High-throughput informed machine learning models for ultrastrong B-N solids

High-throughput (HT) computations and machine learning (ML) algorithms are two fundamental approaches in data-driven paradigms to predict various properties of solids due to their efficiency in data creation and model construction, which however are usually used individually and lack generalization and flexibility. In this paper, we propose a scheme combining HT computations for the efficient creation of consistent data and ML algorithms for the fast construction of surrogate models to screen B-N solids' stability and mechanical properties at ambient and high pressures. Employing HT computations, a standardized database of formation enthalpy, elasticity and ideal strength of thousands of B-N structures with high precision is first established. Then several ML models are comparatively built employing the XGBoost approach with the consideration of four descriptors, i.e., sine matrix, Ewald sum matrix, SOAP, and MBTR. Our results suggest the MBTR provides more accurate estimates of various physical properties except for bulk modulus, which is evaluated with greater precision by the Ewald sum matrix. To further improve the model interpretability, based on the brittleness/ductility criterion of materials, a symbolic model with strong physical significance is successfully built for the ideal strength of the B-N solids through the key descriptors screened by the ML methods, showing great accuracy. Our research demonstrates the possibility of building high-efficient ML models and compact symbolic physical models by incorporating consistent data through HT computations for high accuracy in predicting the thermodynamic and mechanical properties of strong solids with high precision, providing a pathway for inverse design of novel materials.

Automated summary

This paper proposes a scheme that combines high-throughput computations and machine learning algorithms to predict the stability and mechanical properties of B-N solids under ambient and high pressures. The results demonstrate the possibility of building high-efficient machine learning models and compact symbolic physical models with high precision for predicting the thermodynamic and mechanical properties of strong solids.