Session: 03-09 Digital applications

Paper Number: 128704

128704 - Machine Learning Regression of Under-Expanded Hydrogen Jets 

The safety of gas turbine (GT) design is paramount, especially in recent years with the introduction of hydrogen-methane blends to reduce CO2 emissions. Computational Fluid Dynamics (CFD) analysis is therefore crucial to ensure the reliability of ventilation systems in fuel leak scenarios within gas turbine enclosures. This analysis ensures an effective dilution of leaked fuel, minimizing the risks of deflagration. However, each leak scenario is influenced by various factors, like the location and direction of the leak, its cross-section, gas pressure in the pipelines, gas composition, and interactions of the jet with the objects inside the enclosure. Therefore, considering all the possible combinations of these factors, conducting a detailed and throughout investigation would require a number of simulations too high for industrial studies. Hence, there is a need for a streamlined tool capable of providing quick responses regarding fuel dispersion for leak analysis. In this context, advancements in machine learning have proven the effectiveness of surrogate models. These models offer faster responses compared to conventional CFD approaches, especially for specific case studies. In this context, in a previous work (Cerbarano, et al. 2023) a study on hydrogen and methane high-pressure leaks has been proposed, focusing on main features and CFD modelling aspects. With this aim, a comprehensive dataset of 75 RANS simulations accounting for different hydrogen/methane concentrations in the fuel, storage to ambient pressure ratios (PRs) at the leak section, and crossflow ventilation intensities has been constructed. Based on this dataset, in this work, a machine learning surrogate model was developed to interpolate solutions in the space of input parameters. The approach proposed is based on Graph Neural Networks (GNNs), a class of machine learning models that deal with graph structured data, capturing the dependence of graphs via message passing between the nodes of graphs. In the way they are conceived GNNs can deal directly with mesh-based simulation data, leveraging both domain topology and local fluid features. The model is trained to solve a node-level regression task predicting fuel concentration in space for different high pressure leak scenarios. Our model shows a significant speed up in predicting fuel dispersion with respect to conventional methodology.

Presenting Author: Davide Cerbarano Sapienza University of Rome

Presenting Author Biography: Davide Cerbarano is a PhD student at Sapienza University of Rome. Together with Baker Hughes company, he is developing a project on modeling high-pressure fuel leakages scenarios within gas turbine enclosure, with a particular focus on CFD and data-driven methodologies, to enhance safety evaluations.

Authors:

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