Meet the model that reads a battery's fate to a fifth of a percent

Every battery in your phone, laptop, and electric car is slowly dying — chemically, usually gracefully, but dying nonetheless. The question is not whether it will fade, but how fast, how predictably, and whether we can catch the warning signs early enough to do something about it. A new statistical framework suggests we may be able to predict a battery's decline to within a fifth of a percent — a precision that could reshape how we build, certify, and recycle the energy storage that underpins the clean-energy transition.

The work, from Mark Cary and Charles Bokor (Cary & Bokor, 2026), is not about inventing a better battery chemistry. It is about inventing a better way to read the batteries we already have. The researchers modeled the ageing of ten lithium-ion cells cycled in a controlled laboratory, and their approach predicts state-of-health — the technical term for how much usable capacity a battery has left — to within $\pm 0.191\%$. That number may sound like noise. In a field where battery lifetimes are measured in years and thousands of charge cycles, a prediction error smaller than one fifth of one percent is the difference between confidently telling a driver "your battery has 90% of its original range" and hedging with a guess.

The real breakthrough, though, is architectural. Most battery-ageing models treat each cell as an isolated island, modeling its decline curve on its own. But cells in a laboratory — or on a production line, or in a vehicle pack — are not isolated. They share manufacturing tolerances, testing conditions, and environmental quirks. Ignoring that shared structure, the authors argue, throws away information and produces models that are precise on paper but brittle in practice. Their model acknowledges the obvious: some variation lives within a single cell's ageing history, and some lives between cells. Formalizing that distinction is the heart of the paper.

The Science

To understand what Cary and Bokor did, it helps to know what standard practice looks like. Battery researchers typically age a batch of cells, recording how state-of-health ($SoH$) drifts downward over repeated charge and discharge cycles. They then fit a curve — often a simple mathematical function — to each cell's trajectory. The problem with this approach is that it treats the ten fitted curves as ten unrelated problems. Each fit gets its own parameters, its own errors, and its own confidence intervals, as though the cells had nothing in common.

The repeated-measurements framework rejects that framing. It models all the cells simultaneously, sharing information across them. Each cell's ageing profile is described by a power law — a curve of the form $SoH\propto t^{\beta }$, where $\beta$ captures how quickly capacity degrades. But crucially, the parameters of that power law are not fixed constants. They vary from cell to cell, and that variation is itself modeled — here, using a single-knot cubic B-spline, a flexible mathematical curve that lets the ageing parameters drift smoothly as the cells accumulate cycles.

The resulting model is hierarchical: it has a bottom layer (the individual cell ageing curves) and a top layer (the distribution of ageing parameters across cells). Statisticians call this a two-component-of-variance model, because it explicitly splits the variation in the data into two sources: measurement noise within a profile, and genuine test-to-test or cell-to-cell differences among profiles. This is not a subtle philosophical point; it is a practical one. If you fail to separate those two sources of uncertainty, you cannot tell whether a prediction failure comes from measurement error or from a genuinely different battery — and worse, you tend to underestimate your own uncertainty.

The training data was clean and tightly controlled: ten cells aged at various constant discharge and charge current cycles inside a chamber held at a fixed 25°C. Each cell was driven through its paces repeatedly, and its state-of-health recorded along the way. The modelling then asks a pointed question: given this structure — within-cell noise and between-cell variation — what is the most honest way to estimate every parameter at once?

What They Found

The headline result is the prediction accuracy. Across the featured working range of $SoH\in [0,20]$ (a normalized scale where the model's domain is bounded), the model predicts state-of-health to within $\pm 0.191\%$. To put that in context: a typical commercial lithium-ion cell might be retired when it reaches about 80% of its original capacity. A prediction error of a fifth of a percent means that, for a battery right at the retirement threshold, you can say with real confidence whether it has crossed the line.

The paper's key figures illustrate the model's behavior in ways that matter for real engineering. Figure 5, for instance, shows the fitted ageing trajectories against the raw data — the model's smooth curves running through the scattered measurements of each cell's capacity fade — demonstrating how the power-law-plus-spline structure tracks the observed decline. Figure 7 takes this further, showing how the hierarchical model pools information across cells so that a single noisy cell's fit is pulled toward the batch's collective behavior rather than trusted blindly. And Figure 8 displays the residuals — the differences between predicted and observed state-of-health — across the ageing range, with the scatter hugging zero tightly, visually confirming that 0.191% figure rather than just asserting it.

The innovation in estimation is worth pausing on. The authors used what they call regularised iterative generalized least squares — a mouthful that translates roughly as: go back and forth between estimating the ageing parameters and estimating their variation, using the results of each step to refine the other, and apply a penalty that keeps the model from overfitting to noise. This is not a one-shot calculation. It is an iterative dance, and each round re-estimates its own "hyper-parameters" — the dials that control how aggressively the model regularizes — searching for the settings that best match the data's actual complexity.

The result is a model that is statistically honest in a way many black-box machine-learning approaches are not. Deep neural networks can achieve stunning accuracy on battery data, but they are notoriously opaque: it is hard to say why a network predicts a certain trajectory, and harder still to attach an uncertainty to that prediction. The Cary–Bokor model has clear, interpretable structure at every level: a power law for ageing, a spline for parameter drift, and explicit variance components separating measurement noise from cell-to-cell differences.

Two Components of Variance in the Ageing Model

Label	Value
Within-profile (measurement noise)	1
Between-profile (cell-to-cell)	1

Sources of variance modeled explicitly (conceptual, values indicative of structure, not fitted magnitudes).

State-of-Health Prediction Error

Label	Value
Model prediction error	0.191
Typical BMS SoH error (comparison)	2

Model error per the paper (±0.191%); comparison value is an order-of-magnitude benchmark, not a figure from the paper.

Study Setup at a Glance

Label	Value
Training cells	10
Temperature (°C)	25
Pages in paper	17

Experimental parameters as stated in the paper.

Why This Changes Things

Battery ageing is, in a real sense, the silent bottleneck of the energy transition. Solar and wind are intermittent; storage is what smooths them out. Electric vehicles are only as practical as their batteries' lifetimes. Grid-scale storage is only economical if batteries degrade predictably over thousands of cycles. And behind all of that sits the humble fact that, until recently, much of our understanding of battery ageing has been empirical and piecemeal.

This framework matters for three concrete reasons.

First, manufacturing quality control. When a battery factory produces cells, it wants to know not just whether each cell meets spec at birth, but whether the population of cells will age consistently. The two-component-of-variance structure is directly relevant: a manufacturer can estimate how much variation is just measurement noise and how much reflects genuine cell-to-cell differences in quality. If a batch shows inflated cell-to-cell variation, that is a manufacturing red flag — and the model makes that signal measurable.

Second, battery management systems (BMS) in deployed vehicles and grids. A modern EV's BMS constantly estimates its battery's state-of-health to decide how far the car can go and how fast it can charge. Better models with honest uncertainty translate directly into more usable range, safer charging, and longer effective lifetime. A prediction error of 0.191% is not just academically tidy; it means the BMS can trust its estimate and squeeze extra performance from the same hardware.

Third, second-life and recycling decisions. As EV batteries age out of automotive service, they often retain 70–80% capacity and can be repurposed for stationary storage. Whether a used battery is worth repurposing — and what its remaining lifetime will be — depends on precisely the kind of precise ageing prediction this model enables. Getting the remaining-life estimate right by a fraction of a percent could be the difference between a profitable second life and an unnecessary trip to the recycling plant.

Parts of this picture belong in a definite and careful context. The model was trained in a laboratory at a fixed 25°C with constant currents — conditions that are far more controlled than the real world of cold starts, fast charging, and stop-and-go traffic. Real cells are aged by temperature swings, charging regimes, depth-of-discharge patterns, and calendar time, not just cycle count. The authors openly position this work as a methodological demonstration on controlled data, not a field-ready BMS module. But the layered structure — within-profile noise, between-profile variation, interpretable ageing curves — is exactly the kind of foundation that can absorb more complexity in real deployments.

What's Next

The most exciting direction is obvious: extending the framework from a fixed temperature and constant currents to the messy reality of field operation. Real batteries see temperature gradients, current spikes, and partial cycles. The hierarchical structure here is well-suited to absorbing those effects — instead of a single spline controlling ageing parameters, one could imagine covariates for temperature or depth-of-discharge entering at the between-cell layer. The model's logic generalizes; it wants more structure, not less.

Another frontier is uncertainty calibration at the extremes. The model is validated most confidently across the $SoH\in [0,20]$ working range of its age metric. Real life asks harder questions: what happens near end-of-life, where ageing often accelerates? Does the two-component structure hold when cells begin to fail catastrophically — dendrite growth, internal shorts, sudden capacity cliffs — rather than through the graceful power-law decline of healthy cycling? Those are open empirical questions.

Then there is the question of scale. Ten cells is a rigorous demonstration, but a battery pack can hold hundreds of cells, and a grid site can hold thousands. The computational cost and statistical behavior of this estimation machinery as the batch size grows — and as the between-cell structure gets richer — is a natural next test. If the framework scales, it becomes a practical tool for whole-fleet battery analytics rather than a lab instrument.

What makes this work feel like a genuine step forward is not any single number, impressive as 0.191% is. It is the willingness to model the data the way the data actually look — with two distinct sources of variation, shared across cells, using interpretable curves rather than opaque black boxes. Battery science has long been caught between simple-but-honest statistical models and powerful-but-inscrutable machine learning. This paper argues, convincingly, that there is a middle path: a model structured enough to trust, flexible enough to fit, and honest enough about its own uncertainty to be useful in the field.

As the world leans harder on batteries — to power our cars, buffer our grids, and store our sunshine — the ability to read a battery's fate early and precisely becomes infrastructure in its own right. Predicting ageing to a fifth of a percent does not wake anyone up with a headline. But it quietly extends the working life of every cell we own, sharpens every buy-and-recycle decision, and makes the transition to clean energy just a little less contingent on the capricious decay of electrochemistry. Sometimes the most consequential science is the kind that makes the boring, everyday technology last noticeably longer.