Elucidation and prediction of ion transport in battery materials: A first-principles and machine learning study

Abstract

Facile ionic mobility within host frameworks is crucial to the design of high-energy-density batteries with high-power-densities, where the activation energy required for ion migration (Em) between crystallographic sites within electrodes or solid electrolytes, is a governing factor. Em serves as a critical metric for evaluating the commercial viability of battery materials as it directly related to the ionic diffusivity (D(x)) of the intercalants within the host structure, which in turn affects the rate performance of the batteries. Accurate estimation of Em and D(x) using traditional computational methods such as density functional theory (DFT) based nudged elastic band (NEB) simulations or ab-initio molecular dynamics, are typically hindered by several computational challenges, including convergence issues, the demands associated with sampling larger time and length scales, and the scaling of computational cost with system size.

To mitigate these limitations, we introduce a machine learning based approach designed to efficiently and directly predict Em. Our first objective was to benchmark the accuracy of different-exchange (XC) correlation functionals within the DFT-NEB framework in the prediction of Em against the experimental data. The important observations from this work were used to construct a comprehensive dataset comprising of 619 distinct literature reported DFT-NEB Em values. Secondly, to overcome the limited data availability, we propose a multi-property prediction model (MPT), which leverages a graph neural network (GNN) architecture that was pre-trained (PT) on seven bulk material properties that subsequently utilises transfer learning (TL) to enhance Em prediction accuracy. Finally, we fine-tune (FT) our MPT model on our manually curated literature-derived dataset of 619 datapoints, using four different modifications to the MPT architectures. Our best performing FT model demonstrates substantial improvements in prediction accuracy compared to traditional statistical ML methods and models trained from scratch (i.e., without any pre-training). Notably, the proposed model also effectively distinguishes multiple migration pathways within a given structure. Additionally, through our FT strategy, we illustrate modifications that enable our MPT model to adapt effectively to other data-scarce material properties in materials science. Ultimately, our FT model and the demonstrated modification strategies promise to accelerate material screening efforts, particularly for estimating Em, thereby facilitating faster discovery of novel electrode and electrolyte materials for advanced batteries.