Feature engineering methods for machine learning in heterogeneous catalysis

Abstract

Machine learning is revolutionizing the field of heterogeneous catalysis, transitioning from a supporting tool to a central force in materials discovery and mechanistic understanding. At the heart of this transformation lies feature engineering, which bridges the catalyst structure with predictive modeling capabilities. In this review, we provide a systematic overview of the evolution of feature engineering in heterogeneous catalysis. This progression spans hand-crafted descriptors, symbolic regression methods, graph-based features that capture intricate chemical and geometric relationships, topological data features encoding multiscale structural invariants, and most recently, multimodal representations that integrate textual data and structure into unified feature spaces. Despite these advancements, several challenges remain in feature engineering, including the underdevelopment of multimodal representations, limited model interpretability, and the absence of cross-scale structural descriptors. Emerging strategies aimed at addressing these issues are discussed in detail. We hope that this review will inspire further innovation in feature engineering methodologies tailored to the continued advancement of heterogeneous catalysis.