Practical experience with ML surrogates for CFD and FEA simulations?

An engineering simulation professional seeks real-world deployment experiences of machine learning surrogates to reduce the cost of expensive Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEA) solver runs.

• Architectures: Comparing graph neural networks on mesh data, Fourier Neural Operators, point-cloud approaches, and MLP/CNN surrogates for predicting fields like temperature and stress.
• Data efficiency: Investigating the minimum number of training samples required for useful surrogates and the utility of transfer learning across similar geometries.
• Physics-informed approaches: Assessing whether Physics-Informed Neural Networks (PINNs) are practical for real engineering geometries compared to data-driven methods.
• Generalization: Addressing how to maintain model trustworthiness on geometries and boundary conditions outside the training distribution.

The author aims to identify which approaches provide a usable accuracy-versus-speed tradeoff and where ML surrogates fail, necessitating a return to full solvers.