Pretraining a neural PDE surrogate on one airfoil family can be worth 3.25 times more target data than training from scratch, but that edge shrinks or flips as the target dataset grows to 5,000 samples. The catch: when the new task adds transition modeling instead of matching the old physics, the payoff shows up on a different, less favorable schedule entirely.

— via arXiv — Physics, Pochinapeddi Sai Bhargav, Nithin Somasekharan, Rohit Sunil Kanchi, Sicheng He, Shaowu Pan