Machine Learning Modelling of Impinging Hydrogen Gas Leaks

Abstract The use of hydrogen as fuel in gas turbines (GTs) supports energy transitions but raises safety concerns regarding accidental fuel gas leaks within enclosures. Current GT enclosure safety designs rely on CFD simulations to predict fuel dispersion in leak scenarios and assess ventilation effectiveness in mitigating flammable clouds. However, the high computational cost of CFD limits the exploration of diverse leak conditions, including variations in position, orientation, leak size, gas composition, and interactions with nearby objects. Machine learning offers a promising alternative but often struggles with hardware constraints or geometrical changes that affect the CFD domain or grid resolution. This study introduces a novel machine learning approach to predict gas dispersion from leaks interacting with solid surfaces. The training dataset comprises parametric numerical simulations with varying gas compositions, leak areas, and distances from a flat surface. The model leverages a modified Neural Implicit Flow architecture with two neural networks: ShapeNet, which captures spatial flow complexities, and ParameterNet, which models the dependency on simulation parameters. Graph Neural Networks are integrated into ShapeNet to handle domain geometry variations. Once trained, the model achieves up to three orders of magnitude speed-up compared to CFD, with robust interpolation and extrapolation capabilities.