A physics-informed model, ranked against five industry-standard architectures.
Same data. Same held-out cells. Same evaluation code. Zylectra's Physics AI Model for cell electrochemical health, scored against five architectures representative of how the industry typically builds AI-based battery health estimators.
The lowest error rate, at a fraction of the size.
Zylectra's Physics AI Model for Cell Electrochemical Health was benchmarked against five deep-learning architectures representative of how the battery-analytics industry typically builds AI-based health estimators: an MLP, a Transformer-style Attention model, a GRU, an LSTM, and a residual CNN. All six were evaluated on the exact same held-out cells, with the exact same scoring code. Zylectra's model ranked first, at 1.08% MAE, making 7% fewer errors than the closest competitor and up to 39% fewer errors than the weakest model tested, while using 13–27× fewer parameters and a 10–23× smaller checkpoint.
This benchmark compares Zylectra against the class of standard, purely data-driven models the industry reaches for when it isn't physics-constrained. It does not claim a win over proprietary commercial BMS firmware: those algorithms are closed-source, and no vendor publishes accuracy numbers to compare against.
Before extending the model suite, we wanted an honest, reproducible answer to one question: how does a physics-constrained approach actually compare to the standard data-driven approach most of the industry uses, on identical data, under an identical evaluation protocol? No cherry-picked splits, no home-field advantage for either side.
Ranked by how many fewer errors each model makes.
Error reduction vs. the weakest model tested, on 3,769 held-out cycles across 13 cells, identical for every model
The most accurate model on the leaderboard is also the smallest.
Parameter count · lower is better for edge and OTA deployment cost
No new hardware. No new sensors.
Every model in this report, Zylectra included, runs on the same telemetry most BMS systems already log every cycle: 5 raw signal types, encoded into 19 engineered features. No model gets richer input than another. Where they differ is what each one does with that input:
Error-rate metrics, translated to operating cost.
A reading you can act on
Lowest error rate of six models on identical cells means fewer decisions made on a wrong number: the difference between scheduling a swap and getting a roadside failure.
No blind spot at low health
The strongest baseline shows a clustering failure mode where predictions floor out for a subset of low-health cycles. Zylectra's evaluation shows no equivalent pattern: fewer batteries wrongly read as "fine."
Cheaper to ship, at fleet scale
A checkpoint 10–23× smaller means real bandwidth and storage savings on every OTA push and every edge device in the fleet.
Illustrative: on a 10,000-device OTA push, that's ~8.9 GB less payload than the average baseline checkpoint size. Arithmetic on the real checkpoint sizes above, not a customer figure.
Consistent across the health range and across battery life.
No saturation cluster
Zylectra's predictions track true electrochemical health evenly across the full range. The strongest baseline shows a visible cluster of predictions that floor out near a fixed value for a subset of low-health cycles: a failure mode Zylectra's evaluation does not show.
Steadier through late-life fade
Baseline models' error grows disproportionately in the steep late-life fade region, where the signal gets noisier. A physics prior built into Zylectra's training structurally discourages predicting health increases cycle-over-cycle, exactly the kind of physically-impossible wobble a purely data-driven model can produce late in life.
Evaluated on a second, independent dataset
As part of its normal held-out evaluation, Zylectra's model is also scored on a second, independent cycling dataset: 2.10% MAE, R² 0.871. None of the baseline architectures in this report have a published, populated benchmark on that dataset to compare against. Zylectra is trained as one unified model across both sources rather than one model per data source.
Two orders of magnitude fewer parameters.
Same evaluation hardware (T4 GPU) across every model
| Model | Params | Checkpoint | Train time |
|---|---|---|---|
| Zylectra | 14,202 | 64 KB | 2.1 min* |
| Baseline · MLP | 182,785 | 718 KB | 39.7 s |
| Baseline · Attention | 376,329 | 1.46 MB | 3.7 min |
| Baseline · GRU | 169,089 | 664 KB | 21.0 s |
| Baseline · LSTM | 219,265 | 860 KB | 44.0 s |
| Baseline · CNN | 321,745 | 1.24 MB | 27.4 s |
* Estimated. Zylectra's original training run was measured on a different GPU generation; the figure shown here is scaled to T4 for a like-for-like row, not a re-measured benchmark. All six models train in single-digit minutes either way; the numbers that matter for OTA size and edge deployment are parameter count and checkpoint size.
One benchmark. Four different stakeholders.
Same benchmark, same fleet, walked through for the four people who each look at a battery differently: what actually happens, what it costs, and the arithmetic behind the number.
Catch the failing battery before it strands a vehicle.
The weakest baseline model we tested has a specific blind spot: for a slice of batteries that are already unhealthy, it keeps reporting them as fine, right up until they fail. Zylectra's evaluation shows no equivalent pattern.
A battery that fails without warning means a stranded vehicle, a recovery call, and a rushed replacement instead of a swap you scheduled in advance.
Worked example500-pack LFP fleet. Figures are arithmetic on the assumptions shown in each tab, applied to this report's real error-rate and reliability findings, not results from a deployed fleet.
On identical data, identical held-out cells, and identical evaluation code, Zylectra's Physics AI Model beats five architectures representative of the industry's standard data-driven approach: lowest error rate of six, at 13–27× fewer parameters and a 10–23× smaller checkpoint. The win isn't just a lower average: Zylectra shows no equivalent of the strongest baseline's low-health clustering failure, and tracks late-life degradation more faithfully. Physics-informed modeling isn't a nice-to-have architectural choice; on this evaluation, it's the more accurate approach at a fraction of the size.
Have battery data and a problem worth solving?
Talk to us about what physics-informed health scoring looks like on your fleet.
Zylectra · True electrochemical health benchmark · LFP chemistry← All benchmark reports