Identification and Quantification of Aging Modes with Deep Learning Models

Abstract

Identifying the dominating aging modes and mechanisms uniquely with higher confidence using early life-cycle data is a significant scientific challenge. In general, lithium-ion batteries experience different aging modes depending on use conditions—e.g., loss of lithium inventory and loss of active material in both electrodes, loss of power capability, etc.—due to different relevant aging mechanisms. The existing body of research almost exclusively uses extended-calendar or cycling-life data along with different post-testing to identify and quantify different dominating aging modes and mechanisms. Early identification of battery aging modes/mechanisms will enable accurate forecasting of battery state-of-health along with the root cause and shorten the design and testing cycle for new battery chemistries and cell designs for emerging applications. In this presentation, we present a deep learning (DL) modeling framework to enable early identification and quantification of dominating cell aging modes using battery incremental capacity (IC) information at different aging states (Figure 1). The DL model utilizes coin cell data with altered offset and N:P ratio simulating different aging conditions to train and construct two categories of DL models, serving two consecutive steps in the hierarchical learning framework: (i) a classification model to identify the dominating aging modes, e.g. LLI + LAMPE or LLI + LAMNE, (ii) a regression model to quantify the percentages for dominating aging modes, e.g., percentages of LLI and LAM. The deep learning models are built using a 1-D convolutional neural network structure. The deep learning model’s performance is verified with experimental Gr/NMC pouch cells data with different loadings and utilization conditions. It is demonstrated that the classification model can identify dominating aging modes successfully for current experimental validation sets within 50 cycles, and the quantification model can further determine their percentages of LLI and LAM under aging modes. Figure 1