Search

Structural search and feature extraction are a central subject in modern materials design, the efficiency of which is currently limited, but can be potentially boosted by machine learning (ML). Here, we develop an ML-based prediction-analysis framework, which includes a symmetry-based combinatorial crystal optimization program (SCCOP) and a feature additive attribution model, to significantly reduce computational costs and to extract property-related structural features. Our method is highly accurate and predictive, and extracts structural features from desired structures to guide materials design. As a case study, we apply our new approach to a two-dimensional B-C-N system, which identifies 28 previously undiscovered stable structures out of 82 compositions; our analysis further establishes the structural features that contribute most to energy and bandgap. Compared to conventional approaches, SCCOP is about 10 times faster while maintaining a comparable accuracy. Our new framework is generally applicable to all types of systems for precise and efficient structural search, providing new insights into the relationship between ML-extracted structural features and physical properties.

Structural prediction for the discovery of novel materials is a long sought after goal of computational physics and materials sciences. The success is rather limited for methods such as the simulated annealing method (SA) that require expensive density functional theory (DFT) calculations and follow unintelligent search paths. Here a machine-learning based crystal combinatorial optimization program (CCOP) with a fitting-search style is proposed to drastically improve the efficiency of structural search in SA. CCOP uses a graph neural network energy prediction model to reduce the DFT cost and a deep reinforcement learning algorithm to direct the search path. Tests on six multi-alloys show the energy prediction model is capable of extracting the bonding characteristics of the complex alloys to achieve interpretability. It also achieves high accuracy with a tiny training set (an increment of 30 samples per iteration) by active learning in less than 5 iterations. Comparison with a few conventional methods shows that CCOP finds the lowest-energy structures with the smallest number of search steps. CCOP cuts the computing cost of SA by two orders of magnitude, while providing better search results than SA. CCOP is promising for serving as a broadly applicable tool for the efficient crystal structure predictions.