Dual-scale Model Enabled Explainable-AI toward Decoding Internal Short Circuit Risk of Lithium Metal Batteries

Abstract

The commercialization of lithium metal batteries (LMBs) is blocked by the dendrite-induced internal short-circuits (ISC). However, its risk assessment is hampered by trial-and-error testing and original structure-destructive-induced misleading data. Here, we develop an explainable physical-based data-driven framework, where the transparent assessment of Li dendrite-induced ISC risk is achieved from two aspects. In physics, a dual-scale model integrating microscopic lithium (Li) dendrite simulations with macroscopic ISC model, thus enabling the interpretable connection among the internal microstructure evolution, the cell voltage, and ISC risk, which is not attainable by conventional cell-level ISC models without modeling internal states. In the artificial intelligence (AI) perspective, different from traditional machine learning (ML) models as a “black box", explainable-AI (XAI) analyses over an ML-based ISC surrogate model can quantify both global and local insights into the importance of various factors in ISC risk. SHAP (SHapley Additive exPlanations) analysis identifies grain boundary defects and electrolyte thickness as the most influential factors, followed by charging rate, stack pressure, grain size, contact loss, and ionic conductivity. PDP (Partial Dependence Plots) provides local insights, revealing safety thresholds where higher grain boundary defects (>16.93 GPa), longer electrolyte thickness (>200 µm), charging rate near 0.91C, and grain size around 100 µm significantly mitigate ISC risks. The explainable physical-based data-driven framework is general and readily customized to various batteries and energy systems.