Machine Learning for NOx Abatement Catalysts: From Data and Descriptors to Predictive Models and Inverse Design

Selective catalytic reduction (SCR), selective catalytic oxidation of NH 3 (NH 3 -SCO), and NO oxidation are core technologies for controlling NOx emissions from both stationary and mobile sources. However, the development of relevant catalysts has long relied on empirical rules and repeated trial-and-error, while the vast composition-structure space, long development cycles, and high costs make it difficult to achieve the synergistic optimization of competing performance metrics, such as low-temperature activity, selectivity, and resistance to SO2/H2O poisoning. Machine learning (ML) offers a data-driven, closed-loop research paradigm for addressing these challenges, enabling the construction of quantitative composition-structure-property relationships and supporting both forward prediction of catalytic performance and inverse design of catalyst compositions. In this article, we systematically review research on NOx catalysts from the four perspectives of data, descriptors, models, and applications, and introduce representative and transferable case studies from other materials systems to summarize the current status and methodological progress of ML in NOx abatement catalysis and related fields. We first outline the performance bottlenecks of typical systems, including NH 3 -SCR, NO oxidation, NOx storage and reduction (NSR), and NH 3 -SCO, and analyze the critical need for ML-based modeling in these systems. We then summarize the major sources of data for NOx reactions, including experimental measurements, density functional theory (DFT) calculations, and literature-mined datasets, and, drawing on mature experience from other catalytic systems, compare four categories of feature engineering strategies based on composition, local structure, energy descriptors, and graph neural networks (GNNs). On this basis, we discuss the applicability and performance of classical ML models and uncertainty-aware models for forward prediction tasks. We further review the main routes and representative examples of inverse design, including surrogate models, Bayesian optimization, and generative models. Most of these representative studies originate from other materials systems, yet their methodological frameworks can directly inform the intelligent design of NOx catalysts. Finally, we analyze four key challenges currently facing this field, namely data scarcity and incompleteness, descriptor mismatch, the difficulty of cross-scale model coupling, and limited interpretability, and we highlight high-throughput experimentation and self-driving platforms as promising pathways for the rational design of NOx catalysts.