A large lithium-ion cell, such as a 135 Ah "blade" cell, is better described as many small cells wired in parallel inside one can than as a single uniform object. Regions near the tabs run hotter, carry more current, and age faster than regions far from them. A battery management system cannot see this. It measures one terminal voltage, one total current, and a few surface temperatures, and these readings average over all the regions. Single-cell biology has the same problem: a bulk RNA measurement averages over thousands of cells, and in that field a whole toolkit exists for "deconvolution", meaning the recovery of the hidden constituents from the averaged signal. This note proposes carrying that toolkit's central idea into batteries. We infer a probability distribution over the internal regional states of a cell from its external signals, and we force every sample to satisfy the physical sum rules exactly: regional currents must add up to the terminal current, every region must sit at the shared terminal voltage, and regional heats must add up to what the surface sensors imply.
The practical question we target is one that bulk measurements cannot answer. When a cell has lost 10% of its capacity, did every region lose 10%, or did a few regions near the tab lose most of it? Biology faces the identical question when it asks whether therapy resistance is a uniform adaptation of the tumor or the expansion of a rare subpopulation (Xue Lin et al.). The two cases have the same terminal signature but very different safety implications. Concentrated damage is where lithium plating and early internal shorts begin, and Chai et al. show that slow, localized fault evolution is a more dangerous route to thermal runaway propagation than sudden abuse. We will deliver a method that reports, with calibrated uncertainty, which of these two situations the data support. It will also report when the data cannot tell them apart.
Core Claim
We model the cell as $N$ parallel regions with parameters $\theta_j$ (capacity, resistance, solid diffusivity, reaction rate, and an optional internal-short conductance). Each region also has a current $i_j(t)$ and a temperature $T_j(t)$. Inference must respect the following constraints:
$$ I(t)=\sum_{j=1}^{N} i_j(t),\qquad V(t)=U(\theta_j, i_j, T_j, t)\ \text{for every } j,\qquad \mathbf{T}_{\text{surf}}(t)=\mathcal{H}\big(q_1,\dots,q_N\big)(t), $$
Here $I$ and $V$ are the measured terminal current and voltage. $U$ is the regional voltage given by a coupled electrochemical-thermal model. $q_j$ is the regional heat source (ohmic, polarization, and entropic terms in the Bernardi form). $\mathcal{H}$ maps the regional heat sources to the readings of the surface thermocouples through heat conduction.
We claim three things. First, a prior over plausible regional trajectories, learned from a coupled simulator, combined with sampling projected onto these constraints, recovers regional state of charge, current share, and temperature well enough to separate "uniform" from "concentrated" degradation. Second, the posterior spread, read as a function of excitation profile and sensor layout, gives an identifiability map: it shows which heterogeneity patterns are visible at all from the outside. Third, a residual channel outside the learned prior can flag local faults that the simulator library never contained, rather than forcing them onto a "healthy" explanation.
The third point is taken directly from biology's own warning. Reference-based deconvolution can hide pathological novelty by projecting it onto known healthy states (Johansen et al.; Venkat et al.). In a battery, the novel state is usually the dangerous one.
Relation to Prior Work
The battery side supplies the forward physics and motivates the gap. Chen et al. build a 3D coupled electro-thermal model of a 135 Ah blade cell and show that temperature and state-of-charge gradients grow with charging rate and concentrate at the tabs. Their model is a forward simulation with fixed parameters and cannot infer the internal distribution of a particular cell from its terminal data; we solve that inverse problem. KUANG Ke et al. calibrate a 153 Ah cell by tearing it down and testing button cells across four temperatures. They argue that terminal data cannot separate the internal parameters, and they treat the harvested electrodes as uniform. We take their claim seriously as a constraint: our identifiability map is a direct test of how far non-destructive inference can go, region by region.
Allendes et al. show that lumped thermal parameters can be recovered from noisy surface temperatures with under 1% error. Their model has one node, and we ask what happens when the cell has many. The low-temperature aging study (Dilemma of C-Rate and Cycle Life) supplies the degradation physics we simulate. In that model, plating dominates capacity loss at −20 °C and 3C, while SEI growth dominates at 45 °C. Because the competing mechanisms depend so strongly on temperature, a hot region and a cold region in the same cell age by different mechanisms. The study itself assumes uniform particles and forward simulation only.
