BlendedNet: A Blended Wing Body Aircraft Dataset and Surrogate Model for Aerodynamic Predictions

The dataset

Bringing the DrivAerNet recipe to aircraft

Blended wing body (BWB) configurations promise major fuel-efficiency gains, but designing them requires expensive CFD campaigns over unconventional geometry. BlendedNet applies the data-driven playbook pioneered in automotive aerodynamics to aircraft: a public dataset of parametrically generated BWB geometries with high-fidelity CFD, paired with a neural surrogate for rapid aerodynamic prediction.

The numbers behind it: 999 BWB geometries, each simulated across roughly nine flight conditions, yielding 8,830 converged RANS cases with the Spalart–Allmaras turbulence model at 9–14 million cells apiece — with detailed pointwise surface quantities on every one. The surrogate is a two-stage pipeline: a permutation-invariant PointNet first recovers the geometric design parameters from a sampled surface point cloud, then a FiLM-conditioned network uses those parameters plus the flight condition to predict pointwise C_p, C_f,x, and C_f,z — turning a meshing-and-solving campaign into a forward pass.

BlendedNet: parametric blended wing body geometries with simulated aerodynamic fields.

The BWB planform parameterization: chord lengths C1 to C4, span segments B1 to B3, and sweep angles S1 to S3 defined on the planform view. — The parametric BWB planform: section chords (C1–C4), span segments (B1–B3), and sweep angles (S1–S3) fully define each geometry.

Contributions

What the paper delivers

Public BWB dataset

Parametric blended-wing-body geometries with high-fidelity CFD — openly released on Harvard Dataverse for the aircraft-design community.

Aerodynamic surrogate

A learned model predicts aerodynamic performance directly from geometry, replacing hours of simulation in early-stage design.

Design-space coverage

Systematic parametric sampling spans the practically relevant BWB configuration space, enabling generalizable learning.

Fields & surrogate

Surface fields the surrogate learns to reproduce

Each simulation carries full surface distributions — pressure coefficient and all three skin-friction components — across the design space. The neural surrogate learns these fields directly from geometry, matching CFD closely on unseen designs.

Grid of BWB designs showing surface pressure coefficient and the three skin-friction components across five different geometries. — Surface C_p and skin-friction components (C_f,x, C_f,y, C_f,z) across designs from the dataset.

Side-by-side comparison of CFD ground truth, surrogate prediction, and absolute error for pressure and skin-friction fields on a BWB geometry. — Ground truth vs. surrogate prediction vs. absolute error: the learned model reproduces C_p and skin-friction fields with near-uniformly low error.

Reference

Citation

@proceedings{sung2025blendednet,
  author = {Sung, Nicholas and Spreizer, Steven and Elrefaie, Mohamed and
            Samuel, Kaira and Jones, Matthew C. and Ahmed, Faez},
  title  = {BlendedNet: A Blended Wing Body Aircraft Dataset and Surrogate
            Model for Aerodynamic Predictions},
  series = {IDETC-CIE},
  pages  = {V03BT03A049},
  year   = {2025}
}

Keep exploring