cs.AIOct 1, 2026
SaveAtoms to Processes: The Role of Artificial Intelligence and Machine Learning in Chemical Engineering
Organizations: McKetta Department of Chemical Engineering and Oden Institute for Chemical Engineering and Sciences, The University of Texas at Austin, Austin, TX 78731, USA · Department of Chemical and Biological Engineering, Northwestern University, Evanston, IL 60208, USA · Department of Chemical and Biomolecular Engineering, University of Delaware, Newark, DE 19716, USA · Advanced Analytics and AI, Baker Hughes · Smart Operations, Global AI, Linde plc, Tonawanda, NY, 14150, USA · Schneider Electric Research Institute, Schneider Electric, Boston, MA 02108, USA · Institute for Global Sustainability, Boston University, Boston, MA 02215, USA · Department of Chemical Engineering, ETSEQ, Universitat Rovira i Virgili, 43007 Tarragona, Spain · School of Chemical and Biomolecular Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA · Department of Chemical Engineering, Indian Institute of Technology Hyderabad, Kandi - 502 204, Sangareddy, Telangana, India · Department of Chemical and Biological Engineering, University of Wisconsin, Madison, WI 53706, USA · Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore 117585 · Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Chemical Engineering, The University of Manchester, Manchester M13 9PL, United Kingdom · Department of Computing, Imperial College London, London SW7 2AZ, United Kingdom · Department of Chemical and Biological Engineering, Monash University, Clayton, Victoria 3800, Australia
Abstract
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.
Figures & tables
Figure 1: AI/ML methodological classes relevant to chemical engineering applications
| Current capabilities | Remaining barriers | Future opportunities |
|---|---|---|
| MLIPs reproduce DFT-level adsorption energetics ( 0.2–0.5 eV) at orders-of-magnitude lower cost, enabling active-site, nanoparticle- and micron-scale surface simulations | MLIP accuracy is bounded by the underlying DFT training data; long-timescale trajectories are difficult to validate, and UQ outside the training domain remains immature | Tighter active-learning loops coupling MLIPs with DFT, and systematic UQ benchmarks for long-timescale/ out-of-domain predictions |
| AI-based mechanism generators (e.g., RMG-Cat, Genesys-Cat) and diffusion-based adsorbate-placement models automate construction of catalytic reaction networks | Estimating rate coefficients for every elementary step, adsorbate configuration, and facet remains combinatorially difficult; many ML rate models are not mechanistically interpretable | Broader use of physics-constrained (mass-action, site-conservation, thermodynamically consistent) kinetic ML to reduce reliance on empirical rate rules |
| ML surrogates for coverage-dependent kinetics and microkinetic/reactor coupling (e.g., lattice CNNs, AMUSE) enable reactor-scale simulation with retained mechanistic fidelity | Validation of learned kinetic models still relies heavily on comparison to DFT rather than synthesis and testing of real catalysts | Extending multiscale, DFT-to-reactor automation frameworks to routine industrial reactor design workflows |
| ML directly extracts rate coefficients and full rate laws from reactor/experimental data (SINDy-type and kinetics-informed neural networks), and LLM-assisted tools (e.g., SKAI) translate reaction descriptions into model equations | Data quality/quantity requirements for reliable kinetic discovery are often unmet in typical experimental datasets | Development of shared kinetics data benchmarks (e.g., Catechol, Summit, ORDERLY) to enable reproducible comparison across methods |
| High-throughput ML screening, generative (VAE, diffusion) and LLM-based (CatGPT, CataLM) models, and autonomous “self-driving lab” platforms accelerate catalyst discovery and optimization | Generated catalyst candidates are not always synthetically accessible; autonomous platforms remain limited to specific reaction classes and equipment configurations | Closer integration of generative catalyst design with autonomous synthesis-testing loops and multiobjective (activity/selectivity/sustainability) optimization |
| Emerging kinetics data benchmarks (e.g., Catechol, Summit, ORDERLY) provide shared datasets for reaction-kinetics ML | Code, trained MLIP/kinetic-model parameters, hyperparameters, and data splits are rarely released alongside published results, making it difficult to reproduce reported rate coefficients or reimplement mechanism-generation pipelines | Journal- and community-level requirements for code, model weights, and versions of software/environments as a condition of publication, mirroring practice in adjacent computational chemistry venues |
Table 1: Kinetics, Catalysis and Reaction Engineering: current capabilities, remaining barriers, and future opportunities.