The biology side supplies the method. Count Bridges (Fishman et al.) treat unit-level states as latent variables in an EM loop and use projection-guided sampling so that generated constituents sum exactly to the observed aggregate. We reuse the projection idea and the Energy Score training objective, but our constituents are continuous trajectories with an extra shared-voltage constraint, not integer counts. scProjection (Johansen et al.) learns a generative model of valid constituent states from a reference atlas and optimizes latent codes so that their mixture reconstructs a contaminated measurement. Our "atlas" is a library of simulated regional trajectories. POPPY (Venkat et al.) and SubNetDL (Ha et al.) mainly motivate the work. SubNetDL's finding that shuffling mutations across subclones destroys predictive power is the biological version of our hypothesis that the arrangement of degradation inside a cell carries information that the cell's total degradation does not.
Research Program
The first step is building the reference library. We will implement a reduced coupled model, with single-particle or P2D electrochemistry per region and a conduction network between regions. Its parameters will be anchored to the published cell descriptions: the LFP blade geometry and materials library from Chen et al., and the NCM kinetic and Arrhenius parameters from KUANG Ke et al.. We accept the reduced model only if its terminal fit matches the published full-model accuracy for those cells. That accuracy is a voltage RMSE below 7.5 mV and temperature error below 2 °C for the blade cell, and voltage error below 10 mV and temperature error below 1.1 °C for the NCM cell. We will then sample heterogeneity patterns over the regions, including tab-concentrated aging, random spread, plating-dominated cold regions, and SEI-dominated hot regions, and simulate each pattern under constant-current, drive-cycle, and multi-frequency pulse profiles.
The second step is inference. A generative model trained with the Energy Score learns the distribution of regional parameter vectors and their trajectories. At test time, we sample from its posterior given the measured terminal and surface data. After each update, we project the samples onto the set of trajectories that satisfy the sum rules. The posterior includes an explicit residual term: an additive regional current or heat source that lies outside the learned prior and carries a sparsity penalty. A large, localized residual is our fault flag.
The third step is the identifiability analysis. For each sensor layout (terminal signals only, two thermocouples, or a dense thermocouple array) and each excitation profile, we measure how much the posterior contracts relative to the prior for each heterogeneity pattern. The output is a map of what external measurement can and cannot reveal. We expect, for example, that the voltage signature of tab-concentrated resistance growth is separable under pulse excitation, while a uniform loss of capacity spread evenly is not.
Baselines
The naive reference point is the standard lumped equivalent circuit model with a single thermal node, fitted to the same data. Its estimate of every region is the cell average. Any proposed method that does not beat this on regional error for concentrated patterns has failed. KUANG Ke et al. already note that such models lack insight into internal processes.
The strongest classical competitor is direct Bayesian inversion of the same $N$-region distributed model. We will run it with MCMC and, separately, with an ensemble Kalman filter, using the identical simulator, sensors, and priors, and matching our method's wall-clock budget. This baseline isolates whether the learned reference prior and projection add anything beyond simply having the right physics.
The strongest data-driven competitor is a supervised network trained on the same simulated library to map terminal signals directly to regional states. It uses the CNN-BiLSTM-Attention architecture that Xiaoyu Li et al. found best for thermal runaway temperature prediction, where it reported MAPE below 5%. That number concerns a different task, a single temperature target, and is not directly comparable. We found no reported result on Lacuna for recovering regional internal states of large-format cells from external signals, so the comparison will be established here.
Three ablations isolate the key ingredients. Removing the projection step tests whether exact enforcement of the sum rules matters. Replacing the learned prior with a Gaussian smoothness prior tests whether the reference library matters. Removing the residual channel tests whether faults absent from the library are otherwise forced onto healthy explanations.
To count as beating these baselines, the method must achieve lower regional state-of-charge and temperature error on concentrated patterns at equal compute. Its posterior intervals must cover the truth at the nominal rate. It must flag an internal short earlier than threshold rules on terminal signals do. And it must be no worse than the lumped model on uniform patterns.
