Accuracy of Machine Learning Potential for Predictions of Multiple-Target Physical Properties

The accurate and rapid prediction of materials' physical properties, such as thermal transport and mechanical properties, are of particular importance for potential applications of featuring novel materials. We demonstrate, using graphene as an example, how machine learning potential, combined with the Boltzmann transport equation and molecular dynamics simulations, can simultaneously provide an accurate prediction of multiple-target physical properties, with an accuracy comparable to that of density functional theory calculation and/or experimental measurements. Benchmarked quantities include the Grüneisen parameter, the thermal expansion coefficient, Young's modulus, Poisson's ratio, and thermal conductivity. Moreover, the transferability of commonly used empirical potential in predicting multiple-target physical properties is also examined. Our study suggests that atomic simulation, in conjunction with machine learning potential, represents a promising method of exploring the various physical properties of novel materials.