.1038/s41586-018-0307-8 CASPubMedWeb of Science®Google Scholar * 201K. Abdel-Latif, R. Epps, F. Bateni, S. Han, K. Reyes, M. Abolhasani, Advanced Intelligent Systems 2021, 3, 2170022. https://doi.org/10.1002/aisy.202170022 10.1002/aisy.202170022 Google Scholar * 202R. Epps, M. Bowen, A. Volk, K. Abdel-Latif, S. Han, K. Reyes, A. Amassian, M. Abolhasani, Adv. Mater. 2020, 32, 2070222. PubMedGoogle Scholar * 203J. Li, J. Li, R. Liu, Y. Tu, Y. Li, J. Cheng, T. He, X. Zhu, Nat. Commun. 2020, 11, 2046. 10.1038/s41467-020-15728-5 CASPubMedWeb of Science®Google Scholar * 204J. Li, Y. Tu, R. Liu, Y. Lu, X. Zhu, Adv. Sci. 2020, 7, 1901957. 10.1002/advs.201901957 CASGoogle Scholar * 205D. Salley, G. Keenan, J. Grizou, A. Sharma, S. Martín, L. Cronin, Nat. Commun. 2020, 11, 2771. 10.1038/s41467-020-16501-4 CASPubMedWeb of Science®Google Scholar * 206H. Tao, T. Wu, S. Kheiri, M. Aldeghi, A. Aspuru-Guzik, E. Kumacheva, Adv. Funct. Mater. 2021, 31, 2106725. Google Scholar * 207B. MacLeod, F. Parlane, C. Rupnow, K. Dettelbach, M. Elliott, T. Morrissey, T. Haley, O. Proskurin, M. Rooney, N. Taherimakhsousi, D. Dvorak, H. Chiu, C. Waizenegger, K. Ocean, M. Mokhtari, C. Berlinguette, Nat. Commun. 2022, 13, 995. 10.1038/s41467-022-28580-6 CASPubMedWeb of Science®Google Scholar * 208B. P. MacLeod, F. G. L. Parlane, T. D. Morrissey, F. Häse, L. M. Roch, K. E. Dettelbach, R. Moreira, L. P. E. Yunker, M. B. Rooney, J. R. Deeth, V. Lai, G. J. Ng, H. Situ, R. H. Zhang, M. S. Elliott, T. H. Haley, D. J. Dvorak, A. Aspuru-Guzik, J. E. Hein, C. P. Berlinguette, Sci. Adv. 2020, 6, eaaz8867. 10.1126/sciadv.aaz8867 CASPubMedWeb of Science®Google Scholar * 209P. Nikolaev, D. Hooper, N. Perea-López, M. Terrones, B. Maruyama, ACS Nano 2014, 8, 10214. 10.1021/nn503347a CASPubMedWeb of Science®Google Scholar * 210L. Roch, F. Häse, C. Kreisbeck, T. Tamayo-Mendoza, L. Yunker, J. Hein, A. Aspuru-Guzik, PLoS One 2020, 15, e0229862. 10.1371/journal.pone.0229862 CASPubMedWeb of Science®Google Scholar * 211J. Deneault, J. Chang, J. Myung, D. Hooper, A. Armstrong, M. Pitt, B. Maruyama, MRS Bull. 2021, 46, 566. 10.1557/s43577-021-00051-1 Google Scholar * 212F. Häse, L. M. Roch, C. Kreisbeck, A. Aspuru-Guzik, ACS Cent. Sci. 2018, 4, 1134. 10.1021/acscentsci.8b00307 CASPubMedWeb of Science®Google Scholar * 213F. Häse, L. M. Roch, A. Aspuru-Guzik, ArXiv preprint 2020, arXiv: abs/2003.12127, https://api.semanticscholar.org/CorpusID:214693268 (accessed: May 2024). Google Scholar * 214M. Aldeghi, F. Häse, R. J. Hickman, I. Tamblyn, A. Aspuru-Guzik, Chem. Sci. 2021, 12, 14792. 10.1039/D1SC01545A CASPubMedWeb of Science®Google Scholar * 215R. Epps, M. Abolhasani, Applied Physics Reviews 2021, 8, 041316. 10.1063/5.0061799 CASWeb of Science®Google Scholar * 216P. S. Gromski, A. Henson, J. M. Granda, L. Cronin, Chemistry 2019, 3, 119. Google Scholar * 217R. Pollice, G. dos Passos Gomes, M. Aldeghi, R. Hickman, M. Krenn, C. Lavigne, M. Lindner-D'Addario, A. Nigam, C. T. Ser, Z. Yao, A. Aspuru-Guzik, Accounts of Chemical Research 2021, 54, 849. 10.1021/acs.accounts.0c00785 PubMedGoogle Scholar * 218H. Tao, T. Wu, M. Aldeghi, A. Aspuru-Guzik, E. Kumacheva, Nat. Rev. Mater. 2021, 6, 1. 10.1038/s41578-021-00337-5 Google Scholar * 219A. Nicholson, S. Webber, S. Dyer, T. Patel, H. Janicke, Computers & Security 2012, 31, 418. 10.1016/j.cose.2012.02.009 Web of Science®Google Scholar * 220M. Iaiani, A. Tugnoli, S. Bonvicini, V. Cozzani, Reliability Engineering & System Safety 2021, 209, 107485. 