Prospects for using artificial intelligence to understand intrinsic kinetics of heterogeneous catalytic reactions
Andrew J. Medford, Todd N. Whittaker, Bjarne Kreitz, David W. Flaherty, and John R. Kitchin, “Prospects for using artificial intelligence to understand intrinsic kinetics of heterogeneous catalytic reactions”, Current Opinion in Chemical Engineering 51, 101232 (2026). https://doi.org/10.1016/j.coche.2026.101232
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
Most uses of artificial intelligence in catalysis accelerate well-defined computational tasks (screening with foundation interatomic potentials, reaction network analysis, literature extraction) or search for new catalyst formulations. Few are designed to address the full multiscale chain from electrons to reactors, or to piece together the incomplete and uncertain data that fundamental catalysis science actually works with. The authors ask how AI could help with the central goal of that science: a consistent picture of the active site and mechanism that quantitatively reproduces measured kinetics and spectra, a task that today relies heavily on human intuition and effort (Medford et al. 2026).
Approach
This is a short opinion and review article in a themed issue on AI and chemical engineering. It proposes the idea of “self-driving models”, the computational counterpart of self-driving laboratories: AI systems that automate some or all of the construction, refinement and validation of a multiscale catalysis model by comparing it directly to measured kinetic and spectroscopic data. The authors frame the problem, then review recent progress in three areas that such systems would connect: atomistic, kinetic and reactor modeling; multimodal experimental data from operando and transient methods; and model fitting, generative models and agentic AI (Medford et al. 2026).
Key findings
The core difficulty is an ill-conditioned inverse problem. Kinetic models can have hundreds or thousands of parameters, and adding reactor, measurement and atomistic models multiplies them, so many different parameter sets, mechanisms or atomic-scale models can reproduce the same observables (a “many-to-one” mapping). The authors argue that modern computing, machine-learned interatomic potentials and kinetic solvers make it conceivable to solve roughly 106 to 108 forward problems to find ensembles of models consistent with experiment, but only if the workflows run without substantial human intervention (Medford et al. 2026).
On the modeling side they highlight a study by Kreitz and co-workers in which the BEEF-vdW ensemble of exchange-correlation functionals was propagated through adsorption energies, automatic mechanism generation with RMG and reactor modeling, producing 2000 light-off profiles for exhaust gas oxidation over Pt(111) in a few hours on a cluster; comparison with experiment identified the most consistent functional. Existing catalysis frameworks (VLab, CATKINAS, AMUSE, Genesys-Cat, RMG) automate much of model construction but still need human effort to install, set up, debug and refine, which is where language models could help (Medford et al. 2026).
On the experimental side, operando X-ray absorption, infrared and Raman spectroscopy and transient methods (TAP, SSITKA, TPSR) give rich multimodal data, illustrated by semiquantitative agreement between transient DRIFTS on Pd/γ-Al2O3 during CO oxidation and microkinetic simulations. Such data are heterogeneous and poorly standardized, and the measurement or instrument model (for example deconvolution of mass spectra) is usually omitted, adding uncontrolled uncertainty. The authors suggest that self-driving models could include measurement models, work from raw data in varied formats via LLM-written analysis scripts, flag mutually inconsistent data subsets, and so encourage raw data sharing (Medford et al. 2026).
For fitting and generation, they point to an adaptive optimization and machine learning loop for a profile reactor (Kouyate et al.), to generative models for optimization (Alves and Kitchin) as a way to sample multiple solutions and address many-to-one mappings, to in-context learning with pre-trained LLMs for catalyst discovery, and to an agentic system that set up, converged and evaluated DFT calculations for CO on Pt(111) as a proof of concept. Model-based design of experiments could then choose the measurements that best discriminate among the resulting ensembles of mechanisms (Medford et al. 2026).
Significance
The article positions self-driving models as an extension of the self-driving lab idea to the computational domain, with an outcome that is inherently interpretable because it is a collection of physical or statistical models that can be shared, generalized and used in process models. The authors argue that the core components already exist and that coupling generative and agentic AI with existing multiscale tools could broaden mechanistic exploration, reduce bias, improve reproducibility and, in the long run, act as community “knowledge engines”. No new data or calculations are reported (Medford et al. 2026).