Interpretable Prediction of Thermal-Abuse Response Metrics of Cylindrical Li-Ion Cells with Leakage-Free Validation and Calibrated Uncertainty

Thermal-abuse response metrics of lithium-ion cells are measured one destructive test at a time, and the conditions that produced them are often reported incompletely, so values from different laboratories are hard to compare.This work asks how far three such metrics of cylindrical cells can be predicted from pre-test information, and which part of that predictability is an artefact of the test protocol.Total heat output, the average cell temperature at the instant the abuse was initiated, and mass ejection came from 365 fractional thermal runaway calorimetry records of the NREL Battery Failure Databank V2 [1,2,8] (310 18650 and 55 21700 cells; nail, heater with internal short circuit, and heater-only triggering) and regressed with gradient boosting on 28 pretest features.Imputation and scaling were fitted within each fold; accuracy was compared with available-capacity and stored-energy linear baselines under record-level and repeated cell-modelgrouped cross-validation, with paired cluster-bootstrap intervals over the 18 cell models, and prediction intervals were built by split conformal calibration with the finite-sample corrected quantile over ten grouped partitions.Record-level cross-validation gave R 2 = 0.931 for heat, 0.801 for the trigger temperature and 0.601 for mass ejection; grouping by cell model reduced these to 0.823, 0.693 and 0.329, with paired gains over the linear proxy of +0.123 [0.094, 0.189], +0.694 [0.534, 1.037] and +0.442 [0.265, 0.841].The decisive result is negative in kind: a trigger-only model already reaches R 2 = 0.783 on the temperature column while withinmechanism models reach 0.134, -0.063 and 0.209, and all 56 sub-30°C values are nail tests.Conformal coverage was 0.795 [0.764, 0.823] for heat and 0.684 [0.646, 0.722] for the trigger temperature at the 80% nominal level, so the intervals are approximately calibrated for heat and anti-conservative for temperature rather than uniformly conservative.Transfer between formats was asymmetric, and adding three to ten target-format records raised mass ejection from -0.368 to 0.087-0.208.These results bound what such models can mean, and they yield seven minimum DOI: 10.