10.1016/j.ress.2021.107485 Google Scholar * 221 RISI, The Repository of Industrial Security Incidents. https://www.risidata.com/Database/event_date/desc/P30 (accessed: May 2024). Google Scholar * 222A. Nguyen, J. Yosinski, J. Clune, in Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition, IEEE. 2015 pp. 427–436. Google Scholar * 223R. S. Siva Kumar, M. Nyström, J. Lambert, A. Marshall, M. Goertzel, A. Comissoneru, M. Swann, S. Xia, in 2020 IEEE Security and Privacy Workshops (SPW), IEEE. 2020 pp. 69–75. Google Scholar * 224L. Huang, A. D. Joseph, B. Nelson, B. I. Rubinstein, J. D. Tygar, in Proc. of the 4th ACM Workshop on Security and Artificial Intelligence, AISec 11, Association for Computing Machinery, ACM Press, New York, NY. 2011 pp. 43–58. Google Scholar * 225N. Pitropakis, E. Panaousis, T. Giannetsos, E. Anastasiadis, G. Loukas, Computer Science Review 2019, 34, 100199. 10.1016/j.cosrev.2019.100199 Web of Science®Google Scholar * 226I. Goodfellow, J. Shlens, C. Szegedy, ArXiv preprint 2014, arXiv 1412.6572, https://api.semanticscholar.org/CorpusID:6706414 (accessed: May 2024). Google Scholar * 227N. Carlini, D. A. Wagner, in IEEE Symposium on Security and Privacy (SP), IEEE. 2017 pp. 39–57. Google Scholar * 228N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, A. Swami, in 2016 IEEE European symposium on security and privacy (EuroS&P), IEEE. 2016 pp. 372–387. Google Scholar * 229S.-M. Moosavi-Dezfooli, A. Fawzi, P. Frossard, in Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, IEEE. 2016 pp. 2574–2582. Google Scholar * 230W. Brendel, J. Rauber, M. Bethge, ArXiv preprint 2017, arXiv: 1712.04248, https://api.semanticscholar.org/CorpusID:2410333 (accessed: May 2024). Google Scholar * 231N. Papernot, P. McDaniel, I. Goodfellow, S. Jha, Z. B. Celik, A. Swami, in Proc. of the 2017 ACM on Asia Conf. on Computer and Communications Security. 2017 pp. 506–519. Google Scholar * 232J. Su, D. V. Vargas, K. Sakurai, IEEE Transactions on Evolutionary Computation 2017, 23, 828. 10.1109/TEVC.2019.2890858 Google Scholar * 233O. Ibitoye, R. Abou-Khamis, ArXiv preprint 2019, arXiv:1911.02621, https://api.semanticscholar.org/CorpusID:207848033 (accessed: May 2024). Google Scholar * 234H. Lee, S. Han, J. Lee, ArXiv preprint 2017, arXiv:1705.03387, https://api.semanticscholar.org/CorpusID:6222110 (accessed: May 2024). Google Scholar * 235F. Tramér, A. Kurakin, N. Papernot, D. Boneh, P. Mcdaniel, ArXiv preprint 2017, arXiv:1705.07204, https://api.semanticscholar.org/CorpusID:21946795 (accessed: May 2024). Google Scholar * 236B. Zhou, A. Khosla, ß. Lapedriza, A. Oliva, A. Torralba, ArXiv preprint 2014, arXiv:1412.6856, https://api.semanticscholar.org/CorpusID:8217340 (acessed: May 2024). Google Scholar * 237W. Xu, D. Evans, Y. Qi, ArXiv preprint 2017, arXiv:1704.01155, https://api.semanticscholar.org/CorpusID:3851184 (accessed: May 2024). Google Scholar * 238M. Barreno, B. Nelson, A. D. Joseph, J. D. Tygar, Machine Learning 2010, 81, 121. 10.1007/s10994-010-5188-5 Web of Science®Google Scholar * 239X. Yuan, P. He, Q. Zhu, X. Li, IEEE Transactions on Neural Networks and Learning Systems 2019, 30, 2805. 10.1109/TNNLS.2018.2886017 PubMedWeb of Science®Google Scholar * 240A. Rawal, D. B. Rawat, B. M. Sadler, Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications III 2021, 11746, 701. Google Scholar * 241A. B. Arrieta, N. Díaz-Rodríguez, J. D. Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S. Gil-Lopez, D. Molina, R. Benjamins, R. Chatila, F. Herrera, Information Fusion 2020, 58, 82. 10.1016/j.inffus.2019.12.012 Web of Science®Google Scholar * 242G. Manca, N. Bhattacharya, S. Maczey, D. Ziobro, E. Brorsson, M. Bäng, in Frontiers in Artificial Intelligence and Applications (Eds: P. Lukowicz, S. Mayer, J. Koch, J. Shawe-Taylor, I. Tiddi), IOS Press, Munich, Germany 2023, p. 401. https://doi.org/10.3233/FAIA230110 Google Scholar * 243R. R. A. Harinarayan, S. M. Shalinie, Process Saf. Environ. Prot. 2022, 165, 463. 