Multi-fidelity surrogate modeling for lithium-ion battery pack thermal runaway propagation considering propagation occurrence in adaptive sampling

Thermal runaway propagation (TRP) in lithium-ion battery packs is a critical safety concern that demands accurate yet computationally efficient modeling. This paper proposes a multi-fidelity (MF) surrogate modeling framework with a tailored adaptive sampling strategy to predict TRP outcomes under varying battery pack design parameters, including the stacking angle, inter-cell gap, and composite phase change material (CPCM) mass fraction. The framework integrates low-fidelity (LF) 2D and high-fidelity (HF) 3D thermal simulations through an autoregressive formulation, employing Bayesian neural networks (BNNs) to capture predictions and uncertainties of the key outcomes: propagation occurrence, time to second ignition ( t 2 ), and total propagation time ( t total ). The discontinuity associated with propagation occurrence is explicitly handled through a dedicated classifier integrated into the adaptive sampling strategy, which prioritizes new HF simulation samples by jointly considering the predictive response via an upper confidence bound, predictive uncertainty, and propagation probability, thereby optimally allocating the limited simulation budget. Two case studies demonstrate the framework: a modified Forrester benchmark function and a TRP simulation case. In the TRP case, the framework achieves 90% classification accuracy for propagation occurrence, with RMSE values of 2.6497 and 19.5183 for t 2 and t total respectively, and corresponding calibrated prediction standard deviations of 14.0087 and 67.8687. Compared to pure HF surrogates and Gaussian process-based MF surrogates, the proposed model reduces prediction errors by approximately 20%. The adaptive sampling strategy yields more accurate surrogates than baseline methods for the same number of HF runs, and ablation studies confirm the contribution of each framework component. This approach offers a computationally efficient TRP modeling tool that can guide battery pack safety design by enabling fast and reliable TRP risk evaluation. • Multi-fidelity surrogate modeling framework designed for thermal runaway propagation. • BNN-based surrogate with uncertainty quantification for safety-critical TRP. • Tailored adaptive sampling strategy considering propagation probability.