Comparative study of crystal structure prediction approaches based on a graph network and an optimization algorithm

Abstract

The combination of a database, graph neural network, and an optimization algorithm is an effective approach for crystal structure prediction (CSP). Considering that there are multiple options for databases, graph network architectures, and optimization algorithms, a benchmark is required to compare the performances of different approaches. We selected 100 crystal structures from the materials database and established a test benchmark for CSP (TBCSP) aimed at rapidly and accurately assessing the performance of various CSP approaches. We found that a combination of the Materials Project database, M3GNet architecture, and Bayesian optimization could achieve a prediction accuracy of up to 40%. These results are encouraging considering the limited amount of training data, diverse elements and crystal structures. This paper provides a promising way to further enhance the accuracy by properly increasing the training data. (Figure presented.)