Lifetime prediction for lithium-ion batteries undergoing fast charging protocols

Abstract

This thesis describes the application of Porous Electrode Theory and supervised machine learning to lifetime prediction for 18650 lithium iron phosphate (LiFePO₄ LFP)/graphite cells subject to mixed galvanostatic and potentiostatic fast charging policies. Porous Electrode Theory is used to predict battery lifetime by parameteric reductions of effective solid-phase Fickian diffusivities, electrolytic Stefan-Maxwell diffusivity, and Butler-Volmer exchange currents. Parameter estimation and uncertainty quantification are formulated as least squares optimization over galvanostatic discharge curves with Bayesian estimation of uncertainties. A battery lifetime approach from the literature is extended with identifiability analysis to enhance fidelity of the inverse problem, the attribution of degradation modes, and the accuracy of parametric power-law lifetime predictions. Multiphase Porous Electrode Theory (MPET) is also explored in this thesis. In MPET, each particle of the porous electrode ensemble is described by generalized Allen-Cahn-Hilliard dynamics. Single-particle dynamics are governed by firstprinciples free energy landscapes as opposed to inductive fits to open-circuit battery voltages. Multiscale parameter estimation and central limit theorem analysis are implemented, enhancing the suitability of MPET for capacity fade predictions. Supervised machine learning algorithms utilizing feature-based correlations for battery lifetime are described. Electrochemical features that go beyond the discharge-only model provide improved lifetime predictions, generalized voltage analysis indiscrimant of (dis)charge protocol or data, and a clear connection between battery physics and machine learning, and suggest an optimal charging protocol.