TLDR: A new AI model, Anchored-Branched Universal Physics Transformers (AB-UPT), significantly speeds up complex fluid dynamics simulations for cars and aircraft. Evaluated on new datasets (SHIFT-SUV for cars, SHIFT-Wing for aircraft), AB-UPT outperforms existing models, providing highly accurate predictions of aerodynamic forces in seconds and training within a day on a single GPU. It can even make accurate predictions from simple CAD geometries, making it highly practical for industrial design optimization.
Computational Fluid Dynamics (CFD) simulations are crucial for designing vehicles and aircraft, but they come with a significant drawback: they are incredibly time-consuming and computationally expensive. Imagine needing days or even weeks to simulate how air flows around a new car design or an airplane wing. This challenge has driven researchers to seek faster, more efficient alternatives, leading to the development of neural surrogates – AI models that can approximate these complex simulations in a fraction of the time.
A recent technical report, titled “AB-UPT for Automotive and Aerospace Applications” by Benedikt Alkin, Richard Kurle, Louis Serrano, Dennis Just, and Johannes Brandstetter, introduces an advanced solution called the Anchored-Branched Universal Physics Transformer (AB-UPT). This innovative model has demonstrated remarkable capabilities in replicating automotive and aerospace CFD simulations, requiring significantly less computational power than traditional methods. The paper highlights AB-UPT’s strong performance across various applications, notably achieving near-perfect predictions of integrated aerodynamic forces within seconds, starting from a simple geometric representation. What’s more, it can be trained within a day on a single GPU, making it highly suitable for industrial-scale applications.
Understanding AB-UPT
AB-UPT is a transformer-based model designed to handle large-scale physical simulations. It treats the input geometry, like a car or an airplane, as a collection of points (a point cloud). It then processes this information through separate “branches” for surface and volume fields, allowing for specialized handling of different parts of the simulation. A key to its efficiency is the use of “anchor points” – a drastically reduced set of input points used during training. This approach keeps the computational cost of its powerful attention mechanism in check, while still allowing for efficient and accurate inference on the full set of points.
The original AB-UPT work showcased its effectiveness in high-fidelity automotive CFD simulations, outperforming other models. This new report further solidifies its position by evaluating it on two new, challenging datasets.
New Frontiers: SHIFT-SUV and SHIFT-Wing Datasets
The researchers trained and evaluated AB-UPT on two recently introduced datasets from the Luminary Cloud SHIFT program: SHIFT-SUV and SHIFT-Wing. These datasets combine high-quality data generation with state-of-the-art neural surrogates, pushing the boundaries of what’s possible in CFD modeling.
The SHIFT-SUV dataset focuses on automotive CFD simulations, featuring various car designs. It includes over four thousand transient simulations of geometric variants of the AeroSUV vehicle platform. These variants were created using “morphing boxes” to deform the car’s shape, exploring changes in hood height, vehicle height, windshield angle, and more. The dataset includes both full-scale and quarter-scale models, with training primarily conducted on the full-scale simulations. AB-UPT consistently outperformed other transformer-based models like Transolver and Transformer, as well as the point-cloud-based neural operator DoMINO, in predicting surface and volume variables for both estate and fastback car types. It showed high accuracy in predicting aerodynamic drag and lift forces, although lift force prediction for some estate car types proved more challenging due to less represented data in the training set.
The SHIFT-Wing dataset is built around the NASA Common Research Model (CRM) and is tailored for high-speed transonic commercial aircraft aerodynamics. This dataset explores variations in wing and fuselage geometry, angle-of-attack, and Mach numbers (0.5 and 0.85). A unique aspect of SHIFT-Wing is the use of Luminary Mesh Adaptation, a proprietary technology that automatically adjusts the mesh to accurately capture sharp flow features like transonic shocks. For this dataset, separate AB-UPT models were trained for different Mach numbers, as combining them did not improve performance. AB-UPT achieved almost perfect accuracy in modeling aerodynamic drag and lift forces on the SHIFT-Wing test set. Remarkably, it demonstrated strong performance even when trained with significantly fewer simulations, showing high accuracy with as few as 56 samples for lift force prediction. Furthermore, AB-UPT proved capable of making accurate predictions from a simple, isotropically tessellated geometry representation (like a CAD model), rather than requiring the complex, adaptively refined CFD mesh, which is a significant practical advantage.
Also Read:
- Universal Forecasting with Chronos-2: Beyond Single Series Predictions
- Hybrid Optimization Solver Delivers Unprecedented Speed and Trust for Power Grids
Practical Implications and Future Outlook
The results presented in this report underscore AB-UPT’s potential to revolutionize design optimization in automotive and aerospace industries. Its ability to provide highly accurate CFD simulations orders of magnitude faster than traditional methods means engineers can iterate on designs much more quickly and cost-effectively. For instance, calculating aerodynamic forces for new designs can take as little as 0.6 seconds. The flexibility of AB-UPT to work with simpler CAD inputs for inference further enhances its practical usability.
This work solidifies AB-UPT as a leading model for neural surrogates in external CFD simulations, paving the way for its widespread adoption in industry. For more in-depth technical details, you can refer to the full research paper here.
