Efficient Crystal Structure Prediction Using Universal Neural Network Potential with Diversity Preservation in Genetic Algorithms

Abstract

Crystal structure prediction (CSP) is crucial for identifying stable crystal structures in given systems and is a prerequisite for computational atomistic simulations. Recent advances in neural network potentials (NNPs) have reduced the computational cost of CSP. However, searching for stable crystal structures across the entire composition space in multicomponent systems remains a significant challenge. Here, we propose an improvement of genetic algorithm (GA) -based CSP method using a universal NNP. Our GA-based methods are designed to efficiently expand convex hull volumes while preserving the diversity of crystal structures. Our hull-informed filtering and elitist-selection procedures incorporate an aging mechanism that prioritizes recently improved compositions. We also employ niching to prevent convergence to a small set of stoichiometries, thereby preserving a diverse, high-quality population. Our evaluation shows that the present method outperforms the symmetry-aware random structure generation and existing CSP methods, achieving a larger convex hull with fewer trials. We demonstrated that our approach, combined with the developed universal NNP (PFP), can accurately reproduce and explore phase diagrams obtained through DFT calculations; this indicates the validity of PFP across a wide range of crystal structures and element combinations. This study, which integrates a universal NNP with a GA-based CSP method, highlights the promise of these methods in materials discovery.

Figures (12)

• Figure 1: Biased sampling of the convex hull in the Ti--O system by a standard GA-based CSP method.
• Figure 2: Schematic diagram of the standard GA-based CSP method.
• Figure 3: The similarity between the convex hull in the energy--composition space and the Pareto front in the multi-objective optimization problem.
• Figure 4: The relationship between the parent population, elite population, filtered population, next elite population, and the next population.
• Figure 5: Compositional dispersion history of the Br--Pb--Rh system. The Average Nearest Neighbor Index (ANNI) is plotted against the generation number. The blue line indicates the case with population filtering, while the red line indicates the case without population filtering. The shaded area represents the standard deviation of ANNI values across 10 experiments. Lower ANNI values indicate greater compositional skewness.