Boosting Screening of Nonequiatomic High-Entropy Electrocatalysts by Inverse Design via Active Graph Learning

High-entropy alloys are promising for enhancing the performance of electrocatalysts. However, the screening and rational design of such catalysts face formidable challenges due to their vast compositional space and diverse local atomic arrangements. Traditional bottom-up approaches of deep-learning algorithms remain limited by their heavy reliance on data sets prepared by density functional theory (DFT), bottlenecking the rapid screening of high-entropy catalysts. To address these limitations, we present an inverse-design strategy based on an active-learning (AL) framework that integrates conditional generative adversarial networks, atomic graph attention networks, k-nearest neighbors, and high-throughput DFT calculations. This top-down framework reduces the required training data set size for high-performance, high-entropy electrocatalyst (HEEC) design. Using the established AL workflow, we systematically explore the compositional space of HEECs composed of Ni, Co, Fe, Pd, and Pt and identify optimized nonequiatomic compositions with high hydrogen evolution activity. Moreover, electronic structure analyses reveal that Pd and Pt are active species, while Ni, Co, and Fe contribute to triggering the “cocktail effect”, which distinguishes HEECs from ordered metals or alloys. Based on these findings, we propose two design principles to guide the discovery of high-performance HEECs: (i) retaining Pd/Pt as essential reaction centers and (ii) utilizing Fe, Co, and Ni to boost entropy for cocktail synergy. This AL-driven approach offers a powerful platform for the accelerated discovery and design of next-generation electrocatalysts.