High-entropy alloys (HEAs) present considerable potential for catalytic applications. However, their vast compositional spaces render exhaustive first-principles screening impractical. Herein, a two-stage, coarse-to-fine workflow is introduced, integrating a linear-quadratic (LQ) additive model with a graph neural network (GNN). The approach is applicable wherever reactivity can be captured via DFT-based descriptors. As a demonstrative case, materials are screened for the electrocatalytic desulfurization of thiophene. The additive model is fitted using 150 DFT-evaluated compositions to nominate 1,000 five-component HEAs. Subsequently, 350 additional calculations are employed to train a GNN, which ranks the remaining candidates. Relative to computing 3,500 compositions (representing 30% of the 19-metal space), this strategy reduces DFT cost by 7-fold while preserving ranking fidelity, attaining a mean absolute error of 0.1 eV for the top 50 alloys. The approach accelerates discovery by focusing computational resources on the most informative candidates and is readily transferable to other HEA-catalyzed reactions and descriptor-based discovery workflows.
Churuksaev et al. (Fri,) studied this question.