BlendedNet++: A Large-Scale Blended Wing Body Aerodynamics Dataset and Benchmark

The dataset

From dataset to benchmark, at aircraft scale

BlendedNet++ scales the original BlendedNet effort into a large-scale aerodynamics dataset and benchmark for blended wing body aircraft — broader geometric coverage, more high-fidelity simulations, and standardized evaluation protocols so neural surrogates for aircraft aerodynamics can finally be compared on equal footing.

Conceptual BWB design has always been squeezed from two sides: the aerodynamics are complex enough to demand serious CFD, and the design space is high-dimensional enough that you can never afford as many runs as you'd like. BlendedNet++ attacks both ends. 12,492 unique geometries come with RANS-computed forces and dense surface fields, and on top of them the paper benchmarks five surrogate architectures — with Transolver emerging as the most accurate for field prediction. Then it goes a step further than prediction: a conditional-diffusion inverse-design pipeline with gradient-based refinement generates multiple feasible aircraft that hit specified lift-to-drag targets with R² above 0.99, confirmed by CFD — a shift from iteratively analyzing designs to directly generating them.

BlendedNet++ overview: large-scale BWB geometry generation, CFD simulation, and benchmark evaluation.

A 5-by-5 grid of blended wing body planforms illustrating the geometric diversity of the BlendedNet++ design space. — Geometric diversity: sampled planforms range from slender high-aspect-ratio wings to compact deltas.

Contributions

What the paper delivers

Large-scale BWB corpus

A substantially expanded set of blended-wing-body geometries and high-fidelity aerodynamic simulations.

Benchmark protocol

Defined splits, metrics, and baselines turn aircraft surrogate modeling into a reproducible, comparable research problem — mirroring what CarBench did for cars.

Foundation for aircraft SciML

Together with DrivAerNet++ and CarCrashNet, completes a family of open benchmarks spanning automotive aero, crash, and aircraft aerodynamics.

Inside the data

Full surface fields across the flight envelope

Each case pairs the geometry with surface pressure and skin-friction fields, sampled across a realistic flight envelope — including a held-out Group 3 UAS regime that tests whether surrogates generalize beyond their training conditions.

A BWB geometry with its surface pressure coefficient, streamwise skin friction, and vertical skin friction fields. — One sample, four modalities: geometry plus surface C_p, C_f,x, and C_f,z fields.

Flight envelope plot in Mach number and altitude showing full-envelope training cases, sub-envelope Group 3 UAS training cases, and Group 3 UAS test cases. — Envelope splits: 9,992 full-envelope training cases, a 3,860-case Group 3 UAS sub-envelope, and 2,500 held-out test cases.

Reference

Citation

@article{sung2025blendednetpp,
  title   = {BlendedNet++: A Large-Scale Blended Wing Body Aerodynamics
             Dataset and Benchmark},
  author  = {Sung, Nicholas and Spreizer, Steven and Elrefaie, Mohamed and
             Jones, Matthew C. and Ahmed, Faez},
  journal = {arXiv preprint arXiv:2512.03280},
  year    = {2025}
}

Keep exploring