Advancing Thermal Conductivity Prediction of Metallic Materials by Integrating Molecular Dynamics Simulation with Machine Learning

Molecular dynamics (MD) simulation has an intrinsic limitation when calculating thermal conductivity of metals. That is, only phonon (lattice) thermal conductivity can be derived from trajectory of atoms, which is obtained by solving the Newton 's equations of motion numerically. Therefore, significant contribution of electrons in metals remains unaccounted for. In this study, a Light Gradient Boosting Machine (LightGBM) regression model is employed to predict thermal conductivity of metals using heat flux calculated by MD simulation, electrical conductivity and others as input variables. The LightGBM model successfully predicts the complex non-linear Green-Kubo relation for thermal conductivity calculation even though the underlying physical mechanisms are not entirely clear. The model predicts various temperature dependences of thermal conductivity of metals accurately. Furthermore, the model trained with known compositions of Al-Cu alloys is proved to estimate the thermal conductivity of alloys with unknown compositions. The model also demonstrates a certain level of predictive ability for alloys with different compositions and temperatures. This study demonstrates the potential of a data-driven approach as an efficient method for uncovering complex relationships between incomplete data from MD simulations and true materials properties, especially in cases where the underlying physics is elusive.