Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Constructing accurate machine-learned potentials and performing highly efficient atomistic simulations to predict structural and thermal properties

Junlan Liu, Qian Yin, Mengshu He, Jun Zhou

Abstract

The $\text{Cu}_7\text{P}\text{S}_6$ compound has garnered significant attention due to its potential in thermoelectric applications. In this study, we introduce a neuroevolution potential (NEP), trained on a dataset generated from ab initio molecular dynamics (AIMD) simulations, using the moment tensor potential (MTP) as a reference. The low root mean square errors (RMSEs) for total energy and atomic forces demonstrate the high accuracy and transferability of both the MTP and NEP. We further calculate the phonon density of states (DOS) and radial distribution function (RDF) using both machine learning potentials, comparing the results to density functional theory (DFT) calculations. While the MTP potential offers slightly higher accuracy, the NEP achieves a remarkable 41-fold increase in computational speed. These findings provide detailed microscopic insights into the dynamics and rapid Cu-ion diffusion, paving the way for future studies on Cu-based solid electrolytes and their applications in energy devices.

Constructing accurate machine-learned potentials and performing highly efficient atomistic simulations to predict structural and thermal properties

Abstract

The compound has garnered significant attention due to its potential in thermoelectric applications. In this study, we introduce a neuroevolution potential (NEP), trained on a dataset generated from ab initio molecular dynamics (AIMD) simulations, using the moment tensor potential (MTP) as a reference. The low root mean square errors (RMSEs) for total energy and atomic forces demonstrate the high accuracy and transferability of both the MTP and NEP. We further calculate the phonon density of states (DOS) and radial distribution function (RDF) using both machine learning potentials, comparing the results to density functional theory (DFT) calculations. While the MTP potential offers slightly higher accuracy, the NEP achieves a remarkable 41-fold increase in computational speed. These findings provide detailed microscopic insights into the dynamics and rapid Cu-ion diffusion, paving the way for future studies on Cu-based solid electrolytes and their applications in energy devices.

Paper Structure

This paper contains 3 sections, 5 figures.

Table of Contents

  1. INTRODUCTION
  2. RESULTS AND DISCUSSION
  3. CONCLUSION

Figures (5)

  • Figure 1: RMSEs of energies (a) and atomic forces (b) in the training and validation sets of $\text{Cu}_7\text{P}\text{S}_6$.
  • Figure 2: (a) The evolution of the loss functions (total, L1, L2, energy, and force) during the training process. (b) The energy calculated by NEP versus DFT. (c) The force calculated by NEP versus DFT.
  • Figure 3: The calculated pair distribution function of different atomic pairs in $\text{Cu}_7\text{P}\text{S}_6$ at 300 K obtained from averaged over 200 ps MLMD trajectory with the MTP and NEP potentials, compared with AIMD simulations.
  • Figure 4: The calculated DOS of $\text{Cu}_7\text{P}\text{S}_6$ is obtained from the Fourier transform of the velocity autocorrelation function obtained from MLMD simulations with the MTP and NEP potentials at 300 K, and the results of AIMD calculations as a reference.
  • Figure 5: The computational speed of the NEP (running with one NVIDIA RTX 3080 implemented in GPUMD) is compared to those of the DP (running with one NVIDIA RTX 3080), NEP (CPU version), MTP, and Empirical potential (both running with 64 Intel Xeon Platinum 8375C CPU cores). Note that DP, MTP, and NEP (CPU version) are all implemented in the LAMMPS packageplimpton1995fast.