Machine Learning Assisted Design of Metal–Nitrogen–Carbon Catalysts for the Oxygen Reduction Reaction

Metal–nitrogen–carbon (M-N-C) catalysts are emerging as promising candidates for electrochemical reactions (e.g., oxygen reduction reaction) which are critical for clean and sustainable energy devices. However, due to a large chemical design space, myriad possible structural configurations, and dynamic structure evolution of the metal centers under reaction conditions, the design of these catalysts has been challenging and cost-prohibitive for both experiments and computations. Here, using high throughput density functional theory (DFT) calculations combined with machine learning, we rapidly and efficiently evaluate over 20,000 dual-atom M1M2-N-C catalysts for the oxygen reduction reaction. We first generate a DFT database of a subset of the dual-atom catalysts, and validate our computational predictions of the structure, stability, and catalytic activity with experimental data where available. With this benchmarked database, machine learning models using graph neural networks were trained and applied to identify promising dual-atom catalysts in the search space which possess higher activity than the state-of-the-art Pt catalysts. The computational framework developed in this work can be generally extended to other important electrochemical reactions including carbon dioxide reduction reaction and hydrogen evolution reaction for sustainable energy conversion.