A strategy borrowed from generative AI — train cheaply on the familiar, then fine-tune on the hard problem — can cut the number of expensive physics simulations needed by nearly a factor of ten. But a ...