Multi-Model Architecture for Achieving Extended Battery Life Predictions

Abstract

Listen

Translate article

Artificial Intelligence tools such as Machine Learning (ML) will no doubt revolutionize how the battery community approaches many facets of battery research and management. This is particularly true in the area of battery life predictions over periods of time relevant to long-term applications (say, 10-20 years). Yet, with the emergence of ML tools, there is an abiding need to maintain a presence of physics-based “truth” that can guide ML practitioners toward realistic outcomes. This work covers a multi-model architecture that employs a combination of data-driven and materials-driven components that, when in tandem, provide a highly effective framework for achieving life predictions well past the end of test data. Data-driven techniques involve physics-guided deep learning and Bayesian methods, while a materials-driven model predicts fundamental degradation pathways based on materials interactions at the conditions of use, enabling generation of key synthetic data. Results will be presented for various Li-ion cell chemistries (NMC/Gr, LTO/LMO, others), wherein contributions from foremost degradation mechanisms will be identified (loss of lithium inventory, LLI, and loss of active material, LAM). With this we obtain electrode-specific LLI and LAM diagnostics as well as long-term aging predictions that can span several years. The functionality of this multi-model architecture supports materials selection, electrode design, diagnostic aging evaluations, as well as appraisal and optimization of duty cycle conditions to minimize aging trends of the chosen battery chemistry.