Session: 36-09 Machine Learning & Artificial Intelligence Methods - Part 4

Submission Number: 178657

GT Enclosure Dispersion Analysis With Hybrid CFD-AI Tools 

The fuel gas leakage inside the enclosure of a gas turbine is an undesired event which could occur even if all preventive actions are taken by design. In such cases, the dilution of the leak is crucial, and the ventilation system is designed to this aim. CFD tools are usually employed for analyzing the dispersion from accidental fuel gas leakages in relevant scenarios, to improve and eventually update the package design. The complexity of the dispersion simulations, in terms of preprocessing and computing times, makes it possible to run a few cases, carefully identified after a detailed analysis of the air velocity field inside the enclosure. The use of AI based surrogate models can provide a significant improvement to this methodology, allowing to run a significant number of different scenarios in a fraction of the time, providing a speed-up on the current procedure, as well as enabling a more robust evaluation of all the potential leak configurations. However, building a data-driven model for the 3D dispersion analysis is a challenging task. Historical data are scarce, and creating new datasets is not easy because each simulation may take days to converge, due to domain size and time step requirements. In the current work, an available dataset of less than 100 scenarios from historical analysis has been employed, considering variations in package and turbine geometries, leak position, direction and mass flow rate. To tackle the limitations of the reduced size of the training dataset, and improve the robustness of the models, several techniques have been developed and adapted based on physical understanding of the problem and latest research findings. Significant focus has been directed towards generating relevant and informative features to be used as input to the model, such as the pointwise undisturbed flow velocity (obtained from the preliminary CFD simulation of the hot flow), the distance function from the solid walls and the leak source and other preprocessing steps. After the highly customized feature engineering process, different models have been compared, from classical computer vision derived algorithms, such as a U-Net based on 3D convolutions, to Transformer based solutions, Graph Neural Networks and Neural Operators. The comparison has been performed to verify the robustness of the selected pipeline against changes in geometry, flow conditions, and leak properties, in terms of equivalent flammable volume (as per requirements) and in terms of morphology of the dispersed gas phase. The best performing approaches demonstrate a satisfactory agreement with the detailed CFD results, in terms of both volume and shape, and demonstrate to be robust to changes in leak direction and diameter, handling the limited data availability thanks to the strong physics-based feature engineering steps that make the models more robust.

Presenting Author: Leonardo Pulga Baker Hughes

Presenting Author Biography: Leonardo Pulga is Senior AI Specialist for Engineering and Controls at the IET AI Team of Baker Hughes. His main activities involve the integration of AI and simulation tools for improved product and process design and optimization.

Authors:

Leonardo Pulga Baker Hughes
Laure Barriere Baker Hughes
Andrea Panizza Baker Hughes
Elena De Leo Baker Hughes
Stefano Minotti Baker Hughes