BlendedNet: FiLM-Based Aerodynamic Field Prediction

This repository contains the codebase used in the ASME IDETC 2025 paper:

"BlendedNet: A Blended Wing Body Aircraft Dataset and Surrogate Model for Aerodynamic Predictions"
Presented at ASME IDETC/CIE 2025, Anaheim, CA

The project introduces a public high-fidelity dataset for blended wing body (BWB) aircraft, as well as a two-stage surrogate model combining PointNet and FiLM-based neural networks to predict pointwise aerodynamic coefficients.

The dataset is publicly available at: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/VJT9EP

Also, check out:

• Our paper on Arxiv: https://arxiv.org/abs/2509.07209
• A short video overview on YouTube: https://www.youtube.com/watch?v=-SkEXb0ndD0&t=1s

---

🚀 Highlights

• 999 BWB geometries × ~10 flight cases → 8,830 high-fidelity simulations
• Surface-level CFD quantities: Cp, Cfx, Cfz with corresponding point coordinates and normals
• PointNet-based encoder to recover geometric design parameters from a sampled surface
• FiLM-modulated neural field for predicting pointwise aerodynamic coefficients
• Detailed error metrics and R² plots included

---

📂 Repository Structure

.
├── models/
│   ├── film_model_v1.py                # Early FiLM model (ReLU, no residuals)
│   └── film_model_v2.py                # Final FiLM model (SIREN-style with sine + residuals)
├── dataset.py                          # Dataloading
│   train_model.ipynb  # Training pipeline (PointNet + FiLM)
│   test_model.ipynb # Evaluation
└── README.md

---

📊 Dataset Overview

BlendedNet is the first open dataset to provide high-resolution pointwise aerodynamic surface coefficients for BWB aircraft.

Each case contains:

• Geometric design parameters (9 shape parameters)
• Flight conditions (altitude, Mach, angle of attack, Reynolds length)
• CFD-derived outputs:
  • Cp (pressure coefficient)
  • Cf_x, Cf_z (skin friction in x/z)
  • Surface normals

Formats:

• .csv for metadata
• .h5 for coordinates, normals, and coefficients
• .vtk for postprocessing and visualization

The dataset will be hosted on Harvard Dataverse (link pending).

---

🧠 Surrogate Model

1. PointNet Regressor

• Input: Sampled point cloud of the aircraft
• Output: 9 geometric shape parameters
• Permutation-invariant design

2. FiLM Network

• Input: 3D coordinates (+ normals), flight conditions, and shape parameters
• Output: Cp, Cf_x, Cf_z at each surface point
• Modulation via learned scale/shift (gamma, beta)
• Residual connections and sine activations

---

🔧 How to Run

Train the Model

# Inside train_model.ipynb

Evaluate

# test_model.ipynb

---

📈 Performance

PointNet Parameter Prediction (R²)

Parameter	R²
C2/C1	0.9893
C3/C1	0.9896
C4/C1	0.9945
B1/C1	0.9923
B2/C1	0.9948
B3/C1	0.9997
S1	0.9968
S2	0.9914
S3	0.9973

FiLM Prediction Errors (Test Set)

Metric	Cp	Cfx	Cfz
Conditioned on Ground Truth Parameters
MSE	7.86e-03	2.80e-05	1.51e-05
MAE	3.72e-02	1.35e-03	7.96e-04
Rel L1 (%)	13.52%	22.09%	30.01%
Rel L2 (%)	3.11%	7.74%	18.79%
Conditioned on Predicted Parameters
MSE	1.19e-02	1.82e-04	5.72e-05
MAE	4.33e-02	1.98e-03	1.19e-03
Rel L1 (%)	14.99%	24.03%	31.53%
Rel L2 (%)	4.24%	16.78%	21.84%

---

📜 Citation

If you use this dataset or code, please cite:

@inproceedings{sung2025blendednet,
  title={BlendedNet: A Blended Wing Body Aircraft Dataset and Surrogate Model for Aerodynamic Predictions},
  author={Nicholas Sung and Steven Spreizer and Mohamed Elrefaie and Kaira Samuel and Matthew C. Jones and Faez Ahmed},
  booktitle={ASME IDETC/CIE},
  year={2025},
  address={Anaheim, CA},
  number={DETC2025-168977}
}

---

🛠 Acknowledgements

This material is based upon work supported under Air Force Contract No. FA8702-15-D-0001.

© 2025 Massachusetts Institute of Technology.

We also thank the MIT Lincoln Laboratory Supercomputing Center for their HPC resources.

---

📨 Contact

For questions, please contact:

• Nicholas Sung
  Department of Mechanical Engineering, MIT
  [email protected]