The convergence of data science and catalysis: a comprehensive review of machine learning applications in accelerating selective catalytic reduction (SCR) catalyst design and performance prediction

Abstract

Selective Catalytic Reduction (SCR) represents the dominant technology for mitigating nitrogen oxide (NOx) emissions, yet designing superior catalysts remains a complex challenge. Machine Learning (ML) offers transformative potential to accelerate SCR catalyst discovery and optimization. However, its practical impact is constrained by critical limitations: the pervasive scarcity and heterogeneity of high-quality experimental and computational data severely limit model robustness and predictive power, particularly for extrapolating to novel designs. Furthermore, extracting generalizable, causal design principles from predominantly correlative ML models remains difficult. This review provides a critical assessment of traditional ML, Deep Learning (DL), and nascent Large Language Model (LLM) applications in SCR catalysis, specifically analyzing Cu-zeolites, V2O5-WO3/TiO2, and Fe-zeolite systems. We scrutinize how data modalities, feature engineering versus representation learning, and model architectures influence the ability to navigate the SCR design space. We provide detailed analysis of ML-assisted mechanistic studies, including Cu speciation dynamics, zeolite structural evolution, and operando spectroscopic analysis. Crucially, unlocking ML’s full potential necessitates data standardization (FAIR principles), physics-informed ML integration, robust explainability methods yielding causal insights, specific uncertainty quantification techniques, and seamless experimental workflow integration. The goal is leveraging enhanced data-driven tools to catalyze faster, more efficient, knowledge-guided development of next-generation SCR catalysts.