10.1016/j.psep.2022.07.019 CASGoogle Scholar * 244L. C. Brito, G. A. Susto, J. N. Brito, M. A. V. Duarte, Mechanical Systems and Signal Processing 2022, 163, 108105. 10.1016/j.ymssp.2021.108105 Web of Science®Google Scholar * 245K. Jang, K. E. S. Pilario, N. Lee, I. Moon, J. Na, IEEE Transactions on Industrial Informatics 2023, https://doi.org/10.1109/TII.2023.3240601 10.1109/TII.2023.3240601 PubMedGoogle Scholar * 246A. Bhakte, V. Pakkiriswamy, R. Srinivasan, Chem. Eng. Sci. 2022, 250, 117373. 10.1016/j.ces.2021.117373 CASGoogle Scholar * 247P. Agarwal, M. Tamer, H. Budman, Comput. Chem. Eng. 2021, 154, 107467. 10.1016/j.compchemeng.2021.107467 CASWeb of Science®Google Scholar * 248A. Sivaram, V. Venkatasubramanian, AIChE J. 2022, 68, e17687. 10.1002/aic.17687 CASWeb of Science®Google Scholar * 249H. Gandhi, A. White, Explaining Molecular Properties with Natural Language, https://chemrxiv.org/engage/chemrxiv/article-details/633731d1f764e6e535093041 (accessed: May 2024). Google Scholar * 250K. H. Sun, H. Huh, B. A. Tama, S. Y. Lee, J. H. Jung, S. Lee, IEEE Access 2020, 8, 129169. 10.1109/ACCESS.2020.3009852 Google Scholar * 251A. Bhakte, S. Bairi, R. Srinivasan, in 2021 AIChE Annual Meeting, AIChE. 2021. Google Scholar * 252T. Danesh, R. Ouaret, P. Floquet, S. Negny, Comput. Chem. Eng. 2023, 176, 108306. 10.1016/j.compchemeng.2023.108306 CASGoogle Scholar * 253D. Wu, X. Bi, J. Zhao, Ind. Eng. Chem. Res. 2023, 62, 8350. 10.1021/acs.iecr.3c00206 CASGoogle Scholar * 254A. R. N. Aouichaoui, F. Fan, J. Abildskov, G. Sin, Comput. Chem. Eng. 2023, 176, 108291. 10.1016/j.compchemeng.2023.108291 CASGoogle Scholar * 255P. Schwaller, B. Hoover, J.-L. Reymond, H. Strobelt, T. Laino, Sci. Adv. 2021, 7, eabe4166. 10.1126/sciadv.abe4166 PubMedWeb of Science®Google Scholar * 256A. Bhakte, M. Chakane, R. Srinivasan, Comput. Chem. Eng. 2023, 179, 108442. 10.1016/j.compchemeng.2023.108442 CASGoogle Scholar * 257N. Leveson, Safety Science 2004, 42, 237. 10.1016/S0925-7535(03)00047-X Web of Science®Google Scholar * 258T. V. Shajahan, R. Madbhavi, M. A. Shahab, B. Srinivasan, R. Srinivasan, in Computer Aided Chemical Engineering, Vol. 52 (Eds: A. C. Kokossis, M. C. Georgiadis, E. Pistikopoulos), Elsevier, 2023, p. 2043. Google Scholar * 259R. Arunthavanathan, Z. Sajid, F. Khan, E. Pistikopoulos, Digital Chemical Engineering 2024, 11, 100151. 10.1016/j.dche.2024.100151 Google Scholar * 260N. Jones, Nature 2023, 623, 229. 10.1038/d41586-023-03472-x CASPubMedGoogle Scholar * 261E. J. Bjerrum, ArXiv preprint 2017, arXiv:1703.07076, http://arxiv.org/abs/1703.07076 (accessed: May 2024). Google Scholar * 262M. H. S. Segler, T. Kogej, C. Tyrchan, M. P. Waller, ACS Cent. Sci. 2018, 4, 120. 10.1021/acscentsci.7b00512 CASPubMedWeb of Science®Google Scholar * 263K. Kim, S. Kang, J. Yoo, Y. Kwon, Y. Nam, D. Lee, I. Kim, Y.-S. Choi, Y. Jung, S. Kim, npj Comput. Mater. 2018, 4, 67. 10.1038/s41524-018-0128-1 Web of Science®Google Scholar * 264M. H. Segler, M. Preuss, M. P. Waller, Nature 2018, 555, 604. 10.1038/nature25978 CASPubMedWeb of Science®Google Scholar * 265Z. Hao, C. Lu, Z. Huang, H. Wang, Z. Hu, Q. Liu, E. Chen, C. Lee, in Proc. of the 26th ACM SIGKDD Int. Conf. on Knowledge Discovery & Data Mining, ACM. 2020 pp. 731–752. Google Scholar * 266G. Lamb, B. Paige, ArXiv preprint 2020, arXiv:2012.02089, http://arxiv.org/abs/2012.02089 (accessed: May 2024). Google Scholar * 267Y. Pathak, S. Laghuvarapu, S. Mehta, U. D. Priyakumar, in Proc. of the AAAI Conf. on Artificial Intelligence. 2020 pp. 873–880. Google Scholar * 268M. Blomberg, T. Borowski, F. Himo, R.-Z. Liao, P. Siegbahn, Chem. Rev. 2014, 114, 3601. 10.1021/cr400388t CASPubMedWeb of Science®Google Scholar * 269J. Gilmer, S. Schoenholz, P. Riley, O. Vinyals, G. Dahl, Proceedings of the 34th International Conference on Machine Learning, Proceedings of Machine Learning Research, Sydney, NSW 2017, p. 1263. Google Scholar * 270K. Yang, K. Swanson, W. Jin, C. Coley, P. Eiden, H. Gao, A. Guzman-Perez, T. Hopper, B. Kelley, M. Mathea, J. Chem. Inf. Model. 