Early-Cycle Internal Impedance Enables ML-Based Battery Cycle Life Predictions Across Manufacturers

TL;DR

This work tackles the challenge of predicting lithium-ion battery end-of-life across multiple manufacturers by enriching traditional voltage-capacity features with early-cycle Direct Current Internal Resistance (DCIR) measurements. By building a feature set that includes DCIR-derived statistics across SOC and combining it with voltage-based descriptors through SFS and PCA, the authors achieve robust cross-manufacturer generalization using a simple Elastic Net model; they report substantial MAE reductions (e.g., to around $MAE \approx 150$ cycles) on unseen chemistries. The study uses a diverse dataset of 57 commercial cells from three manufacturers under varied temperatures and voltage ranges, and demonstrates that DCIR features better capture degradation mechanisms such as SEI growth and lithium plating than voltage-only features. The authors also release a novel DCIR-compatible dataset to accelerate broader adoption and benchmarking, highlighting the practical potential for industry to leverage existing data with minimal retraining to design new cells.

Abstract

Predicting the end-of-life (EOL) of lithium-ion batteries across different manufacturers presents significant challenges due to variations in electrode materials, manufacturing processes, cell formats, and a lack of generally available data. Methods that construct features solely on voltage-capacity profile data typically fail to generalize across cell chemistries. This study introduces a methodology that combines traditional voltage-capacity features with Direct Current Internal Resistance (DCIR) measurements, enabling more accurate and generalizable EOL predictions. The use of early-cycle DCIR data captures critical degradation mechanisms related to internal resistance growth, enhancing model robustness. Models are shown to successfully predict the number of cycles to EOL for unseen manufacturers of varied electrode composition with a mean absolute error (MAE) of 150 cycles. This cross-manufacturer generalizability reduces the need for extensive new data collection and retraining, enabling manufacturers to optimize new battery designs using existing datasets. Additionally, a novel DCIR-compatible dataset is released as part of ongoing efforts to enrich the growing ecosystem of cycling data and accelerate battery materials development.

Figures (27)

• Figure 1: Distinct cycling behavior across battery manufacturers complicates cross-manufacturer prediction. (a) Discharge capacity as a function of cycle number for battery cells from three manufacturers (LG MH1, Samsung 50E, and BAK CG) under varied operating conditions. Clear differences in degradation rates and capacity loss profiles can be observed across manufacturers. (b-d) Differential capacity (dQ/dV) curves of the first 50 cycles for (b) BAK CG, (c) Samsung 50E, and (d) LG MH1 cells, highlighting the distinct electrochemical signatures of each cell type. (e) t-SNE plot of features derived from cycling data, showing clustering of cells by manufacturer, illustrating the challenge of model generalization across cell types in battery lifetime predictions.
• Figure 2: Schematic representation of feature extraction for battery cycle life prediction. Starting with 102 cycles of measured time-series data for different cell types and operating conditions, features are computed from differences in voltage-capacity profiles using both the full voltage range and current pulses, and effective resistances (DCIR). The feature set dimensions are reduced using sequential feature selection (SFS) and principle component analysis (PCA). Models are trained to predict the number of cycles to a target end-of-life EOL condition.
• Figure 3: Predicted vs. actual cycles to EOL for different feature sets from an elastic net regressor across varying conditions. Panels (a) (d) and (g) depict predictions using the feature set from severson2019data. Panels (b), (e) and (h) show predictions using a downselected set of features from DCIR current pulse differences at various states of charge between cycles 102 and 2. Panels (c), (f), and (i) show results for a down-selection of all features considered in this work. The top, middle, and bottom rows correspond to leave one triplicate out, leave one operating condition out, and leave one manufacturer out cross validation strategies, respectively.
• Figure 4: Prediction Mean Absolute Error across feature sets, end-of-life thresholds, and cross-validation strategies. Mean Absolute Error (MAE) against feature set averaged over all cross-validation folds using (a) leave-one triplicate out (LOTO), (b) leave-one operating condition out (LOOCO), and (c) leave-one manufacturer out (LOMO) cross-validation strategies and for 90% (purple), 85% (green), and 80% (yellow) end-of-life thresholds.
• Figure 5: Feature-based drivers of model performance. SHAP values (in cycles) showing the contribution of various features to the prediction of EOL$_{85}$ for (a) leave one operating condition out and (b) leave one manufacturer out cross-validation strategies. The features are ordered from top to bottom by relative importance for model prediction. Positive SHAP values indicate a positive contribution to longer predicted cycle life, while negative values imply a reduction in predicted cycle life. The color gradient corresponds to the magnitude of feature values, where darker colors indicate higher values. (c) Voltage difference plotted against $\Delta Q_{DCIR,102-2}^{(0)}$ for cells from three different manufacturers (LG - red, Samsung - blue, BAK - green). Darker colors correspond to a greater number of cycles to EOL.