Technical Summary: Lie Generator Networks for EIS-Grade Battery Diagnostics from Pulse Relaxation Data

Problem Statement

Electrochemical Impedance Spectroscopy (EIS) is the gold standard for lithium-ion battery diagnostics, capable of decomposing cell behavior into distinct electrochemical processes (e.g., charge transfer, solid-electrolyte interphase transport, solid-state diffusion) via frequency-resolved spectra. However, EIS requires dedicated hardware, controlled environments, and acquisition times ranging from 10 to 60 minutes, rendering it incompatible with field deployment, fleet-scale monitoring, or rapid sorting of retired cells for second-life applications.

Current alternatives face significant limitations:

• Machine Learning on Cycling Data: Requires tens to hundreds of full cycles and yields scalar predictions without mechanistic decomposition.
• Pulse-to-Impedance Reconstruction: Reduces measurement time but often treats the spectrum as an end product rather than extracting the underlying physical time constants.
• Equivalent Circuit Model (ECM) Fitting: Directly fitting resistance and capacitance parameters to pulse data suffers from ill-conditioned optimization landscapes, where multiple parameter sets can fit the same waveform, leading to instability and non-unique solutions.

There is a critical need for a method that extracts physically interpretable electrochemical time constants directly from seconds of pulse data, with mathematical guarantees on stability, without requiring paired impedance measurements for training, and without retraining when battery chemistry changes.

Methodology: Lie Generator Networks (LGN)

The authors propose applying Lie Generator Networks (LGN), a structure-preserving linear system identification framework, to battery diagnostics.

Theoretical Framework

The method models the post-pulse voltage relaxation of a lithium-ion cell as an autonomous linear system derived from an $n$-th order equivalent circuit model (parallel RC branches in series). The voltage evolution is governed by:
$$\dot{v}(t) = A v(t) \implies v(t) = e^{At}v(0)$$
where $v(t)$ is the vector of branch voltages and $A$ is the system matrix. The measurable terminal overpotential $\eta(t)$ is the sum of these branch voltages.

Network Architecture and Constraints

LGN learns the generator matrix $A$ directly from data. To ensure physical validity and stability:

1. Structure Preservation: The matrix $A$ is parameterized as $A = S - D$, where $S$ is skew-symmetric (energy-preserving) and $D$ is positive-definite diagonal (dissipative).
2. Minimal Relaxation Parameterization: For post-pulse relaxation (purely dissipative), the authors use the minimal form $A = -D$, where $D = \text{diag}(d_1, \dots, d_n)$ with $d_i > 0$.
3. Stability Guarantee: Stability is guaranteed by construction (all eigenvalues are confined to the negative real half-plane) rather than via regularization. This eliminates the ill-conditioning found in unconstrained curve fitting.
4. Time Constant Extraction: The electrochemical time constants are derived as $\tau_i = 1/d_i$. Positivity is enforced via a softplus map on the learned parameters.

Optimization

LGN minimizes the reconstruction loss between measured overpotential and the model output:
$$\mathcal{L}(\theta) = \sum_{k=1}^T |\eta(t_k) - \mathbf{1}^T \exp(A_\theta t_k) v_0|^2$$
Crucially, gradients propagate through the matrix exponential via automatic differentiation, coupling all eigenvalues during optimization. This structural constraint prevents mode-swapping instabilities common in unconstrained multi-exponential fitting.

The model is trained on short pulse relaxation data (e.g., 40–60 seconds) using warm-starting from previous aging states, exploiting the physical prior that electrochemical parameters evolve gradually.

Key Contributions and Results

1. Degradation Tracking with EIS-Grade Fidelity

Validated on the KIT dataset (227 NMC/SiO cells), LGN extracts three time constants ($\tau_1, \tau_2, \tau_3$) and associated resistances from a single 60-second HPPC pulse.