2019, 59, 3370. 10.1021/acs.jcim.9b00237 CASPubMedWeb of Science®Google Scholar * 271K. Schütt, P.-J. Kindermans, H. E. Sauceda Felix, S. Chmiela, A. Tkatchenko, K.-R. Müller, Advances in Neural Information Processing Systems, 2017, 30. Google Scholar * 272C. Lu, Q. Liu, C. Wang, Z. Huang, P. Lin, L. He, in Proc. of the AAAI Conf. on Artificial Intelligence, Vol. 33. 2019 pp. 1052–1060. Google Scholar * 273D. Flam-Shepherd, T. Wu, A. Aspuru-Guzik, ArXiv preprint 2020, arXiv:2002.07087, https://arxiv.org/abs/2002.07087 (accessed: May 2024). Google Scholar * 274W. Jin, R. Barzilay, T. Jaakkola, International Conference on Machine Learning, PMLR, Stockholm 2018, p. 2323. Google Scholar * 275J. Born, T. Huynh, A. Stroobants, W. D. Cornell, M. Manica, J. Chem. Inf. Model. 2022, 62, 240. 10.1021/acs.jcim.1c00889 CASPubMedWeb of Science®Google Scholar * 276J. Born, M. Manica, J. Cadow, G. Markert, N. A. Mill, M. Filipavicius, N. Janakarajan, A. Cardinale, T. Laino, M. R. Martínez, Machine Learning: Science and Technology 2021, 2, 025024. 10.1088/2632-2153/abe808 Google Scholar * 277J. Born, M. Manica, A. Oskooei, J. Cadow, G. Markert, M. R. Martínez, Iscience, 2021, 24, 102269. 10.1016/j.isci.2021.102269 PubMedGoogle Scholar * 278Y. Pathak, K. S. Juneja, G. Varma, M. Ehara, U. D. Priyakumar, Phys. Chem. Chem. Phys. 2020, 22, 26935. 10.1039/D0CP03508D CASPubMedWeb of Science®Google Scholar * 279A. Kadurin, S. Nikolenko, K. Khrabrov, A. Aliper, A. Zhavoronkov, Mol. Pharmaceutics 2017, 14, 3098. 10.1021/acs.molpharmaceut.7b00346 CASPubMedWeb of Science®Google Scholar * 280E. Putin, A. Asadulaev, Y. Ivanenkov, V. Aladinskiy, B. Sanchez-Lengeling, A. Aspuru-Guzik, A. Zhavoronkov, J. Chem. Inf. Model. 2018, 58, 1194. 10.1021/acs.jcim.7b00690 CASPubMedWeb of Science®Google Scholar * 281A. Zhavoronkov, Y. A. Ivanenkov, A. Aliper, M. S. Veselov, V. A. Aladinskiy, A. V. Aladinskaya, V. A. Terentiev, D. A. Polykovskiy, M. D. Kuznetsov, A. Asadulaev, Nat. Biotechnol. 2019, 37, 1038. 10.1038/s41587-019-0224-x CASPubMedWeb of Science®Google Scholar * 282N. H. Park, M. Manica, J. Born, J. L. Hedrick, T. Erdmann, D. Y. Zubarev, N. Adell-Mill, P. L. Arrechea, Nat. Commun. 2023, 14, 3686. Google Scholar * 283A. Izdebski, E. Weglarz-Tomczak, E. Szczurek, J. M. Tomczak, ArXiv preprint 2023, arXiv:2310.02066, http://arxiv.org/abs/2310.02066 (accessed: May 2024). Google Scholar * 284D. Kong, Y. Huang, J. Xie, Y. N. Wu, ArXiv preprint 2023, arXiv:2310.03253, http://arxiv.org/abs/2310.03253 (accessed: May 2024). Google Scholar * 285J. Ross, B. Belgodere, V. Chenthamarakshan, I. Padhi, Y. Mroueh, P. Das, Nature Machine Intelligence 2022, 4, 1256. 10.1038/s42256-022-00580-7 Google Scholar * 286P. Neves, K. McClure, J. Verhoeven, N. Dyubankova, R. Nugmanov, A. Gedich, S. Menon, Z. Shi, J. K. Wegner, J. Cheminf. 2023, 15, 20. 10.1186/s13321-023-00685-0 PubMedGoogle Scholar * 287P. Schwaller, A. C. Vaucher, T. Laino, J.-L. Reymond, Machine Learning: Science and Technology 2021, 2, 015016. 10.1088/2632-2153/abc81d Google Scholar * 288V. Bagal, R. Aggarwal, P. K. Vinod, U. D. Priyakumar, J. Chem. Inf. Model. 2022, 62, 2064. 10.1021/acs.jcim.1c00600 CASPubMedWeb of Science®Google Scholar * 289S. Honda, S. Shi, R. Hiroki, ArXiv preprint 2019, arXiv:1911.04738, http://arxiv.org/abs/1911.04738 (accessed: May 2024). Google Scholar * 290B. Fabian, T. Edlich, H. Gaspar, M. Segler, J. Meyers, M. Fiscato, M. Ahmed, ArXiv preprint 2020, arXiv:2011.13230, http://arxiv.org/abs/2011.13230 (accessed: May 2024). Google Scholar * 291S. Chithrananda, G. Grand, B. Ramsundar, ArXiv preprint 2020, arXiv:2010.09885, http://arxiv.org/abs/2010.09885 (accessed: May 2024). Google Scholar * 292A. Tripp, E. Daxberger, J. M. Hernández-Lobato, Advances in Neural Information Processing Systems 2020, 33, 11259. Google Scholar * 293J. Noh, J. Kim, H. S. Stein, B. Sanchez-Lengeling, J. M. Gregoire, A. Aspuru-Guzik, Y. Jung, Matter 2019, 1, 1370. 