From atoms to systems: machine learning in the clean energy materials revolution

作者：Marco Fronzi, Catherine Stampfl, E. Traversa · 发表于：Materials for Renewable and Sustainable Energy · 年份：2026 · DOI：10.1007/s40243-026-00374-6 · 被引用次数：1

Machine learning (ML) has become a pervasive tool in clean-energy materials research, accelerating virtual screening, inverse design, catalyst discovery, process optimisation, and autonomous experimentation across a range of domains. This review offers a systematic and deliberately critical assessment of what that acceleration has and has not yet delivered. Across twelve application domains spanning CO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_2$$\end{document} reduction, hydrogen evolution, nitrogen reduction, ML interatomic potentials, generative inverse design, battery lifetime prediction, sustainable manufacturing, critical-material recycling, and operando characterisation, we introduce the Materials–ML Maturity Matrix (M4) to rate each domain on four dimensions: data reliability, model robustness, experimental and industrial readiness, and generalisation stability. No domain currently achieves high ratings on all four dimensions simultaneously. Four findings emerge from this analysis. First, ML performance outside training distributions is consistently and substantially lower than benchmark results indicate. Second, experimental validation rates for ML-generated candidates remain below 5–10% in inverse-design workflows, and adsorption-energy predictions in catalysis routinely diverge fr...