BlendedNet: CFD Dataset for BWB

• BlendedNet is a large-scale, high-fidelity dataset and modeling suite for data-driven aerodynamic analysis of unconventional blended-wing-body aircraft.
• It covers a nine-dimensional design space with thousands of simulated cases, providing dense surface fields and integrated aerodynamic coefficients for forward and inverse modeling.
• The suite benchmarks diverse surrogate models and advanced inverse design protocols, demonstrating robust performance metrics and few-shot learning capabilities.

BlendedNet is a large-scale, high-fidelity dataset and modeling suite for data-driven aerodynamic analysis of unconventional blended-wing-body (BWB) aircraft. Developed to address the lack of accessible, field-resolved computational fluid dynamics (CFD) data for rare or novel planform geometries, BlendedNet and its successors enable reproducible benchmarking of surrogate models for both forward prediction and inverse design tasks, supporting rapid exploration of a nine-dimensional parametric design space. The core dataset consists of thousands of simulated BWB geometries with dense surface fields, integrated force and moment coefficients, and code recipes for evaluating pointwise prediction accuracy and inverse design performance under standardized protocols (Sung et al., 8 Sep 2025, Sung et al., 2 Dec 2025, Jiang et al., 25 Apr 2026).

1. Dataset Composition and Parameterization

BlendedNet provides a suite of BWB geometries parameterized in a nine-dimensional design space, with planforms generated via Latin Hypercube Sampling (LHS) to maximize coverage and variation. In the original dataset (Sung et al., 8 Sep 2025), 999 planforms are defined using normalized chord and span ratios, along with sweep angles at critical wing sections:

• $C_2/C_1$, $C_3/C_1$, $C_4/C_1$, $B_1/C_1$, $B_2/C_1$, $B_3/C_1$ specify relative chord and span at inner, mid, and tip stations.
• $S_1$, $S_2$, $S_3$ encode sweep angles (degrees) at corresponding quarter-chord locations.

Each geometry is discretized as an exportable CAD solid, meshed into $3.5\cdot 10^4$ to $C_3/C_1$0 surface points, and simulated using NASA FUN3D with Spalart–Allmaras RANS modeling on 9–14 million cell meshes (ensuring $C_3/C_1$1 for boundary-layer fidelity).

Flight conditions span:

• Mach $C_3/C_1$2–$C_3/C_1$3
• Angle of attack $C_3/C_1$4
• Reynolds length $C_3/C_1$5–$C_3/C_1$6 m
• Altitude $C_3/C_1$7–$C_3/C_1$8 kft

Each geometry is simulated over a grid of $C_3/C_1$9 flight conditions, yielding $C_4/C_1$0 converged cases in BlendedNet and over $C_4/C_1$1 in BlendedNet++ (Sung et al., 2 Dec 2025). Outputs at every surface mesh point include: pressure coefficient $C_4/C_1$2, and skin friction coefficients $C_4/C_1$3, $C_4/C_1$4, $C_4/C_1$5. Integrated lift, drag, and pitching moment coefficients are computed via surface integration with $C_4/C_1$6 normalization.

Subsequent releases (BlendedNet++) introduced expanded parameterization (including outboard break location) and larger datasets (12,490 cases), maintaining stringent mesh convergence and coverage standards (Sung et al., 2 Dec 2025).

2. Data Representation and Preprocessing

Surface meshes are represented as unstructured point clouds. For modeling, these are often subsampled to 2,048 or $C_4/C_1$78,000 points per case, with each point storing its 3D coordinates $C_4/C_1$8 and unit surface normal $C_4/C_1$9. Further geometric embedding extends this to a 9-dimensional feature vector: $B_1/C_1$0 where SDF is the signed-distance to the CAD surface and $B_1/C_1$1 is principal curvature (Jiang et al., 25 Apr 2026).

For machine learning pipelines, training points are normalized, and field outputs $B_1/C_1$2 are z-scored or otherwise standardized channel-wise. During surrogate training, data augmentation is achieved by repeated random subsampling.

All data are released in VTK surface format, enabling direct use in field-based surrogate modeling, visualization, and mesh-aware operator learning (Sung et al., 8 Sep 2025, Sung et al., 2 Dec 2025).

3. Forward Surrogate Modeling Benchmarks

BlendedNet and BlendedNet++ provide benchmarking protocols for forward prediction of surface fields from geometry and flight conditions, comparing multiple machine learning architectures trained on uniform splits (Sung et al., 2 Dec 2025):

• PointNet: Applies a shared MLP to each point, global max pooling for permutation invariance, and geometric regression.
• FiLMNet: A feature-wise linear modulation network, where an MLP receives $B_1/C_1$3 as input, with layerwise scaling and shifting conditioned by a hypernetwork on planform and flight parameters.
• Transolver (Editor's term): Transformer blocks with physics-aware tokenization and per-point features.
• GraphSAGE, Graph U-Net, GNOT: Graph-based models using dynamic $B_1/C_1$4-NN graphs from the mesh, with message passing, pooling, and (for GNOT) operator-style global attention.