10.1016/j.matt.2019.08.017 Google Scholar * 294B. Kim, S. Lee, J. Kim, Sci. Adv. 2020, 6, eaax9324. 10.1126/sciadv.aax9324 CASPubMedWeb of Science®Google Scholar * 295Z. Yao, B. Sánchez-Lengeling, N. S. Bobbitt, B. J. Bucior, S. G. H. Kumar, S. P. Collins, T. Burns, T. K. Woo, O. K. Farha, R. Q. Snurr, Nature Machine Intelligence 2021, 3, 76. 10.1038/s42256-020-00271-1 Google Scholar * 296S. Dieb, Z. Song, W.-J. Yin, M. Ishii, J. Appl. Phys., 2020, 128, 074901. 10.1063/5.0012351 Google Scholar * 297R. Fisher, The Design of Experiments, 3rd ed., Oliver and Boyd, Edinburgh 1942. Google Scholar * 298D. Finney, Annals of Eugenics 1943, 12, 291. 10.1111/j.1469-1809.1943.tb02333.x Web of Science®Google Scholar * 299G. Box, J. Hunter, W. Hunter, Statistics for Experimenters: Design,Innovation, and Discovery, 2nd ed., Wiley, Hoboken, NJ 2005. Google Scholar * 300J. Cornell, Experiments with Mixtures: Designs, Models, and the Analysis of Mixture Data, 3rd ed., Wiley, Hoboken, NJ 2002. 10.1002/9781118204221 Google Scholar * 301B. Jones, C. Nachtsheim, Journal of Quality Technology 2013, 45, 121. 10.1080/00224065.2013.11917921 PubMedGoogle Scholar * 302K. Jørgensen, T. Naes, J. Chemom. 2004, 18, 45. 10.1002/cem.835 CASWeb of Science®Google Scholar * 303L. Párez-Mosqueda, L. Trujillo-Cayado, F. Carrillo, P. Ramírez, J. Muñoz, Colloids Surf., B 2015, 128, 127. 10.1016/j.colsurfb.2015.02.030 PubMedGoogle Scholar * 304M. Reis, P. Saraiva, in Systems Engineering in the Fourth Industrial Revolution (Eds: R. Kenett, R. Swarz, A. Zonnenshain), Wiley, Hoboken, NJ 2019, p. 137. 10.1002/9781119513957.ch6 Google Scholar * 305E. Castillo, M. Reis, Chemom. Intell. Lab. Syst. 2020, 206, 104121. 10.1016/j.chemolab.2020.104121 CASWeb of Science®Google Scholar * 306P. Schwaller, T. Laino, T. Gaudin, P. Bolgar, C. A. Hunter, C. Bekas, A. A. Lee, ACS Cent. Sci. 2019, 5, 1572. 10.1021/acscentsci.9b00576 CASPubMedWeb of Science®Google Scholar * 307X. Liu, P. Li, S. Song, Decomposing Retrosynthesis into Reactive Center Prediction and Molecule Generation. bioRxiv, 677849, https://www.bioarxiv.org/content/10.1101/677849v2.abstract (accessed: May 2024). Google Scholar * 308P. Schwaller, R. Petraglia, V. Zullo, V. H. Nair, R. A. Haeuselmann, R. Pisoni, C. Bekas, A. Iuliano, T. Laino, Chem. Sci. 2020, 11, 3316. 10.1039/C9SC05704H CASPubMedWeb of Science®Google Scholar * 309S. Zheng, J. Rao, Z. Zhang, J. Xu, Y. Yang, J. Chem. Inf. Model. 2020, 60, 47. 10.1021/acs.jcim.9b00949 CASPubMedWeb of Science®Google Scholar * 310X. Wang, Y. Li, J. Qiu, G. Chen, H. Liu, B. Liao, C.-Y. Hsieh, X. Yao, Chem. Eng. J. 2021, 420, 129845. 10.1016/j.cej.2021.129845 CASWeb of Science®Google Scholar * 311Z. Chen, D. Chen, X. Zhang, Z. Yuan, X. Cheng, IEEE Internet of Things Journal 2021, 9, 9179. 10.1109/JIOT.2021.3100509 Google Scholar * 312K. Mao, X. Xiao, T. Xu, Y. Rong, J. Huang, P. Zhao, Neurocomputing 2021, 457, 193. 10.1016/j.neucom.2021.06.037 Web of Science®Google Scholar * 313J. Li, M. Eastgate, Chem. Eng. 2019, 4, 1595. CASGoogle Scholar * 314M. Nielsen, D. Ahneman, O. Riera, A. Doyle, J. Am. Chem. Soc. 2018, 140, 5004. 10.1021/jacs.8b01523 CASPubMedWeb of Science®Google Scholar * 315N. Angello, V. Rathore, W. Beker, A. Wołos, E. Jira, R. Roszak, T. Wu, C. Schroeder, A. Aspuru-Guzik, B. Grzybowski, Science 2022, 378, 399. 10.1126/science.adc8743 CASPubMedWeb of Science®Google Scholar * 316A. Cadeddu, E. Wylie, J. Jurczak, M. Wampler-Doty, B. Grzybowski, Angew. Chem., Int. Ed. 2014, 53, 8108. 