Real-time hydrogen gas explosion prediction from fuel cell vehicles by physics-informed graph neural network

Abstract

Hydrogen fuel cell vehicles (HFCVs) generate electricity from hydrogen via electrochemical reactions and offer an efficient, environmentally friendly pathway to the dual‑carbon targets. Nevertheless, hydrogen leakage can rapidly mix with air to form large flammable vapor clouds. In congested urban areas, the resulting vapor cloud explosions may cause severe structural damage and significant economic losses. Machine learning techniques have been developed to predict the evolution of gas explosions. Nevertheless, accurately capturing the spatial relationships among obstacles and predicting explosion overpressure evolution remain challenging. This study proposes a physics-informed graph neural network model (Physics_ARGNN) for real-time blast wave prediction of hydrogen gas explosion from HFCV in congested urban environment. Computational fluid dynamics (CFD) simulations are conducted to model hydrogen gas explosions within urban blocks and to build a benchmark dataset. Using this dataset, the optimal physics factor λ = 1 × 10 −3 was determined, enabling the proposed Physics_ARGNN to achieve an R 2 of 0.9556 and an MSE of 6.94 × 10 −5 for blast wave prediction, and an R 2 of 0.9806 and an MSE of 1.80 × 10 −4 for peak overpressure prediction, outperforming state-of-the-art models while maintaining real-time capability.