Design high-entropy electrocatalyst via interpretable deep graph attention learning

High-entropy electrocatalysts (HEECs) have been attracting extensive attention because of their multiple merits in heterogeneous catalysis. However, the diverse local environments and vast phase space behind HEECs make experimental and ab initio exploration unaffordable. In this work, we develop an accurate and efficient atomic graph attention (AGAT) network to accelerate the design of high-performance HEECs. The reliability of scaling relations and classical d-band theory is confirmed on HEEC surfaces on a statistical basis. Nonetheless, we prove that HEEC can effectively bypass the scaling relations by providing ample versatile local environments. We apply the model to explore the compositional space composed of Ni-Co-Fe-Pd-Pt, and high-performance compositions are recommended and validated by our experiments. The AGAT is inherently interpretable, as attention scores elegantly explain its behavior, which shows good agreement with physical principles. Through the interpretable AGAT model, this work opens an avenue for rational design and high-throughput screening of high-performance HEECs.