Methodology
Graded against a truth it cannot see.
Ionscore's estimator never observes true state of health. It reads observable BMS signals, predicts a P10 to P90 band, and is then graded against the hidden truth on packs it has never seen. This page is the published result: the model card, error bars included.
- Model
- soh-quantile-1.0.1
- Generated
- 2026-07-12
- Evaluation split
- by pack, three-way
Headline metrics
SoH mean absolute error
1.58pp
held-out, overall
Root mean squared error
2.00pp
held-out, overall
P10-P90 band coverage
80.0%
target near 80
Held-out evaluation set
149packs
9,089 snapshots, no pack overlap
Read this first
Measured on synthetic fleets.
These metrics are measured on synthetic fleets grounded on public battery aging data, so the realism ceiling is set by the simulator, not by field telemetry. The value of this card is the method: the estimator is graded against a latent truth it never sees, the split is by pack, and the band calibration is reported honestly rather than assumed.
Error
Where the error lives.
Held-out error and band coverage, split by cell chemistry and by vehicle segment. The split is by pack: no pack's history appears in both training and test.
By chemistry
| chemistry | N | MAE pp | RMSE pp | Coverage % |
|---|---|---|---|---|
| LFP | 6,466 | 1.41 | 1.77 | 80.8 |
| NMC | 2,623 | 1.99 | 2.47 | 78.1 |
By segment
| segment | N | MAE pp | RMSE pp | Coverage % |
|---|---|---|---|---|
| 2W | 4,148 | 1.43 | 1.82 | 82.7 |
| 3W | 2,318 | 1.78 | 2.21 | 80.3 |
| 4W | 1,159 | 1.98 | 2.40 | 69.7 |
| Commercial | 1,464 | 1.35 | 1.76 | 80.1 |
MAE and RMSE in state-of-health percentage points against the hidden simulator truth. Coverage is the share of truths falling inside the pack's P10 to P90 band.
Calibration
Does the band mean what it says?
One-sided reliability: the share of held-out truths at or below each predicted quantile. A calibrated model tracks the nominal level; after conformal widening, Ionscore runs slightly conservative at every quantile.
One-sided coverage
| Quantile | Nominal percent | Empirical percent |
|---|---|---|
| P10 | 10.0 | 12.6 |
| P50 | 50.0 | 53.4 |
| P90 | 90.0 | 92.6 |
Share of held-out truths at or below each predicted quantile, on a 0 to 100 scale. Empirical slightly above nominal means the band errs on the safe side.
Mean band width
5.08pp
Average P90 minus P10. Honest coverage is cheap with a wide band; this is the width the 80 percent coverage above is achieved at.
Conformal widening added 0.509 SoH points to the raw quantile band.
Training configuration
The setup, reproducibly.
- Split
- By pack id, three-way (train / calibration / test), mutually disjoint
- Training packs
- 345
- Calibration packs
- 106
- Test packs
- 149
- Training rows
- 21,045
- Test fraction
- 0.25
- Calibration fraction
- 0.15
- Model
- HistGradientBoostingRegressor (quantile loss) for P10, P50, P90
- Band calibration
- Conformalized quantile regression (CQR) on the calibration packs
- Conformal widening
- 0.509 SoH points
- Seed
- 42
Run this on your packs.
A pilot scores a sample of your fleet end to end, with the same published error discipline.