🌊 SwiftCFD — Neural Network Flow Field Predictor

SwiftCFD is a deep learning surrogate model for Computational Fluid Dynamics (CFD). It predicts steady-state laminar flow fields (Ux, Uy, pressure) around 2D obstacles in ~50ms — replacing hours of traditional CFD simulation.

---

🚀 Live Demo

👉 Try SwiftCFD on Hugging Face Spaces

---

📦 Resources

Resource	Link
🤗 Model weights	vamsigudipati/deepcfd-model
🌊 Live Demo	SwiftCFD Space
📊 Dataset	DeepCFD — Zenodo
📄 Original Paper	DeepCFD — arXiv

---

🏗️ Architecture

• Model: UNetEx (U-Net with skip connections + max unpooling)
• Input: 3-channel tensor [SDF, Ux_inlet, Uy_inlet] — shape (3, 172, 79)
• Output: 3-channel tensor [Ux, Uy, p] — shape (3, 172, 79)
• Filters: [8, 16, 32, 32] | Kernel size: 5×5
• Parameters: ~500K | Model size: ~3.3MB

---

📊 Training

	Pass 1	Pass 2 (Fine-tune)
Script	train_pass1.py	train_finetune.py
Optimizer	Adam	AdamW (weight_decay=0.02)
Learning rate	1e-3	1e-4
Batch size	64	64
Scheduler	ReduceLROnPlateau (patience=20)	ReduceLROnPlateau (patience=50)
Early stopping	patience=100	patience=200
Max epochs	2000	2000
Best Val MSE	—	0.739

---

📈 Results

Field	R²	Slope
Ux	0.9974	0.9977
Uy	0.9960	0.9983
p	0.7552	0.7647

---

🛠️ Setup

1. Clone the repo

git clone https://github.com/vamsigudipati/SwiftCFD.git
cd SwiftCFD

2. Install dependencies

pip install -r requirements.txt

3. Download the dataset

wget https://zenodo.org/record/3666056/files/DeepCFD.zip
unzip DeepCFD.zip
# → dataX.pkl and dataY.pkl

---

🏋️ Training

Pass 1 — Train from scratch

python train_pass1.py \
  --dataX dataX.pkl \
  --dataY dataY.pkl \
  --output checkpoint.pt

Pass 2 — Fine-tune

python train_finetune.py \
  --checkpoint checkpoint.pt \
  --dataX dataX.pkl \
  --dataY dataY.pkl \
  --output mymodel_v2.pt

---

📊 Visualizations

# Plot sample predictions + scatter + training curves
python visualize.py \
  --checkpoint mymodel_v2.pt \
  --dataX dataX.pkl \
  --dataY dataY.pkl \
  --samples 103 487 354 563 520 \
  --out_dir ./assets

# Pass 1 vs Pass 2 comparison
python visualize.py \
  --checkpoint mymodel_v2.pt \
  --checkpoint1 checkpoint.pt \
  --dataX dataX.pkl \
  --dataY dataY.pkl \
  --plot comparison

---

🖥️ Run the App Locally

streamlit run app.py

---

📁 Repository Structure

SwiftCFD/
├── app.py                  ← Streamlit demo app
├── train_pass1.py          ← Pass 1: train from scratch
├── train_finetune.py       ← Pass 2: fine-tuning
├── visualize.py            ← Visualization scripts
├── requirements.txt        ← Python dependencies
├── README.md               ← This file
├── DOCUMENTATION.md        ← Full project documentation
├── LICENSE                 ← MIT License
├── .gitignore
└── assets/                 ← Result plots and snapshots
    ├── pass1_sample0.png
    ├── pass1_sample1.png
    ├── pass1_sample2.png
    ├── training_curves_v2.png
    ├── finetune_sample103.png
    ├── finetune_sample487.png
    ├── finetune_sample354.png
    ├── finetune_sample563.png
    ├── finetune_sample520.png
    └── scatter.png

Note: Model weights (*.pt) and dataset (*.pkl) are NOT stored in this repository. They are hosted on Hugging Face and downloaded automatically at runtime.

---

⚠️ Known Limitations

• Best results when obstacle is placed in the left-center third of the domain (x = 30–80), matching the training data distribution
• Fixed input resolution of 172 × 79 cells
• Laminar flow only (low Reynolds number)
• 2D predictions only

See DOCUMENTATION.md for full details.

---

🙏 Acknowledgements

Based on DeepCFD by Ribeiro et al.

@article{deepcfd2020,
  title   = {DeepCFD: Efficient Steady-State Laminar Flow Approximation
             with Deep Convolutional Neural Networks},
  author  = {Ribeiro, Mateus D. and Rehman, Auwal and Ahmed, Shahood
             and Dengel, Andreas},
  journal = {arXiv preprint arXiv:2004.08826},
  year    = {2020}
}