PaperlayerView exact revision on arXiv

Source-linked AI summary

Atoms to Processes: The Role of Artificial Intelligence and Machine Learning in Chemical Engineering

Michael Baldea, Linda J. Broadbelt, Marianthi G. Ierapetritou, Akhilesh Jain, Ankur Kumar, Thomas A. Kwan, Fèlix Llovell, Andrew J. Medford, Ilias Mitrai, Joel Paulson, Junyi Qiao, Matthew P. Rivera, Kirti C. Sahu, Lev Sarkisov, Zachary P. Smith, Calvin Tsay, Ching-Mei Wen, Victor M. Zavala, Huacheng Zhang, Dan Zhao

arXiv:2610.02014v1cs.AI

TL;DR

Chemical-engineering AI/ML results can be difficult to compare when computational workflows and materials-informatics benchmarks lack released data splits and model details. This perspective surveys applications and methodological advances, identifying a cross-domain shift toward hybrid and physics-informed models; it concludes that AI/ML amplifies rather than replaces chemical-engineering principles.

  • Problem

    Materials-informatics benchmark results can be difficult to interpret or compare when training/test splits, featurization, tuning, and trained checkpoints are not released.

    Source: §6.4 · p. 31

  • Method

    The paper surveys AI/ML advances across chemical-engineering applications and synthesizes the move toward hybrid and physics-informed frameworks.

    Source: §9 · p. 53

  • Results

    Across application domains, successful AI/ML approaches increasingly embed physical structure, addressing concerns about black-box learning in safety-critical and extrapolative regimes.

    Source: §9 · p. 53

  • Takeaways & Limitations

    AI/ML amplifies chemical engineering when approaches respect physical laws, leverage domain knowledge, and keep human expertise central to decisions.

    Source: §9 · p. 54

    Reproducible comparisons require released computational details, while surrogate-based superstructure optimization still depends on data from rigorous simulations.

    Source: §6.4 · p. 31, §7.1 · p. 36

Abstract

from arXiv · show

The rapid maturation of artificial intelligence (AI) and machine learning (ML) has catalyzed a profound shift in how chemical engineering problems are formulated, analyzed, and solved. Advances in computing, data availability, and learning algorithms have enabled AI/ML methods to impact applications spanning atomic-scale simulations, materials and catalyst discovery, transport and thermodynamics, separations, process systems engineering, and industrial operations. This article provides a perspective on recent methodological developments and representative applications, emphasizing how AI/ML tools are being integrated with first-principles models to address challenges of predictive accuracy, data scarcity, extrapolation, interpretability, and model lifecycle management. Across domains, a unifying trend is the move away from purely black-box approaches toward hybrid and physics-informed frameworks that explicitly respect conservation laws, thermodynamic consistency, and known structural constraints. These approaches not only improve robustness and reliability, but also enable meaningful human-AI collaboration by providing information at an appropriate level of abstraction for the task and decision context. We conclude that AI and ML are not replacing the core principles of chemical engineering; rather, they are amplifying them. As the field advances toward increasingly autonomous, adaptive, and sustainable systems, the thoughtful integration of AI/ML with first-principles understanding and domain expertise will be essential to realizing their full potential across both research and industrial practice.

1 Introduction

Chemical engineering AI/ML adoption has accelerated with advances in computing, data acquisition, and model-training capacity, but effective deployment must address accuracy, physical consistency, interpretability, extrapolation, maintenance, and reproducibility. The article outlines these challenges and argues for minimum standards for code and data availability and computational reproducibility.

1 Introduction

  • GPU-boosted computing, expanded data acquisition and storage, and larger-scale model training have enabled progress across diverse fields, including scientific and engineering applications.High-throughput experimentation and new modeling and computation tools have accelerated data generation; AI/ML have demonstrated utility in catalyst discovery and physicochemical-property prediction.Source: §1 · p. 5
  • Reliable deployment requires accurate predictions and quantified uncertainty, alongside physical consistency, interpretability, extrapolation capacity, and regular model updates to mitigate prediction drift.The examples link prediction errors to consequences in synthesis and product purity, while uncertainty can prompt further validation before synthesis.Source: §1 · p. 5
  • Reproducibility remains a persistent weakness, as missing code, data-processing details, splits, weights, tuning settings, seeds, or software versions impede verification and fair comparison.The authors argue that chemical engineering AI/ML research and publishing venues should move toward explicit minimum standards for code and data availability and computational reproducibility.Source: §1 · pp. 5–6
  • The paper reviews open challenges and promising avenues for applying AI/ML in chemical engineering, following a high-level overview of relevant methodological developments.Source: §1 · p. 6

2 Engineering-Relevant AI/ML Methodology

Chemical engineering AI/ML methods vary in their data use, incorporation of physical knowledge, and balance between interpretability and predictive power. The section surveys task formulation and data representation, supervised and sequential learning, latent representations, and generative and foundation-model approaches.

