Neural network potential for studying the thermal conductivity of Sn

Sn and its compounds have been widely used in the field of new energy, the photovoltaic material and thermal stabilizer, however the traditional empirical potential and other potential functions cannot accurately describe and calculate the properties of Sn materials. In this work, we use a machine learning (ML) method of the atomic energy network (AE net) to construct the potential function of Sn based on the 11,516 training sets of density functional theory (DFT). By error analysis, we prove that the training result is better than other empirical potentials. Moreover, the results obtained by neural network potential in the molecular dynamics (MD) simulation are in good agreement with the ab initio energy data. We further calculate and predict the thermal conductivity by using the optimized potential of Sn and Green-Kubo formula, and the results are in good agreement with the experimental data. It is intended that this method appears to be a useful tool for simulating arbitrary chemical composition materials and then predicts the thermal stability of a material at a variety of temperatures and pressures. The thermal property predictions using MD and neural network potential are applicable to a wide range of materials and provide new perspectives in the explorations of thermodynamic property.

Automated summary

A machine learning method was used to construct the potential function of Sn materials, showing good agreement with experimental data in predicting thermal conductivity. This approach can be applied to simulate arbitrary chemical composition materials and predict thermal stability.