Predictions of realistic hydrogen deflagrations with graph neural networks

This study presents a novel application of Graph Neural Networks (GNNs) to predict pressure, temperature and velocity fields resulting from deflagration of hydrogen-air mixtures in complex, congested and confined environments typical of industrial operations, extending the MeshGraphNets framework to coupled combustion-turbulence physics. The objective is to enable rapid yet accurate explosion predictions for industrial safety assessments where traditional CFD is computationally prohibitive. Our study employs Unsteady Reynolds-Averaged Navier-Stokes (URANS) simulations in OpenFOAM to model the ignition and subsequent deflagration of premixed hydrogen-air mixtures in geometrically complex domains, effectively capturing the interaction between turbulent combustion and congested environments. We train the algorithm on full-physics URANS simulations enabling parameter-free predictions of complete flow fields including velocities. The computational domains comprise volumes with multiple rigid obstacles of random shape, position, and orientation, as well as several gas clouds. The results of these CFD simulations are used to assemble a training dataset for a surrogate model. A GNN architecture is used to learn and predict the evolution of thermodynamic quantities and flow fields initiated by the deflagrations. The model is trained on time histories of pressure, temperature, velocity, and additional parameters related to combustion, diffusion and turbulence. After training, the model demonstrates generalization capabilities, predicting explosion dynamics for unseen geometries. The model achieves computational time savings of up to three orders of magnitude (up to 6200×) compared to CFD solutions, allowing rapid safety assessments in the design of hydrogen systems and infrastructure.