Takuya Shibayama1,Hideaki Imamura1,Katsuhiko Nishimra1,Chikashi Shinagawa1,So Takamoto1
Preferred Networks, Inc.1
Takuya Shibayama1,Hideaki Imamura1,Katsuhiko Nishimra1,Chikashi Shinagawa1,So Takamoto1
Preferred Networks, Inc.1
Crystal Structure Prediction (CSP) based on quantum chemistry calculations is used to discover new materials in various materials fields. However, its applicability is limited by its massive computational cost. An accurate and universal machine learning potential would greatly reduce the computational cost and enable CSP to investigate a wide range of realistic materials. We have developed a CSP system using the versatile and efficient NNP named Preferred Potential (PFP). The latest version, PFP v4.0.0, has been released with improved accuracy, enabling more precise CSP studies.<br/><br/>Our CSP system uses PFP and genetic algorithms (GA) to perform local optimization and evaluation of 10,000 samples within an hour using 100 GPUs (equivalent to 10,000 GPU hours). We modified the GA implemented in the Atomic Simulation Environment (ASE) library by extending the sampling algorithms, as well as mutation and crossover. We improved the GA sampler by considering the similarities between CSP and multi-objective optimization problems, and expanded the non-dominated sorting genetic algorithm II (NSGA-II) for crystal structure search, which is specifically designed for multi-objective optimization. To manage trials effectively and in parallel, we employed Optuna, an open-source software framework for hyperparameter optimization.<br/><br/>We performed CSP search for several binary systems and evaluated new crystal candidates using density functional theory (DFT) calculations. More than 10 structures are confirmed as newly found inorganic crystals, which break known convex hulls of the Materials Project database. The new crystals are discovered in systems such as Ti-O, Li-In, Ca-P, and Mg-Pd. More than 50% of the candidate structures suggested by the CSP system using PFP were confirmed by DFT calculations. These results demonstrate the effectiveness of PFP and GA-based CSP in accelerating materials discovery and design. We have also tested our CSP system on a high-performance computing environment using the MN-Core deep learning accelerator, which provided comparable performance to the NVIDIA V100 GPU.