Generalize the ordering of shift-component effects across architectures

Determine whether the ordering of distribution-shift components observed for the convolutional U-Net—specifically, the relative pretraining gains under geometry shift versus geometry combined with transition-modeling shift—holds across other neural PDE-surrogate architecture families.

Background

The paper evaluates transfer from a pretrained convolutional U-Net trained on fully turbulent airfoils with the Spalart–Allmaras closure to two natural-laminar-flow targets: one with the same closure and one with added e^N transition modeling. The experiments show that the relative pretraining gain depends on the shifted component and on the target-data budget, with the geometry-only and geometry-plus-transition-modeling targets exhibiting different gain profiles.

The architecture comparison is conducted only on the pretraining task, and the convolutional U-Net is selected for all transfer experiments. Consequently, the reported ordering of the effects of geometry and modeled-physics shifts has not been tested with alternative architecture families such as transformers, Fourier neural operators, or other neural PDE-surrogate designs.