From Atoms to Reactors: A Critical Review of a Decade of Machine Learning in Heterogeneous Catalysis (2017-2026).

This critical review traces the evolution of machine learning (ML) across the full value chain of heterogeneous catalysis over 2017-2026, from catalyst discovery to reactor and process design, and across the principal reaction families (CO2 reduction, water splitting, nitrogen chemistry, zeolite-catalyzed transformations). Each stage is evaluated against four criteria - data foundation, method maturity, experimental validation and deployment readiness - through a structured capability scorecard. Digitalization has advanced from the top of the length scale downward: prediction of adsorption energies and atomistic potential-energy surfaces is now mature, with errors of tens of meV relative to the reference DFT calculations and accelerations of three orders of magnitude over DFT, yet experimentally validated catalyst improvements remain rare. The most significant advance is the closed-loop autonomous laboratory coupling Bayesian optimization, automated synthesis and high-throughput testing. Large language models have begun to add a reasoning and orchestration layer, while kinetics, reactor and process design remain the least digitalized stages. We argue that the field's next decade of impact will be decided not by more accurate atomic-scale prediction but by carrying ML into experimental validation and the reactor-and-process scale, where catalysis is hardest to industrialize and where data-driven methods are weakest.