Thermodynamics-Guided Machine Learning Framework with a Multiobjective Optimizer for Catalyst Discovery

Machine learning (ML) holds great promise for discovering catalysts; however, simultaneously possessing high interpretability, prediction accuracy, and catalyst discovery efficiency remains a substantial challenge. Here, an ML framework with thermodynamic guidelines from scratch is constructed to explore superior multimetallic catalysts for the representative reverse water–gas shift (RWGS) reaction. A normalization approach is employed to redefine the elemental physicochemical properties, leveraging the advantages of the elemental and property features of active metals. This not only enables accurate predictions but also derives the value range of each feature through deep insights into the catalytic mechanisms. More importantly, a genetic algorithm (GA)-based multiobjective optimizer is developed for reverse engineering the compositions of multimetallic catalysts. Specifically, by enforcing explicit thermodynamic constraints and self-correcting by experimentally validated results into the ML models, both rigorous discovery of catalysts within the training data set and extrapolation are successfully achieved. After two rounds of self-corrections, optimal ternary-metallic catalysts are experimentally validated with a superior CO yield close to the equilibrium limitation. Therefore, this ML framework is a promising paradigm for catalyst research, as highlighted by intrinsic theoretical guidance and experimental validation.