Aircraft Diffusion CFD - CLI Project

A production-ready command-line tool for generating and optimizing aircraft structures using diffusion models and GPU-accelerated CFD simulation.

Overview

This CLI project combines generative diffusion models, hierarchical structural representation (HRM), and GPU-accelerated computational fluid dynamics to automatically design viable aircraft structures optimized for aerodynamic efficiency.

Key Features

• 🚀 Fast Training: Progressive grid refinement (16³ → 24³ → 32³) with only 4-6 hours on A100
• 💾 Memory Efficient: Optimized for 8-13GB VRAM using gradient checkpointing & sparse grids
• ✈️ Aerodynamic Optimization: Built-in CFD simulator evaluates drag, lift, and structural constraints
• 🎯 Connectivity Constraints: Ensures generated designs are structurally viable
• 📊 Real-time Monitoring: TensorBoard integration for training visualization
• 📦 STL Export: Convert volumetric designs to production-ready meshes

Quick Start

1. Installation

# Clone the repository
git clone <your-repo-url>
cd research-paper/CLI

# Create and activate virtual environment
python -m venv venv
venv\Scripts\activate  # On Windows

# Install dependencies
pip install -r requirements.txt

2. Verify Setup

python aircraft_diffusion_cfd.py info

Expected output shows PyTorch version, CUDA availability, and GPU memory.

3. Train a Model

python aircraft_diffusion_cfd.py train \
  --num-epochs 100 \
  --batch-size 4 \
  --num-samples 100 \
  --save-dir ./checkpoints

4. Generate Designs

python aircraft_diffusion_cfd.py generate \
  --checkpoint checkpoints/final_model.pt \
  --num-samples 10 \
  --output-dir ./designs

5. Evaluate a Design

python aircraft_diffusion_cfd.py evaluate \
  --design designs/design_0.npy \
  --cfd-steps 1000

Commands Reference

train

Train the diffusion model from scratch or resume from checkpoint.

Key Arguments:

• --num-epochs (int): Training epochs at full resolution (default: 100)
• --batch-size (int): Batch size; adjust based on VRAM (default: 4)
• --learning-rate (float): Adam optimizer learning rate (default: 2e-4)
• --latent-dim (int): Latent space dimensionality (default: 128)
• --disconnection-penalty (float): Penalty for disconnected structures (default: 10.0)
• --num-samples (int): Synthetic training data samples (default: 100)
• --resume-from (str): Path to checkpoint to resume training
• --save-dir (str): Directory for saving checkpoints (default: ./checkpoints)

Example:

python aircraft_diffusion_cfd.py train \
  --num-epochs 150 \
  --batch-size 3 \
  --learning-rate 1e-4 \
  --disconnection-penalty 15.0 \
  --num-samples 200

generate

Generate new aircraft designs using a trained model.

Key Arguments:

• --checkpoint (str): Path to trained model checkpoint (required)
• --num-samples (int): Number of designs to generate (default: 10)
• --grid-size (int): Voxel grid resolution (default: 32)
• --output-dir (str): Directory for saving generated designs (default: ./designs)
• --guidance-scale (float): Classifier-free guidance strength (default: 7.5)

Example:

python aircraft_diffusion_cfd.py generate \
  --checkpoint checkpoints/final_model.pt \
  --num-samples 50 \
  --guidance-scale 5.0 \
  --output-dir ./generated_designs

evaluate

Run CFD evaluation on a single design.

Key Arguments:

• --design (str): Path to .npy design file (required)
• --cfd-steps (int): CFD simulation steps (default: 1000)
• --output-file (str): Save evaluation results (optional)

Example:

python aircraft_diffusion_cfd.py evaluate \
  --design designs/design_0.npy \
  --cfd-steps 2000 \
  --output-file evaluation.json

export

Convert voxel designs to STL mesh format.

