How Does Distribution Shift Shape Pretraining Gains in Neural PDE Surrogates?
When AI models trained on old airplane data help predict new ones
Training an AI model on airfoil aerodynamics from one aircraft family before fine-tuning it on a different family cuts the amount of new training data needed by 2–3 times—but only when the physics being modeled stays the same. When researchers added a more complex turbulence model to the target task, the pretraining advantage shrank significantly, suggesting that mismatches between source and target problems can erase what looks like a helpful shortcut.
Computational fluid dynamics simulations are expensive and time-consuming. If engineers can reuse models trained on similar past problems, they could dramatically reduce the cost of designing new aircraft or engines. This research shows pretraining works well for incremental design changes, but warns that it breaks down when the underlying physics or modeled phenomena differ—a critical caveat for deploying these models in real engineering pipelines.