10.1002/anie.201403708 CASPubMedWeb of Science®Google Scholar * 317A. Sato, T. Miyao, K. Funatsu, Mol. Inf. 2021, 41, 2100156. 10.1002/minf.202100156 Web of Science®Google Scholar * 318F. Sandfort, F. Strieth-Kalthoff, M. Kühnemund, C. Beecks, F. Glorius, Chem 2020, 6, 1379. 10.1016/j.chempr.2020.02.017 CASWeb of Science®Google Scholar * 319L. Stops, R. Leenhouts, Q. Gao, A. M. Schweidtmann, ArXiv preprint 2022, arXiv:2207.12051, http://arxiv.org/abs/2207.12051 (accessed: May 2024). Google Scholar * 320Q. Gao, A. M. Schweidtmann, ArXiv preprint 2023, arXiv:2308.07822, http://arxiv.org/abs/2308.07822 (accessed: May 2024). Google Scholar * 321K. Fujiwara, M. Kano, S. Hasebe, A. Takinami, AIChE J. 2009, 55, 1754. 10.1002/aic.11791 CASWeb of Science®Google Scholar * 322C. Shang, X. Huang, J. Suykens, D. Huang, J. Process Control 2015, 28, 17. 10.1016/j.jprocont.2015.02.006 CASWeb of Science®Google Scholar * 323F. Souza, R. Araújo, J. Mendes, Chemom. Intell. Lab. Syst. 2016, 152, 69. 10.1016/j.chemolab.2015.12.011 CASWeb of Science®Google Scholar * 324L. Huang, F. Mao, K. Zhang, Z. Li, Sensors 2022, 22, 841. 10.3390/s22030841 Web of Science®Google Scholar * 325S. Wu, X. Xiao, Q. Ding, P. Zhao, Y. Wei, J. Huang, Advances in Neural Information Processing Systems 2020, 33, 17105. Google Scholar * 326A. Nambiar, C. Breen, T. Hart, T. Kulesza, T. Jamison, K. Jensen, ACS Cent. Sci. 2022, 8, 825. 10.1021/acscentsci.2c00207 CASPubMedWeb of Science®Google Scholar * 327Z. Xiang, B. Xie, R. Fu, M. Qian, Ind. Eng. Chem. Res. 2022, 61, 1531. 10.1021/acs.iecr.1c03883 CASGoogle Scholar * 328K. Liu, Y. Li, J. Yang, Y. Liu, Y. Yao, IEEE Transactions on Instrumentation and Measurement 2020, 69, 8261. 10.1109/TIM.2020.2983595 Web of Science®Google Scholar * 329M. Bauer, J. Cox, M. Caveness, J. Downs, N. Thornhill, IEEE Transactions on Control Systems Technology 2007, 15, 12. 10.1109/TCST.2006.883234 Web of Science®Google Scholar * 330M. Bauer, N. Thornhill, J. Process Control 2008, 18, 707. 10.1016/j.jprocont.2007.11.007 CASWeb of Science®Google Scholar * 331S. Qin, Annual Reviews in Control 2012, 36, 220. 10.1016/j.arcontrol.2012.09.004 Web of Science®Google Scholar * 332L. Chiang, R. Braatz, Chemom. Intell. Lab. Syst. 2003, 65, 159. 10.1016/S0169-7439(02)00140-5 CASWeb of Science®Google Scholar * 333J. Thambirajah, L. Benabbas, M. Bauer, N. Thornhill, Comput. Chem. Eng. 2009, 33, 503. 10.1016/j.compchemeng.2008.10.002 CASWeb of Science®Google Scholar * 334P. Kerkhof, J. Vanlaer, G. Gins, J. Impe, Chem. Eng. Sci. 2013, 104, 285. 10.1016/j.ces.2013.08.007 CASWeb of Science®Google Scholar * 335D. Li, ArXiv preprint 2024, arXiv:1809.04758, http://arxiv.org/abs/1809.04758 (accessed: May 2024). Google Scholar * 336A. Antoniou, ArXiv preprint 2024, arXiv:1711.04340, http://arxiv.org/abs/1711.04340 (accessed: May 2024). Google Scholar * 337G. Pesciullesi, P. Schwaller, T. Laino, J.-L. Reymond, Nat. Commun. 2020, 11, 4874. 10.1038/s41467-020-18671-7 CASPubMedWeb of Science®Google Scholar * 338M. Alberts, T. Laino, A. C. Vaucher, Leveraging Infrared Spectroscopy for Automated Structure Elucidation, https://chemrxiv.org/engage/chemrxiv/article-details/645df5cbf2112b41e96da616 (accessed: May 2024). Google Scholar * 339C. Edwards, T. Lai, K. Ros, G. Honke, K. Cho, H. Ji, ArXiv preprint 2022, arXiv:2204.11817, http://arxiv.org/abs/2204.11817 (accessed May 2024). Google Scholar * 340A. C. Vaucher, F. Zipoli, J. Geluykens, V. H. Nair, P. Schwaller, T. Laino, Nat. Commun. 2020, 11, 3601. 10.1038/s41467-020-17266-6 PubMedWeb of Science®Google Scholar * 341R. OpenAI, ArXiv preprint 2023, Arxiv: 2303.08774, http://arxiv.org/abs/2303.08774 (accessed: May 2024). Google Scholar * 342A. Bender, N. Schneider, M. Segler, W. Patrick Walters, O. Engkvist, T. Rodrigues, Nat. Rev. Chem. 2022, 6, 428. 