Uncertainty Based\nMachine Learning-DFT Hybrid Framework\nfor Accelerating Geometry Optimization

Geometry optimization is an important tool used for computational\nsimulations in the fields of chemistry, physics, and material science.\nDeveloping more efficient and reliable algorithms to reduce the number\nof force evaluations would lead to accelerated computational modeling\nand materials discovery. Here, we present a delta method-based neural\nnetwork-density functional theory (DFT) hybrid optimizer to improve\nthe computational efficiency of geometry optimization. Compared to\nprevious active learning approaches, our algorithm adds two key features:\na modified delta method incorporating force information to enhance\nefficiency in uncertainty estimation, and a quasi-Newton approach\nbased upon a Hessian matrix calculated from the neural network; the\nlater improving stability of optimization near critical points. We\nbenchmarked our optimizer against commonly used optimization algorithms\nusing systems including bulk metal, metal surface, metal hydride,\nand an oxide cluster. The results demonstrate that our optimizer effectively\nreduces the number of DFT force calls by 2–3 times in all test\nsystems.