Why not just…?
Why not just fit a distributed multi-node model with MCMC or a Kalman filter? This is the real threat, and it is why it is our main baseline. If MCMC with the right physics matches our method's accuracy and calibration within the same compute, then our contribution reduces to the testbed and the identifiability map. We will say so. We expect the learned prior to help in two places. One is speed: the posterior is reused across cells rather than requiring a new MCMC chain per cell, which matters for screening at the end of the production line. The other is realism: correlated heterogeneity patterns, such as aging concentrated at the tabs, are hard to encode in a generic prior. The ablation that replaces the learned prior with a smoothness prior inside the same sampler settles which of these effects is real.
Why not just train a network on simulated data to regress the internal map? A regression trained with squared error returns the average of all internal maps consistent with the signal. When the inverse problem is ambiguous, that average is a smeared map, and smearing is exactly the masking we are trying to undo. The regression also gives no indication of when the answer is unidentifiable, and it extrapolates silently on faults outside its training library, the same failure mode that biology documents for reference-based projection (Johansen et al.). The out-of-library fault experiment below will show directly whether this concern is borne out.
Why not just track a single heterogeneity index, such as the spread between thermocouples? A single index might be enough to decide whether a cell needs attention. Chai et al., however, show that averaged module temperature hides slow fault evolution until the pressure relief threshold near 141 °C, after which propagation follows. In biology, Ha et al. find that an aggregate burden score loses to subclonal structure. We will include the thermocouple-spread index as a detector in the fault-lead-time comparison. If it matches our lead time, the reconstruction is unnecessary for detection, though it would still be needed for distinguishing plating from SEI growth.
Proposed Experiments
The first experiment is in simulation. It uses blade-cell and NCM-cell libraries with held-out heterogeneity patterns. Metrics are regional state-of-charge and temperature RMSE, current-share error, accuracy in locating the hot spot, Energy Score, and interval coverage. These are reported separately for uniform and concentrated patterns, and for each sensor layout.
The second experiment is the decisive physical check. It is a battery analogue of the titrated RNA mixtures in CellBench, the benchmark on which scProjection reached 87% accuracy and CIBERSORTx 61% (Johansen et al.). We will connect small cells in parallel into bundles with deliberately mismatched properties: different aging histories, different starting charge, and one cell held in a cold bath. Each branch has a current shunt and a thermocouple that the inference never sees. The bundle exposes only its total terminal signals and two surface sensors, and the hidden branch measurements provide ground truth. The test is whether the method recovers branch currents and temperatures, and whether it correctly labels a bundle as having concentrated versus uniform aging when both have the same total capacity loss.
The third experiment is the novelty test. We introduce a resistive internal-short branch whose type never appears in the library and measure how far ahead of terminal-signal thresholds the residual channel flags it. A positive result in the second and third experiments, with the MCMC baseline not already matching it, would show that biology's unmixing logic transfers to batteries and adds something there. A negative result in the second experiment, with posteriors that stay wide under realistic sensors, would be equally useful: it would quantify how much internal heterogeneity is fundamentally invisible from outside the cell, and would make a measured case for embedded sensors.
References
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- Comprehensive Modeling and Safety Protection Strategy for Thermal Runway Propagation in Lithium-Ion Battery Modules under Multi-Factor Influences
- Projecting clumped transcriptomes onto single cell atlases to achieve single cell resolution
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Study on the Homogeneity of Large-Size Blade Lithium-Ion Batteries Based on Thermoelectric Coupling Model Simulation
- Efficient Approach for Electrochemical-thermal Coupled Modeling of Large-format Lithium-ion Power Battery
- Modeling and Simulation of Thermal Effects on Electrical Behavior in Lithium-Ion Cells
- The Dilemma of C-Rate and Cycle Life for Lithium-Ion Batteries under Low Temperature Fast Charging
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Thermal Runaway Temperature Prediction of Lithium‐Ion Battery Under Extreme High‐Temperature Shock Using Experimental and Virtual Data