| Current capabilities | Remaining barriers | Future opportunities |
|---|---|---|
| Deep learning (CNNs, U-Net, attention/residual architectures) segments droplets/bubbles and reconstructs multiphase flow fields directly from holographic and simulation data, without hand-engineered features | Conventional design/optimization strategies (empirical correlations, iterative experiments) still struggle with the spatiotemporal complexity of multiphase flows, and purely data-driven models can violate physical constraints | Continued convergence of high-speed/holographic sensing with physics-informed ML for real-time, self-optimizing multiphase systems |
| Hybrid and physics-informed frameworks (e.g., PINNs) embed mass/momentum/energy conservation directly into architectures or loss functions, improving robustness under sparse or noisy data | Real-time ML-based control of multiphase flows is still constrained by sensing bandwidth and computational cost | Broader deployment of RL-based control for atomization, bubble-column operation, and spray cooling in industrial settings |
| ML/ANN models predict thermophysical and transport properties (viscosity, thermal and ionic conductivity, diffusion coefficients) competitively with or better than semi-empirical correlations, especially for tunable compound families (e.g., ionic liquids) | Thermophysical/transport property data remain scarce, inconsistent, and heterogeneous across sources, limiting the reliability of purely data-driven models | Curation of standardized, cross-laboratory thermophysical property databases to reduce data heterogeneity |
| Thermodynamics-informed ML (embedding Gibbs–Helmholtz, Clausius–Clapeyron and related identities) and ML-augmented/neural EoS improve internal consistency and extrapolation in low-data regimes | Enforcing full thermodynamic consistency (Maxwell relations, convexity of free energies, positivity of entropy production) remains difficult for generic architectures | Deeper architectural embedding of stability/phase-equilibrium constraints so ML augments, rather than replaces, EoS-based thermodynamic theory |
| Inverse-design workflows map molecular representations (SMILES, fingerprints, -profiles) to EoS parameters, enabling rapid virtual screening for target phase behavior | Extrapolation beyond the training domain, critical for process synthesis and safety, remains a persistent weakness; integration into commercial process software requires robust UQ and validation | Development of transparent, well-validated ML thermodynamic modules with long-term maintainability for integration into standard process simulators |
| Neural EoS surrogates and thermodynamics-informed property-prediction models are increasingly reported in the literature | Trained network weights, training/test splits, and the specific software/library versions used to fit thermodynamic surrogates are seldom shared, hindering independent verification against reported accuracies | Minimum code- and data-availability requirements for published thermodynamic/transport ML models, enabling direct reproduction and fair cross-study comparison |
Table 2: Thermodynamics, Transport and Fluid Mechanics: current capabilities, remaining barriers, and future opportunities.
| Current capabilities | Remaining barriers | Future opportunities |
|---|---|---|
| Generative AI (e.g., VAE-based) has been applied to distillation column design and energy-use optimization; interpretable ML predicts performance of intensified gas–liquid operations | Most ML research targets membranes and adsorption; distillation, still the dominant industrial technology, remains comparatively underserved by AI/ML advances | Extending generative and interpretable ML methods to established gas–liquid operations and complex/impure multicomponent streams |
| Large curated databases (CoRE MOF, Polymer Genome) and ML screening (multitask learning, random forest) enable rapid evaluation of MOF/polymer gas-separation and CO 2 -capture performance across millions of candidates | MOF ML workflows rely largely on simulated (DFT/GCMC) data, whereas membrane ML studies mix experimental and simulated data, complicating cross-domain model transfer | Development of unified, experimentally grounded datasets spanning adsorbents and membranes to improve model transferability |
| Generative models (MOFDiff, VAE-based inverse design, ChatMOF) propose entirely new MOF and membrane candidates targeted at CO 2 capture and gas separation | Generated hypothetical adsorbents/membranes are not always synthetically accessible, and models trained on equilibrium isotherms struggle with dynamic, multicomponent industrial conditions | Coupling generative materials design with synthesis-feasibility filters and models that capture adsorption/desorption kinetics alongside thermodynamics |
| Physics-based ML process emulators (e.g., MAPLE, PANACHE) achieve 100 speedups over conventional solvers while retaining transport-equation fidelity for adsorption/chromatography cycles | Integration of molecular-scale material properties with process-level design and control remains only partially realized | End-to-end frameworks that translate material-level predictions directly into optimal process configurations |
| RL and optimization algorithms autonomously tune membrane and adsorption operating parameters (pressure, concentration, temperature) to improve selectivity and reduce energy use and fouling | Emergent, complex fluid–solid interfacial behaviors (cooperative ion transport, dynamic fouling) remain difficult to capture with conventional analytical or numerical methods | Broader real-time, AI-driven control of desalination, lithium-extraction, and gas-separation operations informed by AI-identified interfacial mechanisms |
| Process-level ML emulators (MAPLE, PANACHE) and generative adsorbent/membrane models are increasingly published with reported performance metrics | Underlying code, isotherm/simulation datasets, trained network parameters, and hyperparameters are inconsistently released, complicating independent verification of speedup and accuracy claims | Adoption of explicit code- and data-availability requirements for separations ML publications, comparable to those emerging in adjacent computational chemistry journals |
Table 3: Separations: current capabilities, remaining barriers, and future opportunities.