10.1038/s41570-022-00391-9 PubMedWeb of Science®Google Scholar * 343B. Sanchez-Lengeling, A. Aspuru-Guzik, Science 2018, 361, 360. 10.1126/science.aat2663 CASPubMedWeb of Science®Google Scholar * 344M. Saebi, B. Nan, J. E. Herr, J. Wahlers, Z. Guo, A. M. Zurański, T. Kogej, P.-O. Norrby, A. G. Doyle, N. V. Chawla, Chem. Sci. 2023, 14, 4997. 10.1039/D2SC06041H CASPubMedWeb of Science®Google Scholar * 345T. Rodrigues, Drug Discovery Today: Technol. 2019, 32, 3. 10.1016/j.ddtec.2020.07.001 PubMedGoogle Scholar * 346V. Bagal, R. Aggarwal, P. K. Vinod, U. D. Priyakumar, Liggpt: Molecular generation using a transformer-decoder model, https://chemrxiv.org/engage/chemrxiv/article-details/60c7588e469df48597f456ae (accessed: May 2024) Google Scholar * 347Y. Tian, R. Yuan, D. Xue, Y. Zhou, X. Ding, J. Sun, T. Lookman, J. Appl. Phys., 2020, 128, 014103. Google Scholar * 348R. Vasudevan, G. Pilania, P. V. Balachandran, J. Appl. Phys., 2021, 129, 070401. 10.1063/5.0043300 Google Scholar * 349K. M. Jablonka, P. Schwaller, A. Ortega-Guerrero, B. Smit, Is GPT-3 all you need for low-data discovery in chemistry?, https://chemrxiv.org/engage/chemrxiv/article-details/63eb5a669da0bc6b33e97a35?trk=public_post_comment-text (accessed: May 2024). Google Scholar * 350K. Ahuja, W. H. Green, Y.-P. Li, J. Chem. Theory Comput. 2021, 17, 818. 10.1021/acs.jctc.0c00971 CASPubMedWeb of Science®Google Scholar * 351M. Olivecrona, T. Blaschke, O. Engkvist, H. Chen, J. Cheminf. 2017, 9, 48. 10.1186/s13321-017-0235-x PubMedWeb of Science®Google Scholar * 352G. Simm, R. Pinsler, J. M. Hernández-Lobato, International Conference on Machine Learning, PMLR, 2020, p. 8959. Google Scholar * 353Z. Zhou, S. Kearnes, L. Li, R. N. Zare, P. Riley, Sci. Rep. 2019, 9, 10752. 10.1038/s41598-019-47148-x PubMedWeb of Science®Google Scholar * 354N. Brown, M. Fiscato, M. Segler, A. Vaucher, J. Chem. Inf. Model. 2019, 59, 1096. 10.1021/acs.jcim.8b00839 CASPubMedWeb of Science®Google Scholar * 355J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, Advances in Neural Information Processing Systems 2022, 35, 24824. Google Scholar * 356L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, Advances in Neural Information Processing Systems 2022, 35, 27730. Google Scholar * 357T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, Advances in Neural Information Processing Systems 2020, 33, 1877. Google Scholar * 358V. Sanh, A. Webson, C. Raffel, S. H. Bach, L. Sutawika, Z. Alyafeai, A. Chaffin, A. Stiegler, T. L. Scao, A. Raja, M. Dey, M. S. Bari, C. Xu, U. Thakker, S. S. Sharma, E. Szczechla, T. Kim, G. Chhablani, N. Nayak, D. Datta, J. Chang, M. T.-J. Jiang, H. Wang, M. Manica, S. Shen, Z. X. Yong, H. Pandey, R. Bawden, T. Wang, T. Neeraj, J. Rozen, A. Sharma, A. Santilli, T. Fevry, J. A. Fries, R. Teehan, T. Bers, S. Biderman, L. Gao, T. Wolf, A. M. Rush, ArXiv preprint 2022, arXiv:2110.08207, http://arxiv.org/abs/2110.08207 (accessed: May 2024). Google Scholar * 359J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, S. Lloyd, Nature 2017, 549, 195. 10.1038/nature23474 CASPubMedWeb of Science®Google Scholar * 360V. Dunjko, J. M. Taylor, H. J. Briegel, Phys. Rev. Lett. 2016, 117, 130501. 10.1103/PhysRevLett.117.130501 PubMedWeb of Science®Google Scholar * 361T. Schick, J. Dwivedi-Yu, R. Dessí, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, T. Scialom, ArXiv preprint 2023, arXiv:2302.04761, http://arxiv.org/abs/2302.04761 (accessed: May 2024). Google Scholar * 362L. Crowell, A. Lu, K. Love, A. Stockdale, S. Timmick, D. Wu, Y. Wang, W. Doherty, A. Bonnyman, N. Vecchiarello, C. Goodwine, L. Bradbury, J. Brady, J. Clark, N. Colant, A. Cvetkovic, N. Dalvie, D. Liu, Y. Liu, C. Mascarenhas, C. Matthews, N. Mozdzierz, K. Shah, S.