Overview
This project uses machine learning to approximate fluid flow. It began as an attempt to speed up topology optimization, where finding a suitable topology may require many computationally expensive simulations.
I developed the project with Ing. Lucie Kubíčková and Ing. Martin Isoz, Ph.D., from the Institute of Thermomechanics at the Czech Academy of Sciences. We published the results through the Topical Problems of Fluid Mechanics conference. The paper is available here.
What the images show
The bitmap above represents a topology. White areas are solid material and black areas are voids. In this fluid mixer, fluid enters from the left and exits on the right. The model predicts the velocity and pressure fields through this geometry, shown in the two images below.
Project details
We generated the training dataset with OpenFOAM and used its simulations as ground truth. Over almost a year, we tested RBF and MLP models, CNNs, autoencoders, and U-Net architectures. We evaluated each model with metrics and by inspecting its predