A data-driven framework for efficient and physically interpretable prediction of RPA-level adsorption energies

Abstract

Accurately predicting the adsorption energies of reaction intermediates is crucial for the rational design of heterogeneous catalysts. However, high-accuracy methods such as the random phase approximation (RPA) are computationally expensive, hindering rapid catalytic materials discovery. Although machine learning models have been proposed to address this, they are often trained on less accurate data or lack physical interpretability. In this work, we introduce a comprehensive framework that integrates three key components: (1) feature importance analysis to identify critical descriptors, (2) a multi-fidelity approach that correlates expensive RPA data with cheap low-fidelity calculations, and (3) symbolic regression to derive transparent, analytic models. Notably, the model not only achieves high predictive accuracy (with an R² value of 0.96) using just 9 % of the RPA data, but also offers fundamental insight by explicitly linking surface energy to an electronic descriptor. Our work thus establishes a rapid and interpretable pathway for catalytic materials discovery.