Towards a Foundation-Model Paradigm for Aerodynamic Prediction in Three-dimensional Design

Paper • 🤗 Hugging Face • Demo

Yunjia Yang, Babak Gholami, Caglar Gurbuz, Mohammad Rashed, Nils Thuerey

---

This work introduces a foundation-model style methodology for efficiently constructing accurate surrogate models for three-dimensional configurations. It includes a two-stage strategy: first, it pre-trains a large-scale model on diverse geometries, and then fine-tunes it with a few more detailed task-specific samples.

We evaluated the method on transonic wings, where the model is pre-trained on SuperWing (including nearly 30000 samples with broad geometric diversity) and fine-tuned to handle specific wing shapes perturbed from the Common Research Model.

We observed:

• pre-trained model learned the dominant aerodynamics, and performs well with a very small task-specific dataset (even zero-shot).

  

• with 450 task-specific samples, the proposed methodology achieves 0.36% error on surface-flow prediction (1.2% error in drag coefficient ${C_D}$), reducing 84.2% compared to training from scratch.

  

We also studied the influence of model configurations and training strategies to guide the effective training and deployment of such models under limited data and computational budgets.

Resources Overview

• 🤗 Dataset collection:
  • SuperWing (Pre-training dataset)
    https://huggingface.co/datasets/yunplus/SuperWing
  • CRMpert (Task-specific fine-tuning dataset) https://huggingface.co/datasets/thuerey-group/CRMpert
• 🤗 Model hyperparameter collection (Pre-trained and fine-tuned):
  https://huggingface.co/thuerey-group/AeroTransformer
• Implementation dependency:
  • Model implementation (FloGen repo) https://github.com/YangYunjia/floGen
  • Wing postprocess and visualization (cfdpost repo) https://github.com/YangYunjia/cfdpost
• Training and simulation source code (here)
• WebWing Online interactive wing design tool:
  • Online version: https://webwing.pbs.cit.tum.de/
  • Source code: https://github.com/YangYunjia/webwing

Repository Scope

This repo is intentionally focused on training source code and simulation-related scripts for the AeroTransformer project.

AeroTransformer/
├── training/
│   ├── pretrain.py      # pre-train AeroTransformer
│   ├── finetune.py      # fine-tune a pretrained model on downstream task data
│   └── postprocess.py   # evaluate field/coefficient errors
├── simulation/
│   ├── gen-mesh.crmpert.py      # surface + volumetric mesh generation for CRMpert dataset
│   ├── gen-mesh.superwing.py    # surface + volumetric mesh generation for SuperWing dataset
│   ├── original_tip.xyz  # template wing tip shape and mesh
│   ├── run-adflow.py     # calling the pyADflow solver
│   └── single-point.py   # main function for simulating single/multiple samples for one shape
├── LICENSE
└── readme.md

Environment Setup

Simulation part requires MDO Lab docker image (at least with ADflow, pyHyp, cgnsutilities)

Citation

@misc{yang2026foundationmodelparadigmaerodynamicprediction,
      title={Towards a Foundation-Model Paradigm for Aerodynamic Prediction in Three-dimensional Design}, 
      author={Yunjia Yang and Babak Gholami and Caglar Gurbuz and Mohammad Rashed and Nils Thuerey},
      year={2026},
      eprint={2604.18062},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.18062}, 
}

License

This repository is released under the license provided in LICENSE.