Why similar batteries fail differently: an empirical analysis of thermal runaway severity variability

Abstract Thermal runaway tests of nominally similar lithium-ion cells can yield substantially different consequences. We examine severity variability using 365 tests from an open battery failure databank and an empirical Thermal Runaway Severity Uncertainty Amplification (TR-SUA) descriptor, defined as the ratio of post-failure to recorded pre-test coefficients of variation (CVs), not as a measured physical gain. Forty-six near-replicate groups (349 tests) were defined by recorded cell description, format, and trigger ( n ≥ 3). The median available three-descriptor input CV was 0.104%, versus output CVs of 7.577% for corrected total energy, 14.730% for ejecta energy, 21.943% for unrecovered mass, and 21.903% for unrecovered-mass fraction. With a 1%-point input-CV floor, median ratios ranged from 7.577 to 21.943 and remained sensitive to alternative input definitions and floors. Exhaustive within-group log decomposition identified 1,564 valid test pairs; among 1,094 pairs with at least 10% ejecta-energy contrast, the absolute EPR term exceeded the total-energy term in 58.0%, which describes an algebraic partition, not mechanistic causality. Cluster-bootstrap P95 intervals were appreciably broad, and model-group/format holdouts revealed poor predictive transfer despite high random-fold classification AUC. The results support reporting severity dispersion and uncertainty-qualified upper-tail outcomes, while unobserved cell/protocol conditions prevent a causal physical-amplification claim.