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

Despite progress in machine learning-based aerodynamic surrogates, the scarcity of large, field-resolved datasets limits progress on accurate pointwise prediction and reproducible inverse design for aircraft. We introduce BlendedNet++, a large-scale aerodynamic dataset and benchmark focused on blended wing body (BWB) aircraft. The dataset contains over 12,000 unique geometries, each simulated at a single flight condition, yielding 12,490 aerodynamic results for steady RANS CFD. For every case, we provide (i) integrated force/moment coefficients CL, CD, CM and (ii) dense surface fields of pressure and skin friction coefficients Cp and (Cfx, Cfy, Cfz). Using this dataset, we standardize a forward-surrogate benchmark to predict pointwise fields across six model families: GraphSAGE, GraphUNet, PointNet, a coordinate Transformer (Transolver-style), a FiLMNet (coordinate MLP with feature-wise modulation), and a Graph Neural Operator Transformer (GNOT). Finally, we present an inverse design task of achieving a specified lift-to-drag ratio under fixed flight conditions, implemented via a conditional diffusion model. To assess performance, we benchmark this approach against gradient-based optimization on the same surrogate and a diffusion-optimization hybrid that first samples with the conditional diffusion model and then further optimizes the designs. BlendedNet++ provides a unified forward and inverse protocol with multi-model baselines, enabling fair, reproducible comparison across architectures and optimization paradigms. We expect BlendedNet++ to catalyze reproducible research in field-level aerodynamics and inverse design; resources (dataset, splits, baselines, and scripts) will be released upon acceptance.

💡 Research Summary

BlendedNet++ addresses two persistent bottlenecks in data‑driven aerodynamics: the lack of a large, diverse, field‑resolved dataset and the absence of a standardized benchmark that spans both forward prediction and inverse design. The authors generate over 12,000 distinct blended‑wing‑body (BWB) geometries by sampling a high‑dimensional design space that includes wing span, sweep, twist, fuselage‑to‑wing ratio, and other shape parameters. Each geometry is simulated once under a fixed flight condition (approximately 10 km altitude, Mach 0.8, standard atmospheric conditions) using steady‑RANS CFD with a fine mesh (~200 k cells) and the k‑ω SST turbulence model. The resulting 12,490 CFD solutions provide (i) integrated aerodynamic coefficients – lift (CL), drag (CD), and pitching moment (CM) – and (ii) dense surface fields: pressure coefficient (Cp) and skin‑friction coefficients in three Cartesian directions (Cfx, Cfy, Cfz). All fields are stored on a common surface mesh, enabling direct use as point clouds or graph structures.

The paper defines a forward‑surrogate benchmark that asks models to predict the full set of surface fields given a geometry. Six state‑of‑the‑art architectures are evaluated under identical data splits (80 % train, 10 % validation, 10 % test): GraphSAGE (message‑passing GNN), GraphUNet (hierarchical GNN with pooling/unpooling), PointNet (permutation‑invariant point‑cloud network), a coordinate‑Transformer in the style of Transolver (self‑attention over spatial coordinates), FiLMNet (coordinate MLP with feature‑wise modulation), and Graph Neural Operator Transformer (GNOT, which merges neural operator theory with transformer attention). Performance metrics include L2 error on the fields, physics‑aware errors (e.g., integrated CL/CD deviation), and computational efficiency (training time, parameter count). The results reveal that hierarchical graph models (GraphUNet) and the transformer‑based approaches achieve the best trade‑off between accuracy and scalability, while pure point‑cloud methods lag in capturing fine‑scale pressure gradients.

For inverse design, the authors formulate a conditional diffusion model that learns a joint distribution over design variables and the target lift‑to‑drag ratio (L/D). At inference time the model samples candidate geometries conditioned on a desired L/D value. Two post‑processing strategies are examined: (a) direct evaluation of the sampled designs using the forward surrogate, and (b) a hybrid pipeline where sampled designs are further refined by gradient‑based optimization on the surrogate (using automatic differentiation of the surrogate’s output with respect to design parameters). The diffusion‑first approach dramatically improves exploration of the design space, providing initial designs that are already close to the target L/D, while the subsequent gradient step fine‑tunes the geometry to meet the exact specification. Compared against a pure gradient‑based optimizer initialized from a random design, the hybrid method converges in fewer iterations and attains higher final L/D values, demonstrating the complementary strengths of stochastic generative sampling and deterministic gradient descent.

The dataset is released with a clear split, accompanying metadata, and scripts for preprocessing, training, and evaluation. All baseline implementations are built on PyTorch Geometric, PyTorch 3D, and HuggingFace Transformers, and are provided under a CC‑BY‑4.0 license to ensure reproducibility and easy adoption by both academia and industry. By delivering a large‑scale, high‑fidelity BWB dataset together with a unified forward and inverse benchmark, BlendedNet++ creates a common ground for fair comparison of emerging aerodynamic surrogates, encourages the development of more accurate pointwise predictors, and opens a pathway toward data‑driven, real‑time aircraft design optimization. Future work may extend the dataset to multiple flight conditions, incorporate structural and control‑surface variables, and explore lightweight surrogate architectures suitable for on‑board design assistance.