Machine learning for catalyst optimization: Outlier detection and material innovation

Abstract

We present a machine learning-driven framework for the discovery and optimization of catalysts in gas adsorption mechanism, focusing on layered heterogeneous catalysts. This approach integrates electronic-structure descriptors with predictive and generative models to explore and evaluate catalyst compositions. By analyzing the adsorption energies of C, O, N, and H, we identify key electronic features that influence chemisorption and govern catalytic performance. Feature attribution methods and permutation importance analyses provide both local and global insights into feature significance, pinpointing critical descriptors that drive material behavior. The generative workflow uncovers novel catalyst candidates and outliers. These outliers — materials situated in low-density regions of the electronic feature space — were analyzed using statistical methods, principal component analysis (PCA), and feature importance techniques to uncover their unique electronic signatures and the potential influence of d-band width and d-band upper edge on catalytic behavior. This strategy accelerates the identification of high-performing catalytic materials, offering a scalable, data-driven pathway for innovation in catalysis and energy storage applications, while ensuring that the discovered materials meet specified adsorption energy ranges for targeted reactions.