| Current capabilities | Remaining barriers | Future opportunities |
|---|---|---|
| GNNs and equivariant architectures (NequIP/Allegro, MACE, M3GNet, CHGNet) predict formation energies, band gaps, and elastic moduli near DFT accuracy at a fraction of the cost | Reported benchmark accuracies (e.g., 0.2 eV MAE) are often sufficient for screening but inadequate for quantitative kinetic modeling | Physics-informed extensions of universal potentials that capture non-equilibrium/kinetic behavior beyond static equilibrium properties |
| Universal ML interatomic potentials (e.g., GNoME) combined with iterative ML–DFT cycles have expanded the known stable-materials database by an order of magnitude | Out-of-distribution generalization is weak for compositions with underrepresented elements or unusual coordination environments | Systematic, topology/scaffold-based (rather than random) benchmark splits and comparison against strong traditional baselines |
| Generative and inverse design (CVAE, transformer/foundation models such as MOFormer, MOFTransformer, Uni-MOF) propose new membranes, surfactants, and reticular materials with targeted properties | Hypothetical materials databases used for training often lack synthesis-feasibility annotations, yielding ”high-performing” but unsynthesizable candidates | Coupling generative design with synthesis-feasibility screening and lifecycle/sustainability objectives embedded directly in the optimization |
| Self-driving laboratories (A-Lab, mobile robotic chemists) autonomously synthesize and characterize new materials, with industrial analogues emerging (e.g., automated battery-material sintering and testing) | Diversity of synthesis routes and characterization techniques across material classes limits broader industrial scaling of autonomous platforms | Wider industrial adoption of closed-loop autonomous discovery platforms across additional material classes |
| Standardized benchmarks (Matbench, OC20/OC22) and infrastructure (cloud HPC) support reproducible, community-wide evaluation of materials ML models | Data heterogeneity, inconsistent metadata, and incomplete FAIR-principle adoption across repositories hinder cross-dataset training | Community-wide data harmonization (e.g., via initiatives such as NOMAD) and multimodal, literature-integrated foundation models for materials discovery |
| Benchmark datasets and leaderboards report standardized accuracy metrics across many groups | Code, trained model checkpoints, exact data splits, and hyperparameter/tuning details behind reported benchmark numbers are frequently withheld, so reported accuracies cannot be independently reproduced or fairly compared across studies | Minimum reproducibility requirements (released code, splits, weights, hyperparameters, software versions) as a condition of publication, as increasingly enforced in adjacent computational chemistry journals |
Table 4: Materials and Interfaces, and Applied Chemistry: current capabilities, remaining barriers, and future opportunities.
