Practical experience with ML surrogates for CFD and FEA simulations?

An engineer seeking real-world experience with ML surrogates for CFD and FEA simulations asks about practical architectures, data efficiency, physics-informed approaches, and generalization. The question highlights the tradeoff between speed and accuracy, noting that expensive solver runs remain necessary when surrogates fail.

I work in engineering simulation CFD and FEA and I’m increasingly interested in using ML to cut down the cost of expensive solver runs. The traditional loop — mesh, solve, post-process — can take hours per design iteration, and I’d love to hear from people who’ve actually deployed ML surrogates in this space rather than just read the papers. A few things I’m trying to get a realistic picture of: Architectures: For predicting fields temperature, pressure, stress over a geometry, what’s worked better in practice — graph neural networks on mesh data, Fourier Neural Operators, point-cloud approaches, or plain MLP/CNN surrogates on parameterized designs? Data efficiency: Simulation data is expensive to generate. How few training samples have people gotten away with for a useful surrogate, and does transfer learning across similar geometries help? Physics-informed approaches: Has anyone found PINNs practical for real engineering geometries, or are they still mostly a research curiosity vs. data-driven surrogates? Generalization: The hard part is a model that holds up on geometries/boundary conditions outside the training distribution. What’s worked for keeping surrogates trustworthy there? Interested in real-world experience — what gave you a usable accuracy-vs-speed tradeoff, and where did ML surrogates fall down and force you back to the full solver?