Surrogate model for temperature fields from a small set of CFD runs: which architecture to start with?

A practitioner building a surrogate model for steady-state temperature fields on heat sinks and electronics boards, trained on roughly 150–300 CFD simulation runs, was advised to stay with PCA plus a small model rather than adopt a Fourier neural operator or graph-based model at that data scale. The recommended fix for the model's missed hot spots near sharp features is to score on worst-case and 95th-percentile error and hotspot location, add a second small head predicting peak temperature and its position, or weight the loss toward high-gradient cells, and only then test a CNN on rasterized geometry against the same held-out geometries.

Hi everyone, I’m trying to build a surrogate model that predicts steady-state temperature fields on a heat sink / electronics board from its geometry parameters, so I can skip running a full simulation for every design change. My setup so far: • About 150–300 simulation runs, each with N geometry parameters fin height, spacing, base thickness, etc. • Output is a temperature field on a structured grid / unstructured mesh with ~X nodes • I’ve tried a simple MLP on the parameters, predicting a downsampled field. Mean error is acceptable, but it misses hot spots near sharp features. With a dataset this small, is it more sensible to go with a Fourier neural operator, a graph-based model, or a CNN on a rasterized geometry? Or should I stay with a simpler model plus a PCA reduction of the field? If there are open datasets or Hub models for physics surrogates that I could fine-tune or learn from, I’d appreciate pointers. I can share my training loop and a sample of the data if that helps. Thanks With 150–300 runs, I would stay with PCA plus a small model for now — an FNO or graph model will likely spend that budget learning the field’s basics, while your current miss has a cheaper cause: a downsampled field trained on mean error can be right on average and still smooth away the peaks. Keep the simple model, but score it on worst-case and 95th-percentile error and whether the hotspot lands in the right place , and add a second small head that predicts peak temperature and its location, or weight the loss toward high-gradient cells near sharp features. If that still misses, test a CNN on the rasterized geometry on the same held-out geometries — that comparison, not the architecture names, will tell you whether geometry representation is really the bottleneck.