Round 2 external capacity validation: plan fixed before downloading/evaluating added cell trajectories
Written 2026-10-06. This is a prospective analysis plan for this revision, not a preregistered study.

Source eligibility: First inspect the original UCL INR18650 MJ1 EIL series for other genuine independent electrochemical cycling records. Imaging records and segments of EIL-015 are not additional cells. If no further eligible trajectories are documented, use the original NASA battery aging experiment containing B0005, B0006, B0007 and B0018; include all four published cells in that named group, never select cells based on fit quality. Include UCL EIL-015 separately as the previously inspected reference; it is not fresh external validation.

Target: reported discharge capacity in Ah versus ordinal discharge-cycle number. Raw files remain unchanged; save download URL, publisher metadata, license and SHA-256 hashes. No external code will be executed. NASA MAT structures will be parsed with locally authored code. Do not infer smartphone chemistry or hardware validation from laboratory cylindrical cells.

Split: For every eligible cell with at least 80 valid capacity observations, take the earliest floor(0.75*n) discharge cycles for calibration and all remaining cycles for temporal testing. No other cutpoint will be selected based on outcomes. Preserve finite positive capacities without smoothing or outlier removal; document missing/invalid capacities. Cycle index counts original discharge events, even when a capacity observation is invalid. UCL's 400-cycle record therefore uses 300/100 as in round 1. Never fit using another cell's future record or combine rows into a random split. This is repeated within-cell forward extrapolation, not a model trained on some cells and deployed unchanged to new cells.

Models fixed before added-data evaluation: (1) nonnegative dual exponential Qinf+A exp(-aN)+B exp(-bN); (2) nonnegative single exponential Qinf+A exp(-aN); (3) global ordinary least squares linear trend; (4) persistence at final calibration capacity; (5) ordinary least squares local linear trend fitted to final max(20,ceil(0.2*n_train)) calibration observations. Local trend is a simple recent-rate extrapolation baseline; no window tuning. Exponential amplitudes/Qinf bounded [0,4] Ah and rates [1e-6,1] per cycle, matching round 1. Use the same 4 single/27 double starts and least_squares tolerances as round 1. Report boundary solutions and optimizer status. No forecast clipping to physical capacity bounds: negative long-range predictions, if any, are visible model failures.

Metrics: Per cell and model held-out MAE, RMSE, signed mean bias in Ah; normalized RMSE and MAE relative to first observed capacity; held-out R2 (undefined if zero target variance). Report test lengths and first/end capacity. Main multicell summary gives equal weight to cells, not rows. Report every model/cell result, including failures and worse baselines. No p-values across correlated time points and no pooled R2 masquerading as cross-cell generalization. Original UCL result is tabulated separately from external NASA aggregate.

Figure: one standalone 6.2-inch scientific figure, per-cell observed capacity and forecast curves with the 75% split, plus an across-cell normalized error panel if legible. Derived CSV/JSON and local analysis script accompany the result. Interpretation: evaluate robustness of short-range capacity extrapolation; do not equate capacity prediction with SOC/TTE prediction, temperature adaptation, or identified degradation mechanisms.
