Pretraining boosts neural CFD models but gains vary by data and physics
How Does Distribution Shift Shape Pretraining Gains in Neural PDE Surrogates?
Machine Learning
Summary
Using neural networks to simulate airflow around wings can be faster if the model is pretrained on similar data first. The authors studied how changing the type of wing shapes and the physics modeled affects the usefulness of this pretraining. They found that pretraining helps more when the new data covers many shapes and uses the same physics as the original data. When the physics model changes, the benefits shift depending on how much new data is available. This work clarifies when pretraining is most effective for fluid dynamics simulations.
What this means in practice
- •For computational fluid dynamics engineers: Choose efficient pretraining strategies for neural surrogates under changing wing shapes and physics models to reduce expensive simulation data needs.
- •For aircraft design teams: Use pretrained neural models more effectively when adapting to new airfoil designs or updated turbulence models to speed up airflow predictions.
Authors
Pochinapeddi Sai Bhargav, Nithin Somasekharan, Rohit Sunil Kanchi, Sicheng He, Shaowu Pan
Abstract
Pretraining a neural PDE surrogate can reduce the amount of new CFD data needed when geometry or modeled physics changes. However, it remains unclear how different components of distribution shift affect this benefit. We pretrain a surrogate on 254,909 RANS solutions from one airfoil family and fine-tune it on a new family under two target settings with matched freestream ranges: the same Spalart-Allmaras (SA) modeling and SA with added $e^N$ transition modeling. At $N=1000$, the pretrained model matches the accuracy of a model trained from scratch on $3.25\times$ as many samples for the same-SA target, but $2.58\times$ as many for the transition-modeled target. By $N=5000$, this ordering reverses ($1.56\times$ versus $1.86\times$). At $N=1000$, sampling more distinct airfoils lowers error on both targets, but only for the same-SA target is the gain increase larger than the observed draw-to-draw variation ($3.3\times$ to $4.0\times$). These results show that pretraining value depends jointly on target-data budget, target-data coverage, and whether source and target differ in modeled physics.