Post #2606324

I recently learned there's this technique in modeling called "neural surrogates". You can basically take an incredibly slow mathematical model, feed its output into a type of neural network, and the neural network learns how to mimic the behavior (and possibly even generalize it). You can then run this neural network efficiently on a GPU and get orders of magnitude speedups. With FNO, a particular technique, you can even switch to a higher resolution than the original training data. That's pretty neat stuff. I suspect there are all kinds of unexpected application domains lurking in there too.