2026 Volume 26 Issue 3 Pages 93-100
Catalysts constitute a core enabling technology for solving a wide range of societal challenges, including resource circulation, mitigation of global warming, and environmental purification. In recent years, data-science-based approaches, particularly machine learning (ML), have attracted significant attention as promising tools to accelerate catalyst development. Despite this expectation, however, there have been very few successful demonstrations in which ML has led to the discovery of genuinely novel catalysts. This limitation arises because newly discovered catalysts typically fall outside the scope of existing datasets and can be regarded as so-called “outliers.” Conventional ML models, which rely on interpolation within known data domains, are fundamentally incapable of making reliable extrapolative predictions for such outliers.
In this review, I first describe our strategy for addressing this challenge by adopting advanced ML methodologies specifically designed to enable extrapolative prediction beyond existing data distributions. I then present a representative example in which a closed-loop exploration framework—iteratively integrating ML-based predictions with experimental investigation—successfully led to the development of a novel multicomponent catalyst for the reverse water–gas shift (RWGS) reaction, which produces CO from a CO2/H2 mixed gas. This closed-loop, ML-guided approach demonstrates the potential to substantially accelerate catalyst discovery and, more broadly, to serve as a powerful paradigm for advancing diverse areas of materials development research.
