(488g) Differentiable Optimization for the Prediction of Ground State Structures in Catalysis

Ground-state structure and property predictions are generally performed using density functional theory (DFT) calculations that apply relaxation to randomly generated initial structures. The DFT relaxation often requires high computational cost, which can be a bottleneck in a large scale search for new materials. Here, we present a graph neural networks framework that incorporates differentiable optimization to approximate ground-state structure of the initial structure by learning and directly relaxing the harmonic potential energy surface. Our method predicts near ground-state structures of diverse metal surfaces consisting of up to 4 elements and covering 32 elements, with and without the presence of adsorbates (H and CO). We further demonstrate that this framework can effectively serve as a pre-processing tool that accelerates DFT relaxation by providing a starting configuration near the ground-state.