Accurate Prediction of Open-Circuit Voltages of Lithium-Ion Batteries via Delta Learning.

Abstract

Accurate prediction of lithium-ion battery capacity before material synthesis is crucial for accelerating battery material discovery. The capacity can be theoretically determined by integrating open-circuit voltage vs state of charge (OCV-SoC) curves of electrode materials. OCV-SoC curves are traditionally computed using first-principles methods, either through geometry optimization (GO) with density functional theory (DFT) or molecular dynamics (MD) simulations of lithiation/delithiation processes using DFT or force fields. While MD simulations incorporate temperature effects that GO lacks, even DFT-based MD simulated OCV-SoC curves show systematic deviations from experimental results due to inherent approximations in DFT functionals. In this study, we performed MD simulations on 43 cathode materials to obtain their OCV-SoC curves. Initial results showed only moderate agreement with experimental data, yielding a coefficient of determination (R2) of 0.249 and a mean absolute error (MAE) of 1.561 V. Considering the scarcity of data, we implemented a delta learning approach to calibrate the MD results without substantial computational overhead, achieving an improved R2 of 0.933 and an MAE of 0.131 V on the testing set. This calibration method significantly enhanced the accuracy of energy density predictions, reducing the MAE from 106.0 to 10.7 Wh/kg. We also developed an automated delta learning platform to make this approach accessible to researchers without machine learning expertise.