-L. Wu, W. Hancock, R. Braatz, S. Cramer, J. Love, Nat. Biotechnol. 2018, 36, 988. 10.1038/nbt.4262 CASWeb of Science®Google Scholar * 363J. Weiner, Why AI/Data Science Projects Fail: How to Avoid Project Pitfalls, Springer Nature, Cham, Switzerland 2022. Google Scholar * 364L. L. Pipino, Y. W. Lee, R. Y. Wang, Commun. ACM 2002, 45, 211. 10.1145/505248.506010 Google Scholar * 365J. Liu, R. Srinivasan, P. SelvaGuru, Computer Aided Chemical Engineering, Vol. 25, Elsevier, Lyon, France 2008, p. 961. Google Scholar * 366 US FDA, Discussion paper: Artificial intelligence in drug manufacturing, notice; request for information and comments, https://www.federalregister.gov/documents/2023/03/01/2023-04206/discussion-paper-artificial-intelligence-in-drug-manufacturing-notice-request-for-information-and (accessed: April 2024). Google Scholar * 367A. Adhitya, S. F. Cheng, Z. Lee, R. Srinivasan, Comput. Chem. Eng. 2014, 67, 1. 10.1016/j.compchemeng.2014.03.013 CASGoogle Scholar * 368S. Pandey, S. Gupta, S. Chhajed, Int. J. Eng. Res. Sci. Technol. 2021, 10, 749. Google Scholar Early View Online Version of Record before inclusion in an issue * FIGURES * REFERENCES * RELATED * INFORMATION METRICS DETAILS © 2024 The Author(s). The Canadian Journal of Chemical Engineering published by Wiley Periodicals LLC on behalf of Canadian Society for Chemical Engineering. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. * Check for updates RESEARCH FUNDING * Jaffer Professorship in Process Systems and Control Engineering * Natural Sciences and Engineering Research Council of Canada. Grant Number: RGPIN-2019-04600 KEYWORDS * artificial intelligence * human-AI interaction * machine learning * process systems engineering PUBLICATION HISTORY * Version of Record online: 06 November 2024 * Manuscript accepted: 10 September 2024 * Manuscript revised: 04 August 2024 * Manuscript received: 15 April 2024 Close Figure Viewer Previous FigureNext Figure Caption Download PDF back ADDITIONAL LINKS ABOUT WILEY ONLINE LIBRARY * Privacy Policy * Terms of Use * About Cookies * Manage Cookies * Accessibility * Wiley Research DE&I Statement and Publishing Policies * Developing World Access HELP & SUPPORT * Contact Us * Training and Support * DMCA & Reporting Piracy OPPORTUNITIES * Subscription Agents * Advertisers & Corporate Partners CONNECT WITH WILEY * The Wiley Network * Wiley Press Room Copyright © 1999-2024 John Wiley & Sons, Inc or related companies. All rights reserved, including rights for text and data mining and training of artificial intelligence technologies or similar technologies. LOG IN TO WILEY ONLINE LIBRARY Email or Customer ID Password Forgot password? NEW USER > INSTITUTIONAL LOGIN > CHANGE PASSWORD Old Password New Password Too Short Weak Medium Strong Very Strong Too Long YOUR PASSWORD MUST HAVE 10 CHARACTERS OR MORE: * a lower case character, * an upper case character, * a special character * or a digit Too Short PASSWORD CHANGED SUCCESSFULLY Your password has been changed CREATE A NEW ACCOUNT Email Returning user FORGOT YOUR PASSWORD? Enter your email address below. Email Please check your email for instructions on resetting your password. If you do not receive an email within 10 minutes, your email address may not be registered, and you may need to create a new Wiley Online Library account. REQUEST USERNAME Can't sign in? Forgot your username? Enter your email address below and we will send you your username Email Close If the address matches an existing account you will receive an email with instructions to retrieve your username The full text of this article hosted at iucr.org is unavailable due to technical difficulties. Close crossmark popup