Atomic cluster expansion force field based thermal property material design with density functional theory level accuracy in non-equilibrium molecular dynamics calculations over sub-million atoms (2309.11026v1)

Abstract: Non-equilibrium molecular dynamics (NEMD) techniques are widely used for investigating lattice thermal conductivity. Recently, machine learning force fields (MLFFs) have emerged as a promising approach to enhance the precision in NEMD simulations. This study is aimed at demonstrating the potential of MLFFs in realizing NEMD calculations for large-scale systems containing over 100,000 atoms with density functional theory (DFT)-level accuracy. Specifically, the atomic cluster expansion (ACE) force field is employed, using Si as an example. The ACE potential incorporates 4-body interactions and features a training dataset consisting of 1000 order structures from first-principles molecular dynamics calculations, resulting in a highly accurate vibrational spectrum. Moreover, the ACE potential can reproduce thermal conductivity values comparable with those derived from DFT calculations via the Boltzmann equation. To demonstrate the application of MLFFs to systems containing over 100,000 atoms, NEMD simulations are conducted on thin films ranging from 100 nm to 500 nm, with the 100 nm films exhibiting defect rates of up to 1.5%. The results show that the thermal conductivity deviates by less than 5% from DFT or theoretical results in both scenarios, which highlights the ability of the ACE potential in calculating the thermal conductivity on a large scale with DFT-level accuracy. The proposed approach is expected to promote the application of MLFFs in various fields and serve as a feasible alternative to virtual experiments. Furthermore, this work demonstrates the potential of MLFFs in enhancing the accuracy of NEMD simulations for investigating lattice thermal conductivity for systems with over 100,000 atoms.