Enhancing atomic structure exploration: The scaled-LCB and similarity-based structure sharing

抄録

Efficient global optimization is crucial in resource-intensive computations, such as atomistic structure search using density functional theory (DFT) calculations. This has been accomplished by combining the DFT with Gaussian process regression within a genetic algorithm called GOFEE. In this work, we modified the GOFEE algorithm to increase its efficiency. We introduced the scaled lower confidence bound acquisition function, which balances the surrogate energy and uncertainty to improve the atomic structure exploration. We implemented similarity-based structure sharing between independent GOFEE calculations to avoid redundant calculations in the target potential. We applied our modifications to find the global minimum in two-dimensional Himmelblau’s function, various hydrocarbon molecules, C₆₀, TiO₂ anatase surface reconstructions, and bulk Co-doped TiO₂ anatase (Co@TiO₂). While our modifications successfully increased the success rate in finding the global minimum of the former systems, they failed to find the ferromagnetic structure of Co@TiO₂, as observed in experiments.