Anchored-branched Universal Physics Transformers

Implementation of AB-UPT, published in Transactions on Machine Learning Research, containing:

• Model implementation src/model.py
• DrivAerML dataset implementation src/drivaerml_dataset.py
• Preprocessing pipeline src/abupt_collator.py

The tutorial.ipynb notebook showcases various aspects of our work:

• DrivAerML data download, inspection, and visualization
• Preprocessing data for AB-UPT
• Running inference with AB-UPT to calculate MSE, L2 error, drag/lift coefficients, and streamline velocity visualizations

We recommend checking it out yourself in an interactive Google Colab runtime.

Update (2026.01): Emmi AI released the Noether Framework. The full AB-UPT training code is part of the framework tutorial.

Further resources

If our work sparked your interest, check out our other works and resources!

Papers

• UPT
• NeuralDEM
• AB-UPT

Videos/Visualizations

• AB-UPT demo
• UPT talk
• AB-UPT talk

Code

• NeuralDEM
• UPT tutorial
• Full AB-UPT training tutorial

Citation

If you like our work, please consider giving it a star :star: and cite us

@article{
alkin2025abupt,
title={{AB}-{UPT}: Scaling Neural {CFD} Surrogates for High- Fidelity Automotive Aerodynamics Simulations via Anchored- Branched Universal Physics Transformers},
author={Benedikt Alkin and Maurits Bleeker and Richard Kurle and Tobias Kronlachner and Reinhard Sonnleitner and Matthias Dorfer and Johannes Brandstetter},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=nwQ8nitlTZ},
note={}
}