Direct simulation of LiNi₀.₈Mn₀.₁Co₀.₁O₂ transport properties using an efficient and accurate machine learning potential

The rate capability of layered lithium nickel manganese cobalt oxide (NMC) cathode materials is decisive for high-power applications such as fast charging, yet the mechanisms governing lithium-ion transport in NMC remain insufficiently understood, both experimentally and computationally. In this work, an efficient and accurate machine learning potential (MLP) was developed for LiNi₀.₈Mn₀.₁Co₀.₁O₂ (NMC811) to enable direct large-scale molecular dynamics (MD) simulations of lithium self-diffusion. The workflow uses a fine-tuned MACE (Message Passing Atomic Cluster Expansion) foundation model as a structure generator together with an active learning strategy applied to a near-ground-state dataset, which allows a reliable MLP to be constructed in a data-efficient manner from a limited number of density functional theory (DFT) reference calculations. MD simulations based on this potential preserve the accuracy of DFT while overcoming its time and length scale limitations, and were used to predict lithium diffusion coefficients in NMC811.

This record contains the DFT reference data used for direct force training and testing of the artificial neural network (ANN) interatomic potential with the atomic energy network (ænet) and ænet-PyTorch packages (https://github.com/atomisticnet/aenet-PyTorch). Each data file contains an atomic structure together with its energy and atomic forces in XCrySDen Structure Format (XSF). The structure and content of the archive are described in README.md.