Exploring compositional-only descriptors in machine learning screening of hydrogen evolution reaction catalysts

The hydrogen evolution reaction (HER) is central to clean energy conversion, as it enables the production of green hydrogen from renewable energy. However, the identification of the particularly suitable electrocatalysts from a broad variety of composition and structural parameters is limited by computational resources, particularly through density functional theory (DFT) calculations of Gibbs free energy. Machine learning (ML) offers a faster route for identifying optimal electrocatalysts. While most ML models rely on large descriptor sets or complex graph networks, here we demonstrate that a few elemental compositional descriptors are alone sufficient to capture the key physics of hydrogen adsorption, the main step in HER. A dataset with hundreds of thousands of entries was filtered, strictly chosen to exclude structural variations, and to ensure that the observed trends are solely governed by elemental composition. Various regression algorithms demonstrated that just a few features give competitive accuracy comparable to algorithms with far larger feature spaces. The paper examines feature importance and identifies that the compositional-only features consist of valence electrons and Coulombic descriptors. Furthermore, tests on an external dataset confirm the robustness of the approach. While these compositional-only descriptors offer an interpretable and universal foundation, incorporating structural descriptors remains crucial for achieving higher accuracy to develop ML models with improved predictivity.