Home // AISyS 2024, The First International Conference on AI-based Systems and Services // View article

An Advanced Surrogate Model Approach for Enhancing Fluid Dynamics Simulations

Authors:
Shubham Kavane
Kajol Kulkarni
Harald Koestler

Keywords: Surrogate Models; Computational Fluid Dynamics (CFD); U-Net; Fourier Neural Operators; Model Parallelism;

Abstract:
The increasing complexity and computational de- mands of 3D fluid dynamics simulations highlight the need for ad- vanced surrogate models that strike a balance between predictive accuracy, computational efficiency, and convergence time. Tra- ditional Computational Fluid Dynamics (CFD) methods, while highly accurate, are often resource-intensive and time-consuming. This research presents advanced U-Net-based surrogate models for 3D fluid flow prediction, aiming to achieve faster convergence and more efficient resource utilization while retaining competitive accuracy relative to traditional CFD solvers. We developed a U-Net model featuring an improved architecture utilizing an advanced attention mechanism known as the Convolution Block Attention mechanism. Considering the high computa- tional demands, the model was trained using multiple GPUs, incorporating both model and data parallelism techniques. The model’s capability was evaluated through overfitting experiments, where it was trained on a limited dataset to assess its ability to accurately replicate true labels. These findings highlight the promise of advanced surrogate models as a viable alternative to traditional CFD methods, providing faster solutions and reduced computational costs with comparable accuracy. Future research will focus on evaluating the current advanced U-net model, trained on an extensive dataset of 10,000 samples, against Fourier Neural Operators and traditional CFD solvers in terms of training time, accuracy, and resource utilization, including energy consumption.

Pages: 1 to 6

Copyright: Copyright (c) IARIA, 2024

Publication date: September 29, 2024

Published in: conference

ISBN: 978-1-68558-192-3

Location: Venice, Italy

Dates: from September 29, 2024 to October 3, 2024