Systematic Feature Design for Cycle Life Prediction of Lithium-Ion Batteries During Formation

TL;DR

This work tackles predicting lithium-ion battery cycle life during the formation phase, a period with limited physical insight and long testing times. It introduces a systematic feature-design framework that selects informative input data, applies fused lasso to obtain a regression template, partitions data into sections, and derives two interpretable $Q(V)$ features from Step B data without extra diagnostic cycles. The two designed $Q(V)$ features achieve a median cycle-life prediction error around 9.2%, outperforming thousands of autoML models while preserving interpretability. A physics-based analysis with a reactive particle ensemble model links these features to formation temperature, electrode utilization, and resistance heterogeneity, providing mechanistic justification and suggesting robustness across protocols. Overall, the framework enables accurate, interpretable extreme early prediction and can accelerate formation optimization by leveraging mechanistic understanding alongside data-driven design.

Abstract

Optimization of the formation step in lithium-ion battery manufacturing is challenging due to limited physical understanding of solid electrolyte interphase formation and the long testing time (~100 days) for cells to reach the end of life. We propose a systematic feature design framework that requires minimal domain knowledge for accurate cycle life prediction during formation. Two simple Q(V) features designed from our framework, extracted from formation data without any additional diagnostic cycles, achieved a median of 9.20% error for cycle life prediction, outperforming thousands of autoML models using pre-defined features. We attribute the strong performance of our designed features to their physical origins - the voltage ranges identified by our framework capture the effects of formation temperature and microscopic particle resistance heterogeneity. By designing highly interpretable features, our approach can accelerate formation research, leveraging the interplay between data-driven feature design and mechanistic understanding.

Figures (15)

• Figure 1: Schematic of formation protocols used for generating the dataset, modified from Figure 1 in Ref. cui_data-driven_2024. Total of 62 protocols were tested by manipulating the six formation protocol parameters: C-rate for the two-step charging ($CC_1$ and $CC_2$), the cutoff voltage between the two CC steps ($CV$), number of cycles between the first charge and last discharge step ($n_{\text{ver}}$), formation temperature ($T$), and the rest time ($t_{\text{OCV}}$). Three common steps among various formation protocols are indicated in blue: the first charge step (Step A), the last discharge step (Step B), and the first discharge step (Step C).
• Figure 2: Systematic feature design framework for extreme early cycle life prediction.
• Figure 3: Determination of $\lambda$ for Outer loop 1 with $Q^{\text{B}}(V)$ based on Table \ref{['tab:downselect_pairs']}. (a) Predictiveness, (b) Robustness, and (c) Interpretability metrics as a function of $\lambda$. The horizontal dotted line indicates the constraint and the vertical dotted lines are the indices where the constraint is activated. The red region is where the constraint is violated and the blue region is where all three constraints are satisfied. (d) Demonstration of $\boldsymbol{\beta}$ at various $\lambda$ values in the blue region. (e) Demonstration of $\boldsymbol{\beta}^{(k)}$'s at smallest $\lambda$ from the blue region for robustness. Colors in panels (a)--(c) and (e) indicate different inner loops.
• Figure 4: Visualization of $\tilde{Q}^{\text{B}}_i(V)$ (colored solid lines) for all cells in the training set, $\boldsymbol{\beta}$ (solid black line), and partition boundaries (vertical dotted lines) for Outer loop 1 with $Q^{\text{B}}(V)$ at $\lambda = 0.4018$. The color of $\tilde{Q}^{\text{B}}_i(V)$ curve indicates the normalized cycle life of the $i$th cell where 1 (red) is for the longest and 0 (blue) is for the shortest in the training set.
• Figure 5: (a) when predicting $\hat{y}_i^{\text{section}}(V_1, V_2)$ with $Q^{\text{B}}_i(V_2)-Q^{\text{B}}_i(V_1)$ and $\text{mean}(Q^{\text{B}}_i(V_1\text{--}V_2))$ in each section in Figure \ref{['fig:partitioning']}. Blue bars indicate the in each section whereas the orange bars indicate the when removing such boundary. The horizontal dotted line indicates the threshold for determining which boundary to remove, which was set to 0.01 in this study. (b) graph after completing the section merging step. All orange bars exceed the threshold. (c) $Q^{\text{B}}(V)$, (d) $\text{d}Q^{\text{B}}/\text{d}V(V)$, and (e) $\text{d}^2Q^{\text{B}}/\text{d}V^2(V)$ with the vertical dotted lines indicating boundaries selected after Algorithm \ref{['alg:merge_sections']} (thin) and the three boundaries near 3.6V (thick). The colors indicate the normalized cycle life and the thick black solid line in panel (e) is for the column-wise average.