2 Engineering-Relevant AI/ML Methodology

  • Chemical engineering AI/ML spans methods that differ in data use, physical-knowledge integration, and the balance between interpretability and predictive power.Model development begins by formulating the learning task and representing chemical or engineering data and domain knowledge as mathematical inputs; data reconciliation can preserve balances and operating constraints.Source: §2 · p. 6

2.1 Classical Supervised Learning: Regression, Forecasting, and Classification

  • Supervised methods cover regression, classification, and forecasting, with hybrid and physics-informed models incorporating physical constraints to improve interpretability and extrapolation behavior.Classical and deep models offer different trade-offs; limited data makes overfitting a challenge, while physics can be integrated through model structure, penalties, or features.Source: §2.1 · p. 7

2.2 Active Learning and Reinforcement Learning

  • Active learning selects informative samples for costly data acquisition, whereas reinforcement learning addresses sequential decisions such as control, scheduling, and process optimization.Bayesian optimization balances exploration and exploitation; hybrid RL approaches use simulated or expert data, physical knowledge, constraints, or controllers to address hazardous exploration.Source: §2.2 · p. 8

2.3 Unsupervised Learning: Dimensionality Reduction and Embeddings

  • Unsupervised methods learn latent structure without labeled outputs, supporting process monitoring and molecular or materials representation learning.Variational autoencoders can provide continuous molecular embeddings for inverse design by optimization in latent space followed by decoding to valid structures.Source: §2.3 · p. 8

2.4 Generative Modeling and Foundation Models

  • Generative models sample molecules, materials, or process systems with desired properties, while foundation models adapt large pretrained architectures to downstream chemical-engineering tasks.Tool-augmented systems can couple language models with simulators, optimization routines, databases, and human input to coordinate multi-step workflows.Source: §2.4 · pp. 8–9

3 Kinetics, Catalysis and Reaction Engineering

AI/ML methods are being used across catalysis, from atomic-scale simulation and mechanism generation to kinetic modeling. The reviewed work reports faster, more detailed modeling of active sites and coverage effects, while also identifying accuracy and validation constraints.

3 Kinetics, Catalysis and Reaction Engineering

  • The section reviews ML used with or instead of optimization to extract kinetic equations and rate parameters, and to discover and optimize catalyst formulations.The reviewed applications include computational catalysis and high-throughput computational and experimental screening.Source: §3 · p. 9

3.1.1 Atomic-scale simulations of active sites and surfaces

  • Machine-learned interatomic potentials (MLIPs) enable faster catalytic simulations and broader sampling, supporting more complex active-site models and surface reaction free energies.They achieve errors of approximately 0.2–0.5 eV on adsorption energies at speeds thousands of times faster than DFT, but cannot exceed the accuracy of their underlying electronic-structure data and long-timescale validation is difficult.Source: §3.1.1 · p. 10

3.1.2 Identification of reaction networks and mechanisms

  • Reaction-network tools generate catalytic mechanisms, with RMG-Cat selecting kinetically significant reactions and Genesys-Cat generating gas- and surface-phase mechanisms in parallel.The approaches address complex reaction networks containing thousands of reactions and species; Genesys-Cat was demonstrated on catalytic cracking of iso-octane for three zeolites.Source: §3.1.2 · pp. 10–11
  • Machine-learning rate-coefficient models have outperformed rate-rule approaches for intramolecular H-migration and can identify uncertain and influential reaction rates in complex gas-phase networks.The models can fit complex mechanism-generation workflows, although the cited models were not readily interpreted mechanistically.Source: §3.1.2 · p. 11

3.2 Treatment of coverage effects in kinetic models

  • ML approaches resolve coverage effects at configurational scales impractical for DFT alone, offering a path toward more realistic and predictive kinetic models.Reviewed strategies learn coverage-dependent adsorption energies or corrections, sample high-coverage configurations, and embed neural networks in kinetic Monte Carlo simulations.Source: §3.2 · pp. 11–12

3.3 Integration of AI/ML with multiscale and reactor models

AI/ML surrogates and automated workflows integrate detailed kinetics with reactor-scale models, while data-driven approaches also estimate kinetic parameters from reactor measurements. The cited examples report computationally efficient alternatives or complements to conventional modeling approaches.

