Real-Time Prediction of Hydrogen Explosions Involving Fuel Cell Vehicles Using a Physics-Informed Transformer Model.

Hydrogen fuel cell vehicles are being increasingly introduced as low-carbon transport systems. Nevertheless, unintended hydrogen release and ignition may still cause severe explosion hazards, especially in densely built urban settings. Conventional data-driven methods may have difficulty representing obstacle-induced spatial interactions and the transient development of explosion fields. To address this issue, this work develops a physics-informed Transformer framework for real-time prediction of hydrogen explosion scenarios associated with fuel cell vehicles. In this study, "physics-informed" refers to a flame-informed regularization mechanism rather than a conventional residual-based physics-informed neural network. Specifically, the intrinsic coupling between flame propagation and overpressure evolution is embedded into the Transformer learning objective, enabling the model to learn the relationship between flame-front development and blast-wave evolution instead of merely fitting isolated pressure-field snapshots. A high-fidelity benchmark database was generated through CFD simulations of hydrogen explosions in representative urban blocks. Based on this database, the optimal weighting coefficient was identified as λ = 0.001. Under this configuration, the model achieved an R 2 of 0.9385 and an MSE of 1.28 × 10-4 for blast-wave prediction and an R 2 of 0.9455 and an MSE of 9.58 × 10-4 for peak-overpressure prediction while requiring only 2.33 s for inference. The proposed approach offers a computationally efficient surrogate for rapid consequence assessment and provides technical support for digital-twin-based urban hydrogen safety management.