Quantitative evaluation on the test set (for $B_1/C_1$5) is summarized below:

Model	Params	MSE ($B_1/C_1$6)	MAE ($B_1/C_1$7)	RelL1 (%)	RelL2 (%)
FiLMNet	0.8 M	0.418	2.147	7.11	1.42
Transolver	2.8 M	0.901	6.075	21.91	3.71
PointNet	0.9 M	0.974	6.541	23.59	4.01
GraphSAGE	0.4 M	1.06	5.65	18.59	18.86
GraphUNet	14 M	2.35	8.96	29.45	28.08
GNOT	11.5 M	9.66	17.04	61.45	39.79

FiLMNet achieves the lowest errors, with MAE $B_1/C_1$8 for $B_1/C_1$9; similar trends are observed for skin friction coefficients (Sung et al., 2 Dec 2025). The best PointNet+FiLM surrogates on BlendedNet report $B_2/C_1$0 MAE $B_2/C_1$1 (ground truth geometry parameters) or $B_2/C_1$2 (inferred parameters) (Sung et al., 8 Sep 2025).

4. Inverse Aerodynamic Design Protocols

BlendedNet++ extends the benchmark to inverse design tasks, seeking planform parameters $B_2/C_1$3 that achieve a target lift-to-drag ratio $B_2/C_1$4 under fixed flight conditions. Three approaches are standardized (Sung et al., 2 Dec 2025):

• Conditional Diffusion Model (CDM): Samples $B_2/C_1$5 conditioned on $B_2/C_1$6, trained via denoising score matching.
• Projected Gradient Descent (Opt): Optimizes $B_2/C_1$7 by minimizing the squared deviation from $B_2/C_1$8 using learned surrogates.
• Hybrid (CDM$B_2/C_1$9Opt): Initializes PGD from CDM samples for diversity.

Key metrics include $B_3/C_1$0 to target $B_3/C_1$1, RMSE, MAE, sample diversity (mean pairwise distance, mean nearest-neighbor), and wall-clock time. The CDM$B_3/C_1$2Opt pipeline achieves $B_3/C_1$3, RMSE=0.017, and MAE=0.016 at runtime $B_3/C_1$4 s per 100 designs. High-fidelity FUN3D validation confirms surrogate predictions with $B_3/C_1$5 (Sung et al., 2 Dec 2025).

5. Advanced Surrogate Architectures and Few-Shot Regimes

GeoFunFlow-3D introduces a physics-guided generative flow matching framework and evaluates its data efficiency on BlendedNet (Jiang et al., 25 Apr 2026). In few-shot conditions (100 or 500 training samples), the framework leverages:

• Geometry-aware latent representations via GNO and Nadaraya–Watson regression.
• Optimal-transport interpolation with a flow matching loss, provably minimizing kinetic energy and avoiding mode collapse (as established by fitting multimodal $B_3/C_1$6).
• No-AD fourth-order discrete differential engines to suppress spectral bias and fit high-frequency surface detail, critical for physical consistency in shock/boundary layers.
• A SATO super-resolution module applying neural MsFEM basis functions and phase masking to reconstruct sharp localized features.

Quantitative results show that with only 100 samples, GeoFunFlow-3D achieves $B_3/C_1$7 MAE of $B_3/C_1$8 (compared to $B_3/C_1$9 for conditional DDPM). With 500 samples, error rapidly plateaus near the FAE warm-up bound, demonstrating stable multimodal fitting without mode collapse (Jiang et al., 25 Apr 2026).

6. Applications, Limitations, and Reproducibility Protocols

BlendedNet supports rapid, data-driven aerodynamic analysis for BWB planforms, including gradient-free and gradient-based optimization, uncertainty quantification, and active learning. It forms a standard pretraining corpus for graph and field-based models, and underpins inverse design of unconventional shapes.

Limitations identified include:

• The presence of unphysical geometries at sampled parameter extremes.
• Surrogate difficulty in capturing very sharp discontinuities or separated flows (noted for modest rises in error in such regions).
• Initial releases omit full 3D volumetric flow fields, focusing on surface data; planned extensions include volume fields with richer encoding schemes and manufacturability constraints (Sung et al., 8 Sep 2025, Sung et al., 2 Dec 2025).

Standardized protocols mandate:

• Geometry-disjoint train/test splits.
• Uniform evaluation and preprocessing.
• Reporting of error metrics across all predicted fields.
• Provision of codebases and scripts for full reproducibility.

BlendedNet and BlendedNet++ together establish the leading open benchmarks for machine learning approaches in field-level aerodynamic prediction and data-driven aircraft design, supporting objective comparison and future methodological advances (Sung et al., 8 Sep 2025, Sung et al., 2 Dec 2025, Jiang et al., 25 Apr 2026).