Key Arguments:

• --design (str): Path to .npy design file (required)
• --output (str): Output STL file path (default: design.stl)
• --simplify (bool): Apply mesh simplification (default: false)

Example:

python aircraft_diffusion_cfd.py export \
  --design designs/design_0.npy \
  --output aircraft_design.stl \
  --simplify true

info

Display system information and GPU/CUDA status.

python aircraft_diffusion_cfd.py info

System Requirements

Hardware

• GPU: NVIDIA CUDA-capable GPU with 8GB+ VRAM
  • 8GB: 16³ training only
  • 10-13GB: Full 16³ → 32³ progressive training
• CPU: Multi-core processor (6+ cores recommended)
• RAM: 16GB+ system RAM

Software

• Python: 3.9+ (3.10/3.11 recommended)
• CUDA: 11.8+ or 12.x
• cuDNN: 8.7+

Dependencies

See requirements.txt. Key packages:

• PyTorch ≥ 2.0.0
• NumPy ≥ 1.24.0
• SciPy ≥ 1.10.0
• scikit-image ≥ 0.22.0
• TensorBoard ≥ 2.13.0
• TrimMesh ≥ 3.20.0

Training Performance

Progressive Training Schedule

Grid Size	Epochs	Memory	Time (RTX 3090)
16³	50	~3GB	2-3 hrs
24³	50	~6GB	4-5 hrs
32³	100	~10GB	8-10 hrs
Total	200	Peak 10GB	~14-18 hrs

Note: A100 GPUs are ~2-3x faster than RTX 3090

Configuration

Config File (config.yaml)

diffusion:
  timesteps: 100
  beta_start: 0.0001
  beta_end: 0.02
  sampling_timesteps: 250
  guidance_scale: 7.5

model:
  latent_dim: 128
  encoder_channels: [32, 64, 128, 256]
  decoder_channels: [256, 128, 64, 32]

training:
  batch_size: 4
  learning_rate: 0.0002
  num_epochs: 100
  disconnection_penalty: 10.0

cfd:
  reynolds_number: 1e5
  mach_number: 0.3
  simulation_steps: 1000

Project Structure

CLI/
├── aircraft_diffusion_cfd.py    # Main CLI entry point
├── advanced_lbm_solver.py        # GPU-accelerated CFD simulator
├── requirements.txt              # Python dependencies
├── config.yaml                   # Default configuration
├── QUICKSTART.md                 # 5-minute setup guide
├── ARCHITECTURE.md               # Technical deep dive
├── README.md                     # Original detailed README
├── checkpoints/                  # Trained model checkpoints
├── runs/                         # TensorBoard logs
└── reference/                    # Reference implementations
    └── complete_amr_d3q27_cascaded_guide.py

Examples

Example 1: Train from Scratch

python aircraft_diffusion_cfd.py train \
  --num-epochs 100 \
  --batch-size 4 \
  --num-samples 100

Example 2: Generate 100 Designs

python aircraft_diffusion_cfd.py generate \
  --checkpoint checkpoints/final_model.pt \
  --num-samples 100 \
  --guidance-scale 7.5

Example 3: Full Pipeline

# Train
python aircraft_diffusion_cfd.py train --num-epochs 100

# Generate
python aircraft_diffusion_cfd.py generate \
  --checkpoint checkpoints/final_model.pt \
  --num-samples 20

# Evaluate best design
python aircraft_diffusion_cfd.py evaluate \
  --design designs/design_0.npy \
  --cfd-steps 2000

# Export to STL
python aircraft_diffusion_cfd.py export \
  --design designs/design_0.npy \
  --output best_aircraft.stl

Monitoring Training

Real-time training metrics are logged to TensorBoard:

tensorboard --logdir ./runs

Then open http://localhost:6006 in your browser to view:

• Loss curves (total, connectivity, aerodynamic)
• Learning rate schedule
• Design sampling previews
• CFD simulation results

Troubleshooting

Out of Memory (OOM)

• Reduce --batch-size to 2 or 1
• Reduce --num-samples to 50
• Start with 16³ grid training only

Slow Training

• Ensure CUDA is available: python aircraft_diffusion_cfd.py info
• Check GPU utilization with nvidia-smi
• Reduce --num-samples to speed up data loading

Poor Design Quality

• Increase --num-epochs to 150+
• Reduce --disconnection-penalty if too restrictive
• Ensure training has converged (check TensorBoard)
• Try higher --guidance-scale in generation

Citation

If you use this project in your research, please cite:

@software{aircraft_diffusion_cfd,
  title={Aircraft Diffusion CFD: Generative Design via Diffusion Models},
  author={Your Name},
  year={2025},
  url={https://github.com/yourusername/research-paper}
}

License

[Your License Here]

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Submit a pull request

Contact

For questions or support, open an issue on GitHub or contact the maintainers.

---

Last Updated: December 2025