3.3 Integration of AI/ML with multiscale and reactor models

  • ML surrogates approximate first-principles microkinetics or rate maps, enabling detailed kinetics to be integrated with reactor-scale mass- and heat-transfer models.Other workflows automate multiscale model construction or explore nanoscale catalyst microstructure effects on reactor performance.Source: §3.3 · p. 12

3.4 Extracting kinetic parameters from measured reactor data

  • For CO2 methanation on a Ni-based catalyst, the best ML regressions outperformed the power law and were comparable to LHHW, particularly on nonisothermal-reactor data.Kinetics-informed neural networks were also used to estimate kinetic parameters from synthetic experimental data by jointly training neural networks and model parameters.Source: §3.4 · p. 13

3.5 Using ML to extract kinetic equations

Machine-learning methods can discover kinetic equations from data, with approaches ranging from sparse identification to integrated estimation of rate laws and kinetic parameters. These methods use domain constraints to restrict learned kinetics to physically plausible models.

3.5 Using ML to extract kinetic equations

  • Lejarza et al. jointly estimate kinetic parameters and identify a governing microkinetic model by eliminating insignificant or redundant elementary steps from candidate mechanisms.Their approach uses experimental or simulated concentration data and can estimate rate constants, activation energies, and pre-exponential factors.Source: §3.5 · p. 13
  • Poly-ODENet discovers rate laws without a known reaction network, including when experimental data are incomplete.The rate laws form the right-hand side of ODEs describing reaction-network kinetics under mass action.Source: §3.5 · p. 13
  • Prabhu and coworkers learn interpretable dynamical models under experimental noise using a derivative-free method that approximates integrals and incorporates mass balance and chemistry information.Source: §3.5 · p. 13
  • Physics-informed kinetic models constrain fitting with mass action, site conservation, thermodynamic consistency, stoichiometry, or known reaction mechanisms.Source: §3.5 · p. 13

3.6 Simultaneous model/experiment optimization of reaction conditions

AI/ML methods support reaction-condition optimization through interpretable modeling, reinforcement learning, and ensemble prediction. Related workflows also address experimental validation and tools for kinetic-model development.

3.6 Simultaneous model/experiment optimization of reaction conditions

  • For CO2 methanation, interpretable ML and multiobjective optimization support reverse design of a low-temperature, highly active catalyst [114].The workflow identifies influential features and integrates the best-performing model with an optimization algorithm.Source: §3.6 · p. 13
  • DDPG agents performed well in optimizing methane partial oxidation for hydrogen production by adjusting temperature, pressure, flow velocity, and substrate composition [115].Source: §3.6 · pp. 13–14
  • An ensemble-learning model accurately predicts coke yield in fluidized catalytic cracking, with a simple structure that is easy to construct [116].Source: §3.6 · p. 14
  • An autonomous electrochemical platform adaptively designs experiments and extracts kinetic information, using negative controls and outliers to discern a molecular electrochemistry mechanism [119].Experiments also serve to validate models or update them through active learning.Source: §3.6 · p. 14
  • Chemfit assesses whether experimental data quality and quantity suffice for reliable kinetic discovery, while the SKAI package uses an LLM to translate reaction descriptions into system equations [123, 124].Source: §3.6 · p. 14

3.7 Catalyst discovery and optimization

AI/ML supports catalyst discovery through rapid screening, formulation optimization, generative design, and autonomous experimentation. These approaches expand candidate exploration, while synthesis accessibility and reproducibility remain important boundaries.

3.7 Catalyst discovery and optimization

  • ML surrogates trained on DFT data screen large catalyst design spaces, while many candidates are validated against DFT rather than through synthesis and testing [66, 83, 125–129].Examples span heterogeneous catalysts, alloys, MOFs, and molecular catalysts; atomistic-model active sites may differ from those of synthesized catalysts.Source: §3.7.1 · pp. 14–15
  • Response-surface models and design-of-experiments map formulation variables to catalyst activity and selectivity while reducing costly synthesis–testing cycles [130–135].Studies include bimetallic compositions, oxide mixtures, and preparation parameters for several catalyst systems.Source: §3.7.1 · p. 15
  • Generative models propose catalyst compositions and structures beyond pre-existing candidates, using VAEs, condition-aware models, and LLMs [136–142].Some approaches generate reaction-tailored formulations or link latent-space directions to chemically meaningful features.Source: §3.7.1 · p. 15
  • Closed-loop robotic platforms have designed, synthesized, and optimized catalysts, including a Rh-based hydroformylation catalyst, and identified a formate-oxidation catalyst within a few iterative loops [143, 144].Related workflows use Bayesian optimization and automated synthesis and testing across other reactions [145–149].Source: §3.7.1 · p. 16
  • Kinetics benchmarks such as Catechol, Summit, and ORDERLY provide shared datasets for reproducible comparison, but data quality and quantity can constrain reliable kinetic discovery.Source: §3.7.1 · p. 17
  • Generated catalyst candidates are not always synthetically accessible, and autonomous platforms remain limited to particular reaction classes and equipment configurations.Source: §3.7.1 · p. 17
  • High-throughput screening, generative models, and self-driving labs accelerate catalyst discovery and optimization.The approaches include VAEs, diffusion models, and LLM-based tools such as CatGPT and CataLM.Source: §3.7.1 · p. 17
  • Code, trained parameters, hyperparameters, and data splits are rarely released, complicating reproduction of rate coefficients and mechanism-generation pipelines.Proposed remedies include publication requirements for code, model weights, and software-environment versions.Source: §3.7.1 · p. 17

