If you have to feed back empirical data to tweak and re-train the transformer/CNN-based model, then it really doesn’t offer an advantage over traditional physics-based models. The goal should be to get rid of as many fudge factors as possible.

I came here to say the same thing. There’s no way a coarse CFD simulation of air over a car takes two weeks to run on even budget HPCs.

The article is AI slop whose only citation is a deleted Reddit post.