Methodology & Validation

Two distinct accuracy numbers describe a battery degradation model and they are not interchangeable. We report both because the difference matters.

• Calibrated-fit RMSE — the residual between the model and a published retention curve after the model has been fit to that curve. Measures whether the engine can reproduce the data it has been calibrated against.
• Held-out forward-prediction RMSE — the residual between the model and the published curve when the model is calibrated using only the early portion of the data (e.g., cycles 0–500 at 100% retention down to 90%) and then extrapolated forward to predict the rest. Measures the metric that matters for cycle-life forecasting.

Calibrated-fit numbers are reported throughout the published literature and the marketing copy of most commercial degradation tools. Held-out forward-prediction is the standard since Severson et al. Nature Energy 2019 introduced it as the appropriate benchmark for cycle-life prediction. We publish both.

Each preset is a hand-tuned PhysicsConfig that fits the published curve. Calibration uses Nelder–Mead in log-space across 4–8 rate constants (SEI, Si-cracking, plating, transport). Six of eight are fully-attributable peer-reviewed curves; two are synthesized references (one proxy fit, one literature composite) and are labeled accordingly in the table below.

Summary (6 attributable + 2 synthesized): 6 A+ (RMSE 0.13–0.32 pp), 2 A on knee-prone cells (RMSE 1.01–1.23 pp). Mean RMSE across all 8: 0.44 pp. The two synthesized references (LG_M50T proxy, SiGr_5pct_45C_1C composite) reproduce literature-typical aging shapes for chemistries where the original cycler data was not publicly available at the time of build — they are useful for regression testing but should not be cited as independent validation against those specific cells. Re-sourcing real data for both is on the post-sprint backlog.

Our grade scheme uses the following thresholds, defined in model/battery_model_v18.py line 1247:

• A+: RMSE < 1.0 percentage points
• A: 1.0 ≤ RMSE < 2.0 pp
• B: 2.0 ≤ RMSE < 3.5 pp
• C: 3.5 ≤ RMSE < 5.0 pp
• D: RMSE ≥ 5.0 pp

Stricter conventions in some national lab reviews use A+ < 0.5 pp / A < 1.0 pp. By those thresholds, our calibrated count would be 3 A+ (NatComms_LPD 0.13, Kirk_Moderate 0.17, HPQ_GEN3 0.18) plus 3 A (Dose_2023 0.23, SiGr_5pct 0.25, LG_M50T 0.32) plus 2 B (Kirk_FastCharge 1.01, NatComms_HPD 1.23). All eight datasets remain within 1.5 pp of the published curves either way.

Held-out protocol: calibrate the model using only data up to the 90% retention cutoff, then run forward prediction to the 70% retention cutoff (typical end-of-warranty checkpoint). Compute RMSE on the held-out portion only.

The architecture sprint May 4–10, 2026 closed the gap on LG_M50T-class cells from 14.10 pp to 4.98 pp by expanding the fit-parameter set to engage v18.3's cathode LAM, electrolyte depletion, and R-overpotential coupling mechanisms (Track 2). The remaining held-out gap on knee-prone cells is the work for the N≥30 cross-cell validation panel beginning 2026-05-11.

Use the calibrated-fit numbers when evaluating whether the model can reproduce a curve you've already characterized — design-of-experiments sweeps, sensitivity analysis, mechanism decomposition on cells you have full data for.

Use the held-out forward-prediction numbers when evaluating whether the model can predict cycle life from limited early-cycle data — warranty modeling, fleet replacement-rate forecasting, BMS calibration from sparse field measurements. Linear-degradation cells are production-ready; knee-prone cells are active research.

The /api/calibrate response surfaces both numbers when you supply enough data to compute them. The dispatch_to field on the pre_screen endpoint flags cells that route to the identifiability-limited branch — those are the ones where the held-out gap matters.

Calibrated envelope: 2.5–20% silicon, 0.5C–2C charge, 25–45°C ambient.Outside this envelope every predict response includes a validation_warnings array with field-specific messages and an out_of_validated_range boolean.

Two known model gaps in the customer-visible warning:

• Cold-charge plating (T < 0°C): the model's plating mechanism does not activate at sub-zero charging. Real silicon-graphite cells plate severely in this regime. The validation warning correctly flags T < 25°C as outside the calibrated envelope; the underlying model behavior underestimates degradation here. Do not use for cold-climate cycle-life forecasting until this is resolved.
• Extreme fast charge (≥3C): mechanical fatigue and transport collapse activate but the magnitude is not validated against published >3C cycling data. Treat predictions as directional, not certified.

Every simulation run produces an AuditRecord with deterministic run_id (SHA-256 of config + protocol), config_hash, model_version, output_schema_version, timestamp, platform info, and invariant-check messages. Run the same config + protocol against the same model version at any point in the future and you get an identical retention curve.

Version retention policy: model versions are pinnable via the ?model_version= query parameter and remain callable for a minimum of 24 months after a successor version ships. See /sla for full details.