| Current capabilities | Remaining barriers | Future opportunities |
|---|---|---|
| Bayesian optimization is widely used for process/product design, with extensions for discrete, graph-structured, and robust/safe design under uncertainty | Generative/AI process-design tools remain largely interpolative; genuine extrapolation beyond training data toward novel designs is unproven | Interactive tools/copilots that help engineers visualize, cluster, and navigate high-dimensional process-design spaces |
| ML-learned dynamic models (Koopman operators, sparse regression, neural/Gaussian-process surrogates) support system identification while increasingly preserving stability and dissipativity properties | Lack of open, shared process-design benchmark datasets (unlike molecular/materials domains) due to proprietary industrial data limits systematic algorithm comparison | Community efforts to establish open process-design benchmarks and metrics, drawing on precedents from self-driving-lab and BO communities |
| RL-based and NN-approximated MPC laws, together with ML-accelerated mixed-integer solvers (e.g., learned Benders warm-starts), speed up online control and scheduling computation | High-dimensional, discrete/mixed-integer design and scheduling spaces remain challenging for BO and RL; deep learning controllers can lack interpretability | Tighter integration of superstructure optimization with data-driven surrogate modeling; extending ML-based solver tuning to nonlinear MPC |
| AI/ML forecasting (LSTM, kernel methods) improves demand/inventory/procurement planning; RL supports dynamic scheduling, predictive maintenance, and vehicle routing under uncertainty | Ensuring safety and constraint satisfaction is difficult once RL policies operate outside their training distribution; online exploration is often unsafe in safety-critical settings | Broader deployment of safe/robust BO and constrained or offline RL for uncertainty-aware, safety-critical design and control |
| Standardized RL benchmark environments (OR-Gym, SafeOR-Gym, PC-Gym) and digital-twin/blockchain-enabled data-governance concepts support reproducible PSE algorithm evaluation and resilient supply chains | Supply-chain AI adoption is limited by data-privacy/sharing constraints, high infrastructure costs, and organizational resistance to automation | LLM-assisted extraction of regulatory/sustainability constraints into optimization formulations, and expanded digital-twin-based, closed-loop decision support |
| Benchmark RL environments provide shared, standardized testbeds for algorithm comparison | Beyond these environments, published PSE control/scheduling studies often withhold trained policy parameters, hyperparameters, random seeds, and software versions, making reported performance hard to reproduce or compare across algorithms | Extending benchmark-environment culture into explicit minimum code-, model-, and environment-release requirements for PSE control/scheduling/design publications |
Table 5: Process Systems Engineering: current capabilities, remaining barriers, and future opportunities.
Figure 2: Application areas of AI/ML in industrial operations with substantial AI penetration.
Figure 3: Factors inhibiting widespread adoption of AI/ML solutions.
| Current capabilities | Remaining barriers | Future opportunities |
|---|---|---|
| Process control and monitoring (ML-augmented APC, PCA/PLS, autoencoder-based anomaly detection) is the most mature AI/ML application domain in chemical plants | Many AI/ML applications remain “add-ons” rather than essential components of the operating toolkit, limiting sustained integration | Embedding AI capabilities as core, trusted components of standard operating and control infrastructure |
| Predictive maintenance using vibration, thermal, acoustic, and process-history data has delivered documented reductions in unplanned downtime at companies such as Shell and Marathon Petroleum | “Last-mile” adoption failure: sophisticated models often fail to influence operator decisions due to distrust of opaque, black-box recommendations | Explainable, uncertainty-aware AI systems that build operator trust and gracefully acknowledge out-of-distribution situations |
| Virtual/soft sensors (NN, Gaussian process, ensemble methods), increasingly physics-informed, infer hard-to-measure variables from routine measurements; ML surrogates enable real-time optimization (RTO) | Model drift from catalyst deactivation, fouling, and equipment changes requires ongoing maintenance; “dashboard fatigue” and platform fragmentation hinder adoption at scale | Formalized model-lifecycle-management frameworks (continuous monitoring, automated drift detection, defined retraining triggers, governance) embedded in unified analytics platforms |
| Hybrid digital twins (mechanistic + ML) are deployed at plant scale for real-time health monitoring and have a growing base of plant-scale case studies across process industries | Training data reflect narrow, closed-loop operating envelopes, limiting applicability to optimization/control tasks requiring broader excitation; comprehensive high-fidelity digital twins remain aspirational for many processes | Expanded hybrid digital twins enabling autonomous management of startups, shutdowns, and transitions, with humans focused on supervisory oversight |
| LLM-based copilots are beginning to lower the barrier to natural-language interaction with plant analytics, contextual knowledge retrieval, and operator knowledge transfer | LLM hallucination, confident-but-incorrect outputs, and lack of intrinsic numerical/physical guarantees pose safety risks if used beyond an orchestrator role | Restricting LLMs to orchestrator roles interfacing with verified simulators/historians/optimizers, with human-in-the-loop review before safety-critical actions |
| Published industrial case studies (e.g., digital twins, predictive maintenance, RL-based control) report quantified operational benefits | Proprietary data and IP concerns mean industrial AI/ML studies rarely release code, trained models, or full data-processing pipelines, making reported gains difficult for independent researchers to verify or reproduce | Development of anonymized benchmark problems and reporting norms that let industrial studies disclose methodology and reproducibility artifacts without exposing proprietary plant data |
Table 6: Industrial Applications of AI: current capabilities, remaining barriers, and future opportunities.
Figure 4: For Table of Contents Only