Engineering simulation practitioners discuss practical ML surrogate models for CFD and FEA

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

The discussion focuses on four key areas: comparing architectures such as graph neural networks, Fourier Neural Operators, and MLPs for field prediction; evaluating data efficiency and transfer learning across similar geometries; assessing the practicality of Physics-Informed Neural Networks (PINNs); and addressing generalization challenges outside the training distribution.

The goal is to identify which approaches provide a usable accuracy-versus-speed tradeoff and where ML surrogates have failed, necessitating a return to full solvers.