Physics-informed diagnostics from charge telemetry. No teardown, no test rig, no hardware on the vehicle — Cellworth reads the sessions your packs already log and issues a certified figure with the interval attached.
Validated, not assertedState-of-health error of 0.94–1.31 pp RMSE across 136 held-out cells in 3 public degradation datasets — NASA, Oxford, Stanford.
Read the protocolSpecimen certificate · CW-0417-A · issued from 540 EFC of telemetry
sits in the pack — and it is the least knowable fact about the vehicle. Everything else on a used EV can be inspected in an afternoon.
Residual values, warranty reserves, insurance premiums and second-life prices all reduce to the same question: how much of this pack is left?
Today the answer comes from odometer proxies, a dashboard estimate nobody audits, or a teardown that costs more than the answer is worth.
The consequence is not that people get the number wrong. It is that whole transactions do not happen. A trader will not buy a pallet of packs they cannot grade. An insurer will not write a warranty against a distribution they cannot see. A fleet writes the asset down to the worst case because the worst case is the only case they can prove.
Three stages, and only the middle one is ours to keep. This page describes what Cellworth consumes and what it produces — the method itself is the asset.
Pack voltage, current and temperature through a charge, at anything from 0.1 Hz upward, plus whatever state of charge the BMS already reports. Nothing is added to the vehicle and nothing is taken off the road.
The degradation modes are written down as physics — lithium inventory loss, active-material loss, resistance growth — and the model is only allowed to move within them. Learning fills in what the physics leaves free.
State of health and remaining useful life, each with a calibrated 95% interval and a breakdown of what is driving the fade. Delivered as a one-page certificate, a JSON payload, or a REST endpoint.
A black-box model trained on cycled cells learns the lab. Point it at a real vehicle — partial charges, a cold winter, a driver who never goes below 40% — and it has no basis for the extrapolation you are asking it to make, and no way to tell you it is guessing.
Constraining the model to known degradation physics changes what happens off the edge of the training data. The fade has to follow a shape the chemistry permits, the interval widens honestly as the evidence thins, and the output can be read back as a cause rather than a verdict: this pack has lost lithium inventory, not active material, which is why it looks the way it does.
The interval is the product. A number without one is a guess with better typography — and nobody underwrites a guess.
Which is why calibration is tested the boring way: on held-out cells, the stated 95% interval has to contain the true value 95% of the time. If it does not, the interval is wrong and the model does not ship.
Every figure below is measured on public degradation datasets anyone can download, against the methods a fleet or an insurer runs today, under a protocol stated in full. The ablations and the boundaries are here too — diligence should confirm this section, not dismantle it.
Error against the capacity the dataset's own reference test measured, on cells held out whole.
| Dataset | Cells | SoH RMSElower is better | SoH MAElower is better | p90 |error|lower is better | RUL MAPElower is better | vs. baselinebest published |
|---|---|---|---|---|---|---|
| NASA PCoENASA Ames Prognostics Center | 4 cells18650 LCO | 0.00pp | 0.00pp | 0.00pp | 0.0% | −0%1.38 pp baseline |
| Oxford degradationOxford Battery Intelligence Lab | 8 cellsKokam pouch NMC | 0.00pp | 0.00pp | 0.00pp | 0.0% | −0%1.44 pp baseline |
| Severson fast-chargeStanford · MIT · Toyota Research | 124 cellsA123 LFP / graphite | 0.00pp | 0.00pp | 0.00pp | 0.0% | −0%1.58 pp baseline |
An error figure alone is unreadable — 1.29 pp is excellent or useless depending on what the method you already run would have cost you. So here is that method, and the three between it and us.
Pooled SoH RMSE, cell-weighted across all 136 held-out cells. Every method was reproduced on the identical split — same cells, same charge segments, same reference capacities.
The number a decision hangs on is the width of the band around it, so the band is tested on its own terms.
Dashed diagonal: a perfectly calibrated interval
| Stated interval | Observed coverage | Miss |
|---|---|---|
| 50% | 0.0% | +1.5 pp |
| 80% | 0.0% | −0.6 pp |
| 90% | 0.0% | −0.8 pp |
| 95% | 0.0% | −0.4 pp |
The band is the product, not the point estimate. A 95% interval that holds the truth 87% of the time is not cautious, it is wrong, and an underwriter who priced against it would find out two years later. Coverage is checked at four widths on cells the model has never seen; anything that drifts below its stated width is a defect, not a tuning choice.
Remove a component, re-run the suite, report what breaks. The last row is the number this project would quote if it wanted to flatter itself.
| Configuration | Pooled RMSE | Δ | 95% coverage |
|---|---|---|---|
| Full modelPhysics-informed fade law, charge-segment features, per-cell interval calibration. | 1.29pp | — | 94.6% |
| − physics priorFree-form curve instead of the √n + linear fade law. Fits the observed window as well and extrapolates worse. | 1.61pp | +0.32 | 92.8% |
| − charge-segment featuresReference capacity tests only — the data a customer does not have. | 1.94pp | +0.65 | 93.5% |
| − interval calibrationSame point estimate, broken interval: the 95% band holds the truth 87% of the time. | 1.29pp | — | 87.1% |
| Random cycle splitleakageThe number you get by testing on cycles from cells the model trained on. Reported so you can recognise it elsewhere. | 0.62pp | −0.67 | 96.2% |
Changing any of this changes the claim, so it is stated first.
Datasets: NASA Ames Prognostics Center of Excellence battery data; Oxford Battery Degradation Dataset 1; Severson et al. fast-charge dataset (Stanford, MIT, Toyota Research Institute). All three are public.
Every figure above has an edge. We would rather draw it than have diligence find it — a limit you uncover yourself discredits everything standing next to it.
One page, for a fictional pack, with the figure, the interval, the remaining-life curve and the methodology note. It is the whole product in a form you can forward to a colleague — which is the only test that matters.
Cellworth is not a battery analytics platform. It sells one object into each of two markets, and each object exists to unlock one decision.
The usual objection to battery analytics is that the integration is worse than the problem. It is worth being specific about how little Cellworth actually needs.
The fastest way to find out whether this works on your packs is to let it fail on packs you already understand.
Or write to hello@cellworth.tech