4 Thermodynamics, Transport and Fluid Mechanics

AI/ML applications in transport and thermodynamics span complex-flow analysis, physics-informed modeling, and property prediction. Across these areas, hybrid approaches address physical consistency, while data limitations, extrapolation, and deployment remain challenges.

4.1.1 Analysis of complex multiphase phenomena

  • ML analyzes complex multiphase flows, including droplet deformation, breakup, coalescence, evaporation, and transport, using imaging and simulation data [150–165].Neural networks can segment overlapping or deformed droplets and quantify size distributions across breakup regimes.Source: §4.1.1 · p. 18
  • ML models support tracking, reconstruction, and prediction in bubble columns, boiling, spray cooling, microfluidics, and environmental flows [168–173].These applications include bubble dynamics and gas holdup, nucleation sites, droplet generation, and aerosol characterization.Source: §4.1.1 · pp. 18–19

4.1.2 Hybrid modeling: Interfacing ML and physics-based models

  • Physics-informed neural networks can reconstruct velocity and pressure fields from sparse measurements, while ML-accelerated CFD and reinforcement learning support flow simulation and control [167, 174].Applications include atomization, bubble-distribution control, spray cooling, and reactor or separation-unit design.Source: §4.1.2 · p. 19
  • Hybrid and physics-informed models incorporate governing equations and conservation laws to improve robustness, physical consistency, interpretability, and extrapolation in multiphase-flow predictions.Constraints can be embedded in model architectures or loss functions, including in PINNs and hybrid CFD–ML frameworks.Source: §4.1.2 · p. 20, §4.2.1 · p. 22

4.2.1 Thermodynamic and transport property prediction

  • Thermodynamics-informed ML embeds relations such as Gibbs–Helmholtz and Clausius–Clapeyron, improving internal consistency and extrapolation in low-data regimes [184, 185].Neural-network surrogates can represent free energies and derive thermodynamic quantities consistently, while other models embed phase-transition and stability criteria [191, 192].Source: §4.2.1 · pp. 20–21
  • ML interatomic potentials reproduce ab initio accuracy at a fraction of the computational cost, enabling larger-system, longer-timescale thermodynamic calculations [186].Statistical-mechanical ML also accelerates estimation of free energies, phase behavior, and phase diagrams [187].Source: §4.2.1 · p. 21
  • Thermodynamic ML remains constrained by scarce, heterogeneous data, difficult consistency requirements, weak extrapolation, and demands for uncertainty quantification, validation, and maintainability [176, 184–192].Reproducibility is also hindered when trained weights, data splits, and software versions are not shared.Source: §4.2.1 · pp. 21–22
  • ML models predict thermophysical and transport properties, including viscosity, thermal and ionic conductivity, and diffusion coefficients, competitively with or better than semi-empirical correlations [177–183].Source: §4.2.1 · pp. 20, 22

5 Separations

Separations research applies ML across distillation, adsorbents, membranes, and process operations, ranging from materials discovery to control. A recurring direction is combining data-driven models with physical constraints to connect material properties, transport, and process performance.

5.1 Gas-Liquid Contact Operations

  • For complex multicomponent distillation, generative AI designs purity-constrained columns (197); ML also optimizes energy use (198, 199) and predicts NOx operation and rotating-contactor hydrodynamics (200, 201).The generative column-design work reports novel configurations, purity constraints, and efficiency gains.Source: §5.1 · p. 23

5.2.1 Adsorbents

  • CoRE MOF’s experimental structures support gas-separation ML (205, 206, 207–213), while generative approaches design new MOFs, including candidates competitive for CO2 separations (216–219).Generative approaches include diffusion-model, variational-autoencoder, and GPT-4-based pipelines.Source: §5.2.1 · pp. 23–24

