Direct Simulation of LiNi0.8Mn0.1Co0.1O2 Transport Properties Using an Efficient and Accurate Machine Learning Potential

TL;DR

This paper introduces an efficient machine learning potential that enables large-scale molecular dynamics simulations to accurately predict lithium-ion diffusion in NMC811 cathode materials, advancing understanding of their transport properties.

Contribution

The study develops a reliable, data-efficient machine learning potential for NMC811, allowing direct simulation of lithium diffusion at scales previously inaccessible to DFT methods.

Findings

MLP accurately predicts lithium diffusion coefficients

Simulations reveal detailed lithium transport mechanisms

Method reduces computational cost compared to DFT

Abstract

The rate capability of layered lithium nickel manganese cobalt oxide (NMC) cathode materials plays a decisive role in high-power applications such as fast charging, necessitating a detailed understanding of lithium-ion diffusion. However, the mechanisms governing lithium-ion transport in NMC remain insufficiently understood, both experimentally and computationally. In this study, we employ an advanced and efficient machine learning potential (MLP) to simulate lithium self-diffusion in LiNi0.8Mn0.1Co0.1O2 (NMC811), enabling direct large-scale molecular dynamics (MD) simulations. The workflow integrates a fine-tuned MACE (Message Passing Atomic Cluster Expansion) foundation model as a structural generator and leverages an active learning strategy applied to a near-ground-state dataset. This approach enables the construction of a reliable MLP for NMC811 in a data-efficient manner using a limited number of density functional theory (DFT) reference calculations. Based on this potential, we performed MD simulations to predict lithium diffusion coefficients. The MLP-based simulations preserve the accuracy of DFT while overcoming its time and length scale limitations, thereby allowing direct simulation of lithium self-diffusion in NMC811.