Efficient optimization of atom/ion arrangements in crystalline solids using genetic algorithms and machine-learning regression

In crystalline materials composed of multiple elements, such as alloys and solid solutions, degrees of freedom for atomic or ionic arrangements arise, making the determination of reasonable atom/ion configurations an important aspect of simulations. However, even in relatively small simulation cells, the number of possible arrangements is vast, rendering exhaustive evaluation infeasible. Although methodologies such as Monte Carlo and special quasi-random structures method have been proposed, genetic algorithm (GA) optimization is particularly useful for identifying stable arrangements, as it is applicable to bulk systems, surfaces, and interfaces. In this study, we improve the search method by combining GA with machine learning (ML), which we refer to as the GA and ML regression analysis (GAML). Specifically, this approach uses ML to screen and evaluate some of the structures generated by a GA, thereby reducing the computational demand of material simulations. This study provides an overview of the GAML, its computational methods, and optimization examples, demonstrating that the GAML achieves optimized structure more efficiently than the conventional GA. Integrating ML into GA significantly enhances the efficiency of optimizing atomic and ionic arrangements in crystalline solids. By achieving stable structures in fewer generations compared with traditional methods, the GAML offers a powerful tool for addressing complex systems with numerous possible configurations, with broad implications for accelerating materials discovery and design, particularly in fields requiring computationally efficient optimization of large and intricate systems.