5.2.2 Membranes

  • Membrane ML spans polymer properties, mixed-matrix screening, and ion transport; Li/Mg extraction demands sub-nanometer discrimination, and homogenized experimental and simulation data can support quality models on smaller datasets.Polymer models predict properties including glass-transition temperature, dielectric constant, gas permeability, diffusion coefficients, and solubility coefficients.Source: §5.2.2 · p. 24

5.3 Module and Process Modeling

  • PANACHE adds governing transport equations and chromatography to MAPLE’s VSA-cycle emulation, achieving 100× faster computation than traditional solvers while retaining physics fidelity.MAPLE optimized a vacuum swing adsorption cycle for CO2/N2 separation from single-component isotherms.Source: §5.3 · p. 25

5.3.1 Multiscale Modeling and Design

  • Physics-constrained ML predicts osmotic pressures, activity coefficients, and hydration enthalpies across ionic-strength and mixed-electrolyte ranges, frequently outperforming empirical correlations (244–252).Hybrid thermodynamic–AI frameworks couple energetic terms with transport kinetics for more predictive membrane descriptions and real-time optimization under variable conditions.Source: §5.3.1 · p. 25
  • Multiscale adsorption models rapidly screened over 2,500 sorbents for dilute CO2 capture, while random-forest predictions across hundreds of thousands of MOFs informed VSA candidate selection (253, 254).The first study reduced its framework to simple algebraic equations; the second combined MOF adsorption predictions with a VSA model.Source: §5.3.1 · pp. 25–26

5.3.2 System Operations

  • AI is proposed for adaptive control, inverse design, and uncovering difficult-to-capture interfacial behavior; materials-to-process integration remains partial, and hypothetical adsorbents may not be synthetically accessible.Reported barriers also include incomplete release of code, data, model parameters, and hyperparameters, complicating independent verification of speed and accuracy claims.Source: §5.3.2 · pp. 26–27

6 Materials and Interfaces, and Applied Chemistry

This section surveys how AI/ML is changing materials discovery through property prediction, generative design, universal potentials, and autonomous experimentation. It also identifies deployment barriers and research priorities for connecting materials models to practical industrial use.

6.1 Benchmark Datasets and Computational Infrastructure

  • Matbench and OC20/OC22 provide benchmark resources; cloud workflows screened over 32 million candidate structures and identified 18 promising solid electrolytes.OC20 contains over 1.2 million DFT relaxations; OC20/OC22 models reach adsorption-energy mean absolute errors approaching 0.2 eV, sufficient for preliminary screening but not quantitative kinetic modeling. The reported cloud workflow used approximately 46 seconds per structure-relaxation step across 38,000 node-hours.Source: §6.1 · p. 29

6.2 Universal Machine Learning Potentials

  • Universal potentials extend across chemical space: M3GNet reports formation-energy errors below 30 meV/atom, while GNoME reported 2.2 million stable structures, including experimentally validated novel phases.Equivariance constraints support high accuracy with limited training data and efficient nanosecond-scale simulations.Source: §6.2 · pp. 29–30

6.3 Autonomous Discovery Platforms

  • Self-driving laboratories combine automated experimentation with ML-guided optimization, and industrial deployments have shortened materials-development and testing cycles.A mobile robotic chemist conducted more than 688 experiments over eight days while adapting its sampling strategy; CATL reduced development cycles by eight months, and Tesla compressed characterization from 70 to 6 hours per sample.Source: §6.3 · p. 30

6.4 Implementation and Deployment Barriers

  • Deployment remains constrained by non-equilibrium behavior, out-of-distribution accuracy degradation, heterogeneous datasets, and limited release of reproducibility materials.Reported accuracy can degrade for underrepresented elements or unusual coordination environments; metadata inconsistencies, missing uncertainty estimates, and incompatible formats impede cross-dataset training. Unreleased splits, featurization, tuning, and checkpoints also make comparisons difficult to interpret.Source: §6.4 · pp. 30–31

6.5 Transfer-Based MOF models

  • MOF Transformers support transfer across properties and operating conditions, but hypothetical databases can yield synthetically inaccessible candidates and performance degrades for dynamic multicomponent conditions.Models integrate structural representations and conditions such as temperature, pressure, and gas type; current models also struggle to distinguish thermodynamic drivers of competitive gas separation when kinetic adsorption rates are identical.Source: §6.5 · pp. 31–32

6.6 Future Trajectories and Research Imperatives

  • Future priorities include linking materials properties to process performance, incorporating lifecycle sustainability into design, and rigorously evaluating generalization and reproducibility.The section also describes LLM interfaces and an ammonia-catalyst workflow integrating literature mining with Bayesian optimization, while noting the need to validate LLM-generated scientific claims.Source: §6.5 · p. 32, §6.6 · pp. 33, 35

7.1 Process and Product Design

Process and product design uses AI/ML to explore alternatives and support engineers in navigating trade-offs, rather than aiming for full design automation. The section discusses generative design, Bayesian optimization, latent representations, and integration with established optimization tools.

7.1 Process and Product Design

  • Process representations that capture topology, phenomena, and properties can support databases and generative models, but extrapolation beyond training data and encoding expert intuition remain open challenges.The paper compares process representations to the language-like representations used by large language models.Source: §7.1 · p. 35
  • Bayesian optimization can jointly optimize process topology and operating conditions while combining experimental data with physical knowledge, but discrete decisions and high-dimensional spaces remain challenging.Its requirement for an explicitly defined design space distinguishes it from generative AI/ML approaches.Source: §7.1 · p. 35
  • Probabilistic latent representations, such as those learned by variational autoencoders, turn complex discrete inputs into continuous structured spaces suitable for gradient-based optimization.Graph convolutional networks and transformer models also map discrete inputs into continuous feature spaces.Source: §7.1 · p. 36
  • Interactive dimensionality-reduction tools could help engineers visualize design families, compare alternatives, and navigate trade-offs beyond a single optimal solution.The paper identifies analogous tools for process structure spaces as a next step after examples in molecular design.Source: §7.1 · p. 36
  • Superstructure optimization encodes design alternatives and uses mixed-integer programming, but detailed process nonlinearities challenge these methods and surrogate construction requires rigorous simulation data.Streamlining simulation, surrogate modeling, and design-optimization workflows remains an open need.Source: §7.1 · p. 36
  • AI/ML design tools should help engineers pose better questions, explore alternatives systematically, and quantify trade-offs and uncertainty; full design automation is unlikely and not the right goal.The section frames these tools as support for engineering judgment, not a replacement for it.Source: §7.1 · p. 36

7.2 Process Control

AI/ML methods support process control by learning dynamic models and controllers, tuning control parameters, accelerating optimization solvers, and identifying control-problem structure. Reported work shows faster MPC computation without compromising solution quality or stability guarantees, while identifying nonlinear MPC as an open direction.

7.2.1 Learning dynamic models

  • When process-model structure is unknown, neural networks or Gaussian processes can learn approximate dynamics, while sparse regression and control-relevant structures target interpretability and properties such as stability.Examples of learned structures include dissipativity, Lyapunov functions, and Koopman operators.Source: §7.2.1 · p. 37

7.2.2 Learning and tuning controllers

  • Controllers can be learned through reinforcement-learning policies, designed with neural-network parameterizations when a process model is known, or approximated from data using supervised learning.Supervised approaches approximate MPC laws or solution operators to avoid solving nonlinear optimization problems online; the same section also covers observer learning and controller-parameter tuning.Source: §7.2.2 · p. 37

7.2.3 Accelerating computation within MPC

  • ML can tune MPC solver hyperparameters, learn warm starts, select algorithms or control architectures, and guide Benders decomposition decisions.These methods learn aspects of the numerical solution process itself, rather than only approximating a controller or process model.Source: §7.2.3 · p. 37
  • AI can accelerate MPC without compromising solution quality or stability guarantees, although extending ML-based solver tuning to continuous nonlinear MPC remains an open direction.Existing work focuses predominantly on mixed-integer MPC.Source: §7.2.3 · p. 38

7.2.4 Learning the structure of optimal control problems

  • Network and statistical methods represent control or optimization problems as graphs to uncover structural interactions, but current approaches omit functional interactions that could improve decomposition.This limitation matters particularly when solution time is the primary performance criterion.Source: §7.2.4 · p. 38

7.3 Supply Chain Planning, Scheduling, and Management

The section surveys AI/ML applications across supply-chain forecasting, scheduling, logistics, sustainability, and decision support. These approaches support adaptive planning and operations, while adoption depends on data, infrastructure, safety, and lifecycle-management conditions.

7.3.1 Data-Driven Forecasting for Inventory, Procurement, and Demand Management

  • Demand, inventory, and procurement forecasting methods often outperform baseline time-series approaches, and upstream shipment histories and production delays can improve pharmaceutical forecasting accuracy.Anticipating shortages and disruptions before they propagate downstream can support proactive risk mitigation and sourcing decisions.Source: §7.3.1 · p. 39

7.3.2 Dynamic Scheduling, Maintenance Planning, and Operational Optimization

  • Deep RL has been used to generate online production schedules under stochastic demand and production-condition losses, outperforming MILP schedulers with reasonable computational costs.Surrogate models also support rapid schedule evaluation, while deep RL has been explored for dynamic inventory policies and predictive maintenance.Source: §7.3.2 · pp. 39–40, §7.3.1 · p. 39

7.3.3 Transportation, Routing, and Distribution Optimization

  • Digital supply-chain twins combined with data-driven methods support monitoring, simulation, and reconfiguration after disruptions; retail and parcel studies report 10–15% transportation fuel reductions with AI/ML-enhanced routing.RL has also been applied to dynamic routing and multi-agent coordination under uncertainty.Source: §7.3.3 · pp. 40–41

7.3.4 Sustainable and Policy-Aware Supply Chain Design

  • Policy-aware frameworks integrate regulatory constraints and sustainability metrics into prediction, optimization, and control for sourcing, production, transportation, and inventory decisions.Reinforcement learning has been explored for adaptive responses to carbon trading and emissions caps while maintaining economic performance.Source: §7.3.4 · p. 41

7.3.5 Emerging AI Methods for Supply Chain Transformation

  • Emerging supply-chain approaches combine forecasts with optimization, digital twins, generative scenarios, reinforcement learning, and blockchain-enabled data governance to support adaptive and resilient systems.The section describes closed-loop decision support that updates models as new operational data arrive.Source: §7.3.5 · pp. 41–42
  • Deployment is constrained by data costs, privacy and sharing limits, organizational resistance, and limited high-quality data; AI lifecycle management requires validation, uncertainty management, monitoring, and periodic updating.The paper calls for supply chains to use continuously updated AI models as components of integrated optimization and control systems, rather than as standalone predictors.Source: §7.3.5 · pp. 42–43

8 Industrial Applications of AI

This section surveys the broadening use of AI/ML in process industries, from control and monitoring to hybrid plant models, digital twins, optimization, conversational tools, and autonomous operations. It also examines deployment barriers and the role of human oversight and first-principles capabilities in future autonomy.

8.1 Historical Evolution and Current State of AI/ML in Industrial Operations

  • Industrial AI has evolved from first-principles expert systems toward data-driven tools, with cloud computing, open-source frameworks, and self-service platforms lowering adoption barriers.High-fidelity models such as CFD, finite element analysis, and kinetic modeling remain useful for complex what-if scenarios but take time and specialized expertise to develop; AI/ML penetration across operational domains remains uneven.Source: §8 · p. 43, §8.1 · p. 44

8.1.1 Current Mainstream Applications of AI/ML in Industrial Operations

  • AI/ML supports nonlinear control, monitoring, predictive maintenance, virtual sensing, and real-time process optimization across chemical plants.Neural networks and reinforcement learning extend control beyond traditional MPC; PCA, PLS, and autoencoders detect faults, while models predict equipment degradation and remaining useful life. Virtual sensors infer hard-to-measure variables, and optimization surrogates adapt operating strategies to changing conditions; Shell and Marathon Petroleum reported substantial reductions in unplanned outages using ML-based predictive maintenance.Source: §8.1.1 · pp. 44–45

8.1.2 Persona-based Applications

  • Industrial AI should present role-appropriate information: actionable, explainable alerts for operators, deeper improvement analyses for engineers, and strategic summaries for managers.Operators need immediate explanations and possible interventions; engineers need root-cause analysis, plant modeling, and historical knowledge for proactive improvement over hours to weeks; managers need concise performance, production, and risk insights. These needs favor layered, persona-specific interfaces over a monolithic system.Source: §8.1.2 · pp. 45–46

8.1.3 Recent Trends

  • Digital twins and hybrid models are advancing plant-scale applications, although comprehensive real-time twins remain difficult to deploy for complex processes.Chevron410 described plant replicas for real-time health assessment; such twins combine first-principles equations with AI/ML. Hybrid approaches have improved extrapolation and transfer learning with limited new data, and a neural-network/kinetic-model hybrid improved accuracy and robustness in a greater-than-100,000-gallon fermenter relative to the mechanistic model alone430.Source: §8.1.3 · pp. 46–47

8.2 Technical and Organizational Challenges to Widespread Adoption

  • Industrial AI adoption is constrained by opaque recommendations, fragmented tools, model drift, and training data that poorly cover operating conditions needed for optimization.Operators may distrust black-box systems or resist additional dashboards, while siloed platforms prevent insights from informing other applications. Process changes can degrade models, and normal closed-loop data often lack the operating-space variation needed for optimization; lifecycle frameworks use monitoring, drift detection, retraining triggers, revalidation, uncertainty, and governance.Source: §8.2 · pp. 47–49

8.3 Large Language Models: A Paradigm Shift in Human-AI Interaction in Process Industry

  • LLMs could connect operators and engineers to analytics through natural-language queries and plant-specific knowledge, but safety-critical use requires verified tools and human oversight.Conversational interfaces can replace static dashboard push with decision-context-driven queries and help retain tacit operational knowledge. Because LLMs can hallucinate and lack guarantees of numerical or physical consistency, the paper recommends using them primarily to orchestrate verified tools, with human checks before safety-critical action.Source: §8.3 · pp. 49–50

8.4 Future Trajectory: Towards Increasingly Autonomous Operations

  • Unified enterprise-wide analytics platforms could replace fragmented solutions with standardized data access and interoperable components that coordinate plant decisions.Such integration requires technical data-exchange standards and architectures with loosely coupled systems and well-defined interfaces.Source: §8.4 · p. 51
  • Reinforcement learning may enable adaptive control, but safe exploration, sparse data, stability, and simulation-to-plant mismatch limit direct deployment.At a Linde air separation plant, an RL controller demonstrated stable control behavior with superior control quality for over 20 days440. Safe and offline RL, safety layers, and RL-assisted MPC constrain actions; RL is presented as an addition to established MPC, not its replacement in safety-critical processes.Source: §8.4 · p. 51
  • Unified analytics and hybrid digital twins could enable plant-wide optimization across operating modes, demand shifts, and equipment aging, and manage well-characterized transitions autonomously under human supervision.Different MPC and RTO models can lead to suboptimal operations, especially during fast transitions; humans would supervise autonomous regimes and handle off-normal situations.Source: §8.4 · pp. 52–53
  • Plant-scale hybrid digital twins already support real-time health monitoring, and process-industry case studies report quantified operational benefits.The passages do not specify the benefit values.Source: §8.4 · p. 53
  • Narrow, closed-loop training data limit applicability to broader optimization and control tasks, while comprehensive high-fidelity digital twins remain aspirational for many processes.Source: §8.4 · p. 53
  • LLM-based copilots are lowering barriers to plant-analytics interaction and knowledge transfer, but hallucinations and absent numerical or physical guarantees pose safety risks.The passages recommend limiting LLMs to orchestrating verified tools and using human review before safety-critical actions.Source: §8.4 · p. 53
  • Proprietary data and intellectual-property concerns limit release of industrial AI studies’ code, models, and processing pipelines, hindering independent verification and reproducibility.Source: §8.4 · p. 53

9 Conclusions

The review finds AI/ML’s impact across chemical engineering increasingly measurable and emphasizes a shift toward hybrid, physics-informed methods. It concludes that integrating these tools with first-principles understanding and human expertise amplifies rather than replaces chemical engineering principles.

9 Conclusions

  • AI/ML’s impact across chemical engineering research and practice is measurable, growing, and increasingly indispensable, spanning atomistic to industrial scales.The review surveys advances from atomistic simulations and materials discovery through process systems engineering and industrial operations.Source: §9 · p. 53
  • Across application domains, hybrid and physics-informed frameworks embed physical structure to improve predictive accuracy, interpretability, extrapolation, and safety.Examples include physics-informed neural networks, hybrid mechanistic–data-driven models, thermodynamically consistent learning architectures, and structure-preserving control models.Source: §9 · p. 53
  • At molecular and materials scales, learned potentials, graph neural networks, and generative models expand design spaces and support materials screening, inverse design, and autonomous discovery.The review also highlights symmetry-aware and equivariant architectures, alongside benchmark datasets and standardized evaluation protocols that support data efficiency, physical fidelity, and reproducibility.Source: §9 · pp. 53–54
  • At process-unit and plant scales, AI/ML supports multiscale simulation, optimization, control, adaptive experimentation, and closed-loop operation.Bayesian optimization, reinforcement learning, and generative models offer approaches to design, scheduling, and supply-chain management under uncertainty; first-principles constraints help retain feasibility and interpretability.Source: §9 · p. 54
  • Industrial applications have delivered tangible value, but last-mile adoption, model lifecycle management, trust, and organizational integration remain challenges.LLMs may reduce cognitive load and dashboard fatigue, but industrial use requires factual grounding, uncertainty awareness, and safe operation.Source: §9 · p. 54
  • The review calls for deeper physical integration, stronger uncertainty and out-of-distribution methods, multiscale frameworks, interoperable platforms, and enforced reproducibility standards.It also recommends anonymized benchmarks and reporting norms that allow disclosure of methods and reproducibility artifacts without exposing proprietary plant data.Source: §9 · pp. 53–54
  • AI/ML should amplify chemical engineering principles by respecting physical laws, leveraging domain knowledge, and keeping human expertise central to context-appropriate decisions.The review identifies this integration as essential to realizing AI/ML’s potential across research and industrial practice.Source: §9 · p. 54

Related papers

Loading 2610.02014v1…