• Correlation: The extracted $\tau_1$ shows a near-perfect monotonic relationship with State of Health (SOH), achieving a median within-cell Spearman rank correlation of $|\rho_s| = 0.999$.
• Efficiency: This accuracy is achieved in 1/13th the time of the best single-frequency EIS feature (which requires ~13 minutes) and without additional hardware.
• Mechanistic Insight: Decomposition reveals that $\tau_1$ captures the product of increasing resistance ($R_1$) and collapsing capacitance ($C_1$), providing a more sensitive degradation coordinate than resistance alone.
• Spectrum Reconstruction: Using only the seven LGN parameters extracted from the pulse, a linear ridge regression model can reconstruct the full complex impedance spectrum (Nyquist plot) with a median Mean Absolute Percentage Error (MAPE) of 2.0% across 227 cells, without using any EIS data from the test cell during fitting.

2. Early-Life Prognosis

Using the TRI Aging Matrix (374 NCA/graphite+SiO$_x$ cells), the authors demonstrated that time-constant trajectories predict end-of-life outcomes better than capacity alone.

• Orthogonal Information: Cells with identical initial capacity and capacity slope showed divergent lifetimes based on the slope of $\tau_3$.
• Prognostic Accuracy: In pairwise tests where cells were indistinguishable by conventional metrics, the $\tau_3$ slope correctly identified the cell with the shorter remaining life with 71–82% accuracy, compared to ~58% for capacity slope.
• Time Reduction: LGN achieves comparable prognostic accuracy to methods requiring 100 full cycles (e.g., Severson et al., 2019) using only 7 pulses of 40 seconds each, representing a 540× reduction in diagnostic time.

3. Manufacturing Quality Control

Applied to a Samsung INR21700-50E dataset (254 cells, 32 batches), LGN served as a rapid manufacturing fingerprint.

• Batch Discrimination: LGN identified an anomalous subpopulation (batches 21–25) with a 21% depression in $\tau_1$, localized to fast interfacial processes.
• Superiority over Curve Fitting: While LGN achieved a Kruskal–Wallis statistic of $H=117$ ($p < 10^{-11}$) for batch discrimination, standard unconstrained curve fitting on the same data yielded $H=35$ ($p=0.28$), failing to reject the null hypothesis that all batches were identical. This highlights LGN's ability to resolve manufacturing variation through stable, structure-preserving identification.

4. Electrochemical Validation

Validated on a Panasonic NCR18650PF cell across 50 operating conditions (13 SOC levels, 5 temperatures), LGN recovered genuine electrochemical processes without physics priors.

• Arrhenius Validation: LGN-derived time constants followed Arrhenius kinetics. The activation energy for $\tau_1$ (234 meV) differed from EIS-derived resistance $R_1$ (335 meV) by exactly the amount expected from the temperature dependence of double-layer capacitance, confirming LGN captures the coupled $R \times C$ physics.
• Butler-Volmer Consistency: The SOC dependence of $\tau_1$ transitioned from SOC-independent at 25°C to a U-shaped dependence at lower temperatures, consistent with the shift from activation-controlled to concentration-limited kinetics.
• Stability: The model maintained sub-1% NRMSE across all conditions with zero ordering violations ($\tau_1 < \tau_2 < \tau_3$), even with stiffness ratios up to 2,386×.

Significance and Claims

The paper claims that the primary contribution is not merely that pulse data can diagnose batteries, but that pulse relaxation data and impedance spectroscopy are connected through a shared mathematical object: the eigenvalues of the relaxation generator.

• Bridging Time and Frequency: LGN operationalizes this connection by providing a stable, training-free algorithm that extracts these eigenvalues from any pulse waveform. This converts a laboratory-only measurement (EIS) into a real-time capability for existing Battery Management Systems (BMS).
• Physics-Informed without Priors: The method recovers genuine electrochemical processes (activation energies, kinetic transitions) without requiring explicit physics priors, chemistry-specific calibration, or retraining for different cell chemistries.
• Scalability: By leveraging data already collected by BMS (voltage relaxation during rest periods), LGN enables real-time health monitoring, rapid second-life grading, production-line quality control, and early-warning prognosis using minutes of measurement rather than hours or days.
• Robustness: The structural constraints of LGN (stability by construction, coupled eigenvalue optimization) solve the ill-conditioning problems of conventional curve fitting, allowing the method to resolve subtle manufacturing variations and degradation trends that other methods miss.

The authors conclude that the diagnostic infrastructure for terawatt-hour scale battery deployment is already in place; the missing element was an algorithm capable of reading the electrochemical information already contained within standard pulse data.