[Paper] Going with the Speed of Sound: Pushing Neural Surrogates into Highly-turbulent Transonic Regimes

Source: arXiv - 2511.21474v1

Overview

A new study pushes the frontier of neural‑network‑based aerodynamic surrogates from ground‑vehicle bluff‑body flows into the highly non‑linear, compressible world of transonic aircraft wings. By releasing a large‑scale 3‑D CFD dataset (≈ 30 k samples) that captures wing‑tip vortices and shock‑wave interactions, the authors show that modern surrogate models—especially the AB‑UPT architecture—can reliably predict lift, drag, and full flow fields even for geometries and flight conditions they have never seen before.

Key Contributions

• First public 3‑D transonic wing dataset (≈ 30 k CFD samples) with volumetric and surface fields, geometry parameters, and inflow conditions.
• Benchmarking of state‑of‑the‑art neural surrogates (Transolver, AB‑UPT) on out‑of‑distribution (OOD) generalization across both geometry and Mach number variations.
• Demonstration that AB‑UPT accurately reproduces drag‑lift Pareto fronts for unseen wing designs, enabling rapid design space exploration.
• Open‑source release of the dataset on Hugging Face, encouraging reproducibility and further research in data‑driven aerospace aerodynamics.

Methodology

1. Dataset Generation – High‑fidelity CFD simulations were run for a diverse set of 3‑D wing geometries (varying sweep, aspect ratio, twist, etc.) and inflow conditions spanning subsonic to transonic Mach numbers (≈ 0.6–0.9). Each simulation provides:
   • Full 3‑D volumetric fields (pressure, density, velocity).
   • Surface quantities (wall shear stress, pressure coefficient).
   • Integrated performance metrics (lift C_L, drag C_D).
2. Neural Surrogate Architectures – Two recent models were evaluated:
   • Transolver – a graph‑based solver that iteratively refines predictions using physics‑inspired message passing.
   • AB‑UPT – an attention‑based U‑shaped transformer that processes both geometry embeddings and flow field tensors.
3. Training & OOD Testing – Models were trained on ~ 80 % of the dataset and tested on the remaining 20 % that contain unseen wing shapes and new Mach numbers. Performance was measured by:
   • Mean absolute error on C_L and C_D.
   • Field‑wise L2 error.
   • Ability to reconstruct the drag‑lift Pareto curve for novel designs.
4. Physical Consistency Checks – Beyond raw error metrics, the authors verified that predicted flow fields respect key transonic phenomena (e.g., shock location, vortex strength) and that the surrogate‑derived Pareto fronts obey known aerodynamic trade‑offs.

Results & Findings

• AB‑UPT outperforms Transolver on both coefficient prediction (≈ 5 % lower MAE) and full‑field reconstruction (≈ 7 % lower L2 error).
• The surrogate can predict shock wave position and wing‑tip vortex structures with visual fidelity comparable to CFD, even for OOD cases.
• When sweeping design parameters, AB‑UPT reproduces smooth, physically plausible drag‑lift Pareto curves for unseen wings, enabling rapid “what‑if” analysis without rerunning expensive CFD.
• Inference time drops from hours (full CFD) to sub‑second on a single GPU, representing a speed‑up of 4–5 orders of magnitude.

Practical Implications

• Accelerated Conceptual Design – Aircraft engineers can now explore thousands of wing variants in minutes, iterating on lift‑drag trade‑offs early in the design cycle.
• Real‑Time Flight‑Envelope Tools – Embedding the surrogate in flight‑simulation software could provide pilots or autonomous systems with on‑the‑fly aerodynamic estimates for varying Mach numbers and attitudes.
• Optimization Pipelines – Gradient‑based or evolutionary optimizers can query the surrogate directly, dramatically reducing the computational budget of multi‑objective aerodynamic shape optimization.
• Cross‑Domain Transfer – The dataset and model architecture lay groundwork for extending neural surrogates to other compressible flow problems (e.g., nozzle design, supersonic intake) where 3‑D effects dominate.

Limitations & Future Work

• Dataset Scope – While diverse, the current set focuses on a single wing family and a limited Mach range; extreme transonic/supersonic regimes and full aircraft configurations remain untested.
• Physics Fidelity – The surrogate captures major flow features but may miss subtle turbulence‑induced losses; coupling with turbulence‑model corrections could improve accuracy.
• Generalization to New Airfoils – OOD tests involve geometry variations within the same design space; truly novel airfoil families may still challenge the model.
• Future Directions – The authors suggest expanding the dataset to include control surfaces, variable Reynolds numbers, and integrating uncertainty quantification to provide confidence bounds for surrogate predictions.

The dataset and code are freely available at https://huggingface.co/datasets/EmmiAI/Emmi-Wing, inviting the community to build the next generation of AI‑driven aerodynamic tools.

Authors

• Fabian Paischer
• Leo Cotteleer
• Yann Dreze
• Richard Kurle
• Dylan Rubini
• Maurits Bleeker
• Tobias Kronlachner
• Johannes Brandstetter

Paper Information

• arXiv ID: 2511.21474v1
• Categories: cs.CE, cs.AI, cs.LG
• Published: November 26, 2025
• PDF: Download PDF