🌊 Learning Mesh-Based Simulation with Graph Networks

Fast, Adaptive, and Physics-Informed Neural Simulators for Complex Fluid Dynamics

This repository provides a PyTorch + PyG (PyTorch Geometric) implementation of MeshGraphNets—a powerful graph neural network framework for learning mesh-based physical simulations. We focus on the flow around a circular cylinder problem, reproducing and extending the groundbreaking work from DeepMind.

🔬 Original Paper:
Learning Mesh-Based Simulation with Graph Networks
Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, Peter W. Battaglia
ICLR 2021

---

✨ Why This Project?

• Physics-aware learning: Leverages mesh structure to respect geometric and physical priors.
• High performance: Runs 10–100× faster than traditional solvers while maintaining fidelity.
• Extensible: Built on PyTorch Geometric—easy to adapt to new PDEs, materials, or domains.

---

🛠️ Requirements

Install dependencies via:

pip install -r requirements.txt

💡 Note: TensorFlow < 1.15.0 is required only for parsing the original TFRecord datasets.

---

🚀 Quick Start

1. Download the Dataset

We use DeepMind’s cylinder_flow dataset:

aria2c -x 8 -s 8 https://storage.googleapis.com/dm-meshgraphnets/cylinder_flow/train.tfrecord -d data
aria2c -x 8 -s 8 https://storage.googleapis.com/dm-meshgraphnets/cylinder_flow/valid.tfrecord -d data
aria2c -x 8 -s 8 https://storage.googleapis.com/dm-meshgraphnets/cylinder_flow/test.tfrecord -d data

2. Parse TFRecords

Convert to PyTorch-friendly format:

python parse_tfrecord.py

Output saved in ./data/.

3. Train the Model

python train.py

FOR MULTI-GPU TRAINING:

export NGPUS=2 # set as your machine's available GPUs
torchrun --nproc_per_node=$NGPUS train_ddp.py --dataset_dir data

4. Run Rollouts & Visualize

Generate long-horizon predictions and render videos:

python rollout.py          # saves results to ./results/
python render_results.py   # generates videos in ./videos/

---

🎥 Demos

Results on DeepMind’s cylinder_flow:

Demo 0	Demo 1

Results on our own CFD-generated data (new geometries & conditions):

Demo 2	Demo 3

✅ The model generalizes well—even to unseen flow regimes and mesh configurations!

---

📬 Contact

Have questions, suggestions, or want to collaborate?
📧 Reach out: [email protected]

---

⭐ If you find this project useful, please consider starring the repo!
Your support helps us keep improving open-source scientific ML tools.