Researchers have taken a closer look at how pretraining artificial intelligence (AI) models can speed up simulations of fluid flow around objects like airplane wings. These simulations, based on complex mathematical equations called partial differential equations (PDEs), are essential in engineering but require huge amounts of computing power and data. The new study explores when pretraining AI models on one set of conditions helps reduce the data needed to adapt them to new, related situations—a question that’s crucial for making AI-driven simulations more practical and efficient.
Key Takeaways
- Pretraining AI models on a large dataset of airflow simulations around one family of airfoils can significantly reduce the amount of new data needed to accurately model a different airfoil family.
- The benefit of pretraining depends on the similarity between the original training conditions and the new target conditions, especially whether the physics being modeled changes.
- For simpler physics models, pretrained AI required 2.6 to 3.3 times less new data to reach the same accuracy compared to training from scratch.
- When the physics model was more complex (adding transition modeling), pretraining still helped but required relatively more new data, and the advantage shifted depending on how much target data was available.
In this study, the researchers focused on neural PDE surrogates—AI models designed to approximate solutions to fluid dynamics equations much faster than traditional computational fluid dynamics (CFD) simulations. Specifically, they looked at Reynolds-Averaged Navier-Stokes (RANS) equations, which are widely used to predict airflow around airfoils (wing-shaped objects). The main challenge is that when the geometry (shape of airfoil) or the physics model changes, these AI surrogates usually need to be retrained with new data, which can be costly and time-consuming.
To tackle this, the team pretrained their neural network surrogate on a massive dataset of 254,909 RANS solutions from one family of airfoils. Pretraining means the model first learns general patterns from this large dataset before being fine-tuned on a smaller dataset from a new airfoil family. They tested two scenarios: one where the physics model remained the same (using Spalart-Allmaras or SA turbulence modeling) and one where the physics model was more complex, adding an “e^N transition” model that captures how airflow transitions from smooth to turbulent.
“Distribution shift” is a key term here—it refers to differences between the data used for pretraining and the data used for fine-tuning. This could mean changes in airfoil shapes, flow conditions, or physics models. The researchers wanted to understand how these shifts affect the usefulness of pretraining. For example, if the physics changes significantly, will the pretrained model still help, or is it better to start from scratch?
The results showed that at smaller target dataset sizes (around 1,000 samples), pretraining provided substantial benefits, requiring roughly 2.6 to 3.3 times less data to achieve the same accuracy. However, as the size of the target dataset grew (up to 5,000 samples), the relative advantage of pretraining changed depending on whether the physics model was the same or different. Interestingly, increasing the diversity of airfoils in the fine-tuning data helped reduce errors, but this effect was more pronounced when the physics model did not change.
These findings highlight that the value of pretraining AI surrogates for fluid simulations depends on multiple factors: how much new data is available, how diverse that data is, and whether the underlying physics model has changed. This nuanced understanding can help engineers decide when it’s worth investing in large-scale pretraining and when it might be better to gather new data and train fresh models.
Looking ahead, this research offers practical guidance for applying AI to accelerate fluid dynamics simulations in aerospace, automotive design, and other fields where running full CFD simulations is expensive. By clarifying when pretraining helps most, it paves the way for more efficient AI tools that adapt quickly to new scenarios without requiring massive new datasets. Future work could explore other types of distribution shifts and physics models, as well as real-world testing in industrial design workflows.
Based on research published on arXiv by Pochinapeddi Sai Bhargav, Nithin Somasekharan, Rohit Sunil Kanchi et al..