(569ah) High Throughput Screening of High Entropy Alloy Catalysts for Oxygen Reduction Reaction through Atomic-Interactive Graph Attention Network

High Entropy Alloy (HEA) has been suggested as one of the solutions which can enhance high activity and stability in Oxygen Reduction Reaction (ORR). However, screening of HEA suffers from time-consuming performing both in experimental and computational methods. Many ML (machine learning) approaches in materials science demonstrated that the search of new candidates can be rapidly achieved with the strong inference capability of ML models. Despite of their remarkable ability, significant numbers do not sufficiently represent real-world conditions in data or algorithms, such as permutation invariance, leading to self-contradiction and low prediction accuracy. Therefore, we developed ML Force field which performs IS2EF task(Initial Structure to Energy and Force) based on graph neural network. Our model has two phases: 1)bulk and 2)adsorption phase, for the sake of lightweight computations. Those commonly employ attention algorithms which represent atomic interaction to predict ground state of energy. Furthermore, adsorption phase has 3-dimensional spatial recognition algorithm to determine the optimal adsorption energy among numerous adsorption configurations. As a result, our model achieved error less than 0.02eV in MAE similar to the level of DFT, and derived new HEA candidates. Our model implies that a model guided by quantum chemical information is superior in its inference ability, leading to the high-throughput screening of HER catalysts for boosting ORR.