Abstract: Aiming at the issue of reduced efficiency caused by separate use of first-principles methods and machine learning models in crystal structure prediction, which hinders the effective utilization of information provided by machine learning models, a stable crystal structure prediction algorithm based on multi-fidelity Bayesian optimization is proposed. The algorithm models the potential energy surface of crystal structures through a surrogate model, and the acquisition function selects sampling points along with their corresponding evaluation fidelity based on the modeling results of the potential energy surface. The evaluation function then assesses the selected sampling points, and the evaluation results are used to update the surrogate model. Upon meeting the termination criteria, the algorithm ceases iteration and outputs the final predicted stable crystal structure. Experimental results demonstrate that the proposed algorithm effectively leverages the information from machine learning models, ensuring both the accuracy and quality of the final prediction results while achieving higher efficiency.

Key words: multi-fidelity, Bayesian optimization, Gaussian process, crystal structureprediction