Benchmarked against the field's most cited physics-informed architecture.
Zylectra's Physics AI model for true electrochemical health, scored against Wang et al. 2024's PINN4SOH, published in Nature Communications and one of the most cited physics-informed architectures in the literature, chosen as the toughest available comparison, not a convenient one.
One model, beating two specialized ones.
Zylectra's Physics AI model was benchmarked directly against Wang et al. 2024's PINN4SOH, the architecture the research community treats as the reference point for physics-informed health estimation. Both models were scored on two public LFP cycling datasets, MIT and HUST, using each project's own held-out test split. Zylectra's single model beat Wang's dataset-specialized checkpoint on both datasets, on every metric measured: MAE, RMSE, and R².
Wang's PINN4SOH ships one specialized checkpoint per dataset. Zylectra trains one model across every dataset it supports, on principle: a fleet is never one dataset. This report tests whether that choice costs anything against a specialized alternative built by the reference implementation itself.
| Model | Type | Params | Training scope |
|---|---|---|---|
| Zylectra Physics AI model | Physics-informed NN | 14,202 | One model, both datasets |
| Wang PINN4SOH: MIT | Physics-informed NN | 13,662 | Specialized, MIT only |
| Wang PINN4SOH: HUST | Physics-informed NN | 13,662 | Specialized, HUST only |
Same architecture family, two different bets on generalization.
MAE and RMSE in health fraction (0–1 scale; 0.01 = 1 percentage point) · lower is better. R² · higher is better.
Wang's per-dataset specialization doesn't buy back what it costs: their two checkpoints together still trail Zylectra's one.
No specialization tax
Zylectra's single model matches or beats each of Wang's specialized checkpoints on its own dataset, without retraining per source.
Built for fleets, not single datasets
A real fleet is never one dataset. Zylectra's one-model architecture is the assumption a multi-source fleet product has to hold; this comparison is evidence it holds without a performance cost.
| Model | Params | Checkpoints | Total params shipped | Total disk |
|---|---|---|---|---|
| Zylectra Physics AI model | 14,202 | 1 | 14,202 | 64 KB |
| Wang PINN4SOH | 13,662 | 2 | 27,324 | 118 KB |
Wang's network is marginally smaller per checkpoint: the gap is Zylectra's wider input layer. But Wang's deployment needs to know which dataset a cell came from and load the matching checkpoint, or ship both. Zylectra ships one file, one code path.
- Zylectra's and Wang's held-out test cells are independently constructed splits over the same public data, not a literal identical cell-for-cell rerun.
- Zylectra's number is the best of a multi-seed training sweep; Wang's figure is a single run, matching how the original paper reports it. This asymmetry favors Zylectra's number being closer to its ceiling.
- No comparison against proprietary commercial BMS firmware exists or is obtainable outside those vendors.
Benchmarked directly against the architecture the field treats as its physics-informed reference, Zylectra's Physics AI model wins on both public datasets, on every metric, and does it with one unified checkpoint against Wang's two specialized ones. Physics-informed modeling built for a fleet, not a single dataset, doesn't cost accuracy to get there.
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 · vs. Wang et al. 2024 PINN4SOH← All benchmark reports