Graph theory and graph neural network assisted high-throughput crystal structure prediction and screening for energy conversion and storage

Prediction of crystal structures with desirable material properties is a grand challenge in materials research, due to the enormous search space of possible combinations of elements and their countless arrangements in 3D space. Despite the recent progress of a few crystal structure prediction algorithms, most of those methods only target a few specific material families or are restricted to simple systems with limited element diversity. Moreover, these algorithms are usually coupled with first principles calculations and thus are computationally expensive and very time consuming. Therefore, establishing a workflow that can generate a large number of hypothetical structures with diverse elements and quickly optimize and screen out stable structures is urgently needed for the crystal structure prediction field. In this study, we take 17 277 compositions involving 63 elements across the periodic table from the open quantum materials database (OQMD) and use a graph theory assisted universal structure searcher (MAGUS) to generate more than 3.4 million hypothetical structures. We employ a pre-trained universal interatomic potential named Crystal Hamiltonian Graph neural Network (CHGNet) to rapidly optimize this large number of hypothetical structures. Subsequently, we validate these optimized structures using density functional theory (DFT) and find 4145 structures are successfully optimized and 2368 structures have energy lower than those from the original OQMD. 647 structures are further identified to be dynamically stable using CHGNet. Moreover, the stability of 123 out of 200 randomly chosen structures are validated by DFT, corresponding to a high success rate of 61.5%. We further use 4706 DFT data points to train 3 graph neural network models to predict lattice thermal conductivity (LTC) and heat capacity. Numerous structures with ultralow LTC and high heat capacity, which are promising for advanced energy conversion and energy storage, have been identified. The success of our workflow demonstrates that combining graph theory with pre-trained universal interatomic potential is highly expected to accelerate the search for new both thermodynamically and dynamically stable structures with target material properties. Prediction of crystal structures with desirable material properties is a grand challenge in materials research. We deployed graph theory assisted structure searcher and combined with universal machine learning potentials to accelerate the process.