Toward Multi-Fidelity Machine Learning Force Field for Cathode Materials

TL;DR

This work tackles the data scarcity challenge in machine learning interatomic potentials for cathode materials by introducing a multi-fidelity graph neural network force field built on CHGNet. The framework fuses low-fidelity non-magnetic and high-fidelity magnetic DFT data through fidelity-dependent embeddings, message passing, readout, and a composition model to improve data efficiency while preserving magnetic moment predictions. Ablation and comparative analyses show the approach yields lower prediction errors across energy, forces, stress, and magnetic moments, and outperforms standard transfer learning and Delta-learning in this context. The method enables accurate, data-efficient simulations of complex battery materials and suggests a path toward general multi-fidelity pre-trained models for battery-relevant systems.

Abstract

Machine learning force fields (MLFFs), which employ neural networks to map atomic structures to system energies, effectively combine the high accuracy of first-principles calculation with the computational efficiency of empirical force fields. They are widely used in computational materials simulations. However, the development and application of MLFFs for lithium-ion battery cathode materials remain relatively limited. This is primarily due to the complex electronic structure characteristics of cathode materials and the resulting scarcity of high-quality computational datasets available for force field training. In this work, we develop a multi-fidelity machine learning force field framework to enhance the data efficiency of computational results, which can simultaneously utilize both low-fidelity non-magnetic and high-fidelity magnetic computational datasets of cathode materials for training. Tests conducted on the lithium manganese iron phosphate (LMFP) cathode material system demonstrate the effectiveness of this multi-fidelity approach. This work helps to achieve high-accuracy MLFF training for cathode materials at a lower training dataset cost, and offers new perspectives for applying MLFFs to computational simulations of cathode materials.

Figures (5)

• Figure 1: Schematic diagram of the multi-fidelity graph machine learning force field architecture. (a) Overall architecture; (b) fidelity-dependent atomic embedding layer; (c) fidelity-dependent message-passing layer; (d) fidelity-dependent readout layer. In (a), the modules related to fidelity are indicated in orange.
• Figure 2: Comparison of test errors between force fields trained using a mixed dataset for multi-fidelity force field training and those trained using only a single high-fidelity dataset. The mixed dataset includes all low-fidelity data and a certain number of high-fidelity data. The abscissa in the figure represents the number of high-fidelity data.
• Figure 3: The impact of the size of the low-fidelity dataset in the mixed dataset for multi-fidelity training on the test error of high-fidelity labels. (a) The number of high-fidelity data is fixed at 400; (b) The number of high-fidelity data is fixed at 50.
• Figure 4: Comparison of training accuracies with combinations of different fidelity-dependent functional modules. N denotes training using only a single high-fidelity dataset (100 frames); F denotes, on this basis, mixing all low-fidelity data and enabling all four modules for multi-fidelity force field training; C, E, M, and R denote respectively only enabling the corresponding fidelity-dependent composition model, atomic embedding layer, message-passing layer, and readout layer modules; $\bar{\text{C}}$, $\bar{\text{E}}$, $\bar{\text{M}}$, and $\bar{\text{R}}$ denote only disabling the corresponding functional modules.
• Figure 5: Comparison of training effects between the multi-fidelity force field architecture and transfer learning. In this experiment, training is conducted using all low-fidelity data and a variable number of high-fidelity data (abscissa).