Session: 03-02 Advanced Modeling and Retrofit Techniques for Hydrogen and Mixed-Fuel Combustion Systems

Paper Number: 154095

Machine Learning Modelling of Impinging Hydrogen Gas Leaks 

The hydrogen safety in gas turbines is a critical concern, particularly in managing the risks of accidental fuel gas leaks within enclosures, which can lead to deflagrations or detonations. Computational Fluid Dynamics is generally used to analyze the development of the flammable cloud in a fuel gas leak scenario. However, computational costs of this method limit the exploration of the full range of potential leak conditions, including variations in leak area, location, gas composition, and interaction with nearby objects.

To address this limitation, in this paper we propose a machine learning surrogate model to predict gas dispersion from leaks interacting with flat surfaces. The training dataset is based on a parametric set of numerical simulations with different gas compositions, leak area and distance from the surface. The model architecture is based on the Neural Implicit Flows that consists of two separate neural networks, namely: ShapeNet, which isolates and represents the spatial complexity of the local flow filed, and ParameterNet, which accounts for the dependence of the solution from simulation parameters. Once trained the model can reproduce the entire flow field given the parameters that characterize the scenario, reducing the computational cost of 2 order of magnitude if compared with CFD.

Presenting Author: Davide Cerbarano Sapienza University of Rome

Presenting Author Biography: Davide cerbarano is a PhD candidate at Sapienza University of Rome, and he is working with Baker Hughes company to develop machine learning methodologies to enhance CFD gas leak analysis inside gas turbine enclosure.

Authors:

Davide Cerbarano Sapienza University of Rome
Lorenzo Tieghi University of Trento
Giovanni Delibra Sapienza University of Rome
Stefano Minotti Baker Hughes
Alessandro Corsini Sapienza University of Rome