Towards physics-guided graph neural network for hydrogen gas explosion simulation at urban scale

As hydrogen and fuel cell technologies become integral to the urban energy transition, their widespread adoption in densely populated areas necessitates robust safety measures. The rapid upscaling of hydrogen energy applications introduces risks associated with accidental hydrogen gas explosion, resulting in substantial blast loads and posing catastrophic threats to both structures and people. Machine learnings have been employed to efficiently evaluate the consequence of obstructed gas explosion, which however exhibits poor accuracy in blast load dynamics prediction due to the lack of considering the interactions between congestion, flame propagation, and blast wave dynamics. This paper aims to develop a physics-guided graph neural network approach, termed Physics_GNN, to simulate the dynamics of hydrogen vapor cloud explosion at the urban scale. The underlying physics of the interactions between congestion, flame propagation, and blast wave dynamics are integrated to enhance prediction accuracy regarding overpressure peaks and their arrival times. The OpenFOAM solvers, validated using public experimental data, are utilized to construct a benchmark dataset. Sensitivity analysis of the empirical coefficient affecting physical interaction is conducted. Results demonstrate Physics_GNN approach achieves a higher prediction accuracy with an R2 of 0.97 compared to the state-of-the-art deep learning models. Additionally, it offers a 1000-fold computational speed-up compared to CFD model for simulating hydrogen gas explosions at urban scale. Physics_GNN approach has the potential to efficiently and accurately analyze the destructive effects of hydrogen gas explosions at urban scale, supporting decision-making to improve urban resilience in the context of the energy transition.