https://github.com/arthur-kitsuragi/fluiddiff
A Denoising Diffusion Model for Fluid Field Prediction
https://github.com/arthur-kitsuragi/fluiddiff
diffusion-model fluid-simulation machine-learning phiflow tensorflow
Last synced: 5 months ago
JSON representation
A Denoising Diffusion Model for Fluid Field Prediction
- Host: GitHub
- URL: https://github.com/arthur-kitsuragi/fluiddiff
- Owner: Arthur-Kitsuragi
- Created: 2025-08-30T14:48:14.000Z (5 months ago)
- Default Branch: main
- Last Pushed: 2025-08-30T17:46:29.000Z (5 months ago)
- Last Synced: 2025-08-30T19:09:53.108Z (5 months ago)
- Topics: diffusion-model, fluid-simulation, machine-learning, phiflow, tensorflow
- Language: Jupyter Notebook
- Homepage:
- Size: 75.1 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# FluidDiff
This repository contains an implementation of **FluidDiff**, a diffusion-based model for fluid dynamics.
The project includes dataset generation, model training, and example usage.
Original paper: [link to the article](https://arxiv.org/pdf/2301.11661)
---
## 📂 Repository Structure
- **`FluidDiff.py`** — core implementation of the FluidDiff model (can be imported as a module).
- **`FluidDiff_dataset.ipynb`** — dataset generation using [PhiFlow](https://github.com/tum-pbs/PhiFlow).
- **`FluidDiff_train.ipynb`** — training notebook for the FluidDiff model.
- **`example.py`** — minimal example of model inference.
- **`example_data.npy`** — example dataset used for testing and inference.
- **`weights.40.weights.h5`** — trained model weights after 40 epochs.
---
📌 Notes
Designed for fluid field simulations (e.g., velocity, pressure).
Trained on synthetic data generated with PhiFlow.
Achieved MAE = 0.1 after 40 epochs.
To predict future condition of fluid use `ddm.generate(n, diffusion_time, data)`,
where `n` is amount of batches, `diffusion_time` is amount of diffusion steps during
sampling, `data` is TensorFlow Tensor (shape = (N, 64, 64, 4), where N is number of pictures,
the first and the second channels should be Noise, the third one - ρ(0), the fourth one -
64x64 matrix where each element = t/40 (t can be 1,...40; in fact it's simulation time))
## Example Generation
Below is an example of FluidDiff denoising a fluid field:
## Architecture
