The SHARP-Raptor Challenge: Surface Pressure Prediction on an F-22 Airfoil Section

The Challenge

Predicting how air pressure distributes itself over a wing is one of the central problems in aerodynamic design. This distribution—the pressure coefficient (C_p) at every point along an airfoil's surface—determines critical aerodynamic performance indicators such as lift, drag, and structural loads.

Typically, C_p curves are obtained by running expensive Reynolds-Averaged Navier-Stokes (RANS) CFD simulations. Since each simulation can take from several minutes to hours, exploring the vast multi-dimensional operating space during early-stage aircraft design is computationally prohibitive.

The SHARP-Raptor Challenge addresses this bottleneck as a high-speed machine learning benchmark for surrogate modeling. Given only two inputs—the freestream Mach number and the angle of attack (α)—your task is to build a model capable of accurately predicting the complete C_p distribution across 13,862 physical points of an F-22 fighter airfoil section.

By substituting expensive physics solvers with instantaneous ML inference, this challenge aims to evaluate the limits of ML models in learning and generalization across highly non-linear subsonic, transonic, and supersonic flow regimes.

Database

All samples in this dataset correspond to the exact same fixed airfoil geometry (an F-22 Raptor airfoil section), evaluated across various flow conditions. The geometry and point distribution remain identical across all cases; only the flow environment changes.

Each flow condition is parameterized by:

Mach number (mach): Freestream Mach number, ranging from 0.3 to 1.4 (covering subsonic, transonic, and supersonic regimes).
Angle of attack (alpha / aoa): Angle of attack in degrees, ranging from 0° to 5°.

For each combination of Mach and α, the pressure coefficient (C_p) is provided at 13,862 fixed points along the airfoil's boundary. The same physical points (mapped via point_id) are preserved across all data files.

Methods

We are currently documenting our baseline architectures, including Multi-Layer Perceptrons (MLPs), Kolmogorov-Arnold Networks (KANs), and diffusion-based models.

Complete methodology documentation, modeling code, and training pipelines will be available soon.

Results

Baseline performance benchmarks and Kaggle leaderboard snapshots will be published here upon the official launch of the challenge. Stay tuned!

BibTeX

@misc{sharpraptor2026,
      title={The SHARP-Raptor Challenge: A High-Speed Machine Learning Benchmark for Surface Pressure Prediction on an F-22 Airfoil Section},
      author={Jaraiz, Miguel and Ramos, David and Guti{\'e}rrez, Ferm{\'\i}n and Rubio, Gonzalo and Lacasa, Lucas and Valero, Eusebio},
      year={2026},
      eprint={coming soon},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={coming soon},
      }

If you use pyLOM, please also cite:

@misc{pyLOM,
      author={Eiximeno, Benet and Begiashvili, Beka and Guti{\'e}rrez, Ferm{\'\i}n and Ramos, David and Jaraiz, Miguel and Yeste, Pablo and Ladr{\'o}n, {\'A}ngel and Becerra, Nicol{\'a}s and Franc{\'e}s-Belda, V{\'\i}ctor and Robledo, Isaac and Nieto-Centenero, Javier and Miro, Arnau and Rubio, Gonzalo and Lacasa, Lucas and Castellanos, Rodrigo and Sanmiguel, Carlos and Andr{\'e}s, Esther and Valero, Eusebio and Lehmkuhl, Oriol},
      title={pyLOM: Low order modelling in Python},
      year={2022},
      publisher={Barcelona Supercomputing Center},
      journal={GitHub repository},
      url={https://github.com/ArnauMiro/UPM_BSC_LowOrder},
      }

Acknowledgements

This work was supported by the following institutions and funding programs.