A Data-Driven Modeling Framework for Lithium Metal Battery Health Estimation

Abstract

The Achilles' heel of LMB, despite its high theoretical energy density, has been its limited lifespan and variability in aging trajectories due to coupling of multiple failure mechanisms. Effective health estimation could enable timely maintenance actions before failure and a rapid LMB development iteration. Data-driven methods capture a wide spectrum of failure mechanisms and incorporates aging behavior variability. Recent work around data-driven LMB health estimation rely on empirical correlations observed specific to the datasets [1][2], and lack adaptability for a wide variety of experimental conditions and cell behaviors, especially those with a flat discharge capacity retention followed by abrupt decline.In this work, we cycled 36 LMB cells across various temperatures and charge/discharge rates. The examined cells exhibit significant cell-to-cell variabilities even under identical experimental conditions, and consistent abrupt capacity fade behaviors. We extracted an extensive inventory of non-invasive features based on cycle-to-cycle variation of LMB external voltage-current measurements throughout aging, and managed to accurately track LMB capacity fade with quantified confidence intervals using these features combined with machine learning tools. The identified features provide not only reliable and robust LMB health monitoring performance, but also starting points for further first-principal modeling and experimental investigation to demystify LMB aging mechanisms. The abundance of discovered features, along with an automatic and time-efficient correlation mining pipeline, enables the generalizability of our proposed framework toward a wide variety of LMB cells.