Machine learning (ML) is transforming the field of computational fluid dynamics (CFD) by addressing the challenges posed by traditional high-fidelity simulations, which often require significant computational resources. At the Technical University of Munich (TUM), under the leadership of Professor Dr. Nikolaus A. Adams, our research team is developing innovative hybrid methods that integrate ML models into established simulation workflows. This approach leverages the accuracy of conventional numerical methods while significantly reducing computational costs, opening new possibilities for applications in fields such as aerospace and environmental engineering.
One of the key focuses of our work is the integration of Fourier neural operators (FNOs) with the lattice Boltzmann method (LBM). To achieve this, we have developed the TorchLBM solver, built on PyTorch, which allows for efficient tensor computing and GPU acceleration. This modular framework enables rapid simulations of complex single- and multiphase flows. By dynamically incorporating FNOs into the LBM framework, we have achieved substantial reductions in computational time while maintaining the physical accuracy essential for high-fidelity simulations.
We demonstrated the effectiveness of our hybrid approach through two case studies: the dynamic evolution of a Kármán vortex street and the steady-state flow through porous media. In both cases, FNOs were employed to reduce time-to-solution by up to 50%, highlighting the efficiency and stability of these hybrid models. This integration of ML into traditional CFD workflows marks a significant improvement in our ability to solve complex flow problems efficiently.
Our work with NVIDIA Modulus and the development of these hybrid models represent a step forward in fluid dynamics research. As we continue to refine these models and scale up our simulations using multi-GPU setups, we are opening new avenues for AI-enhanced CFD applications that can significantly improve performance and accelerate innovation across various industries.
Source: https://developer.nvidia.com/blog/transforming-cfd-simulations-with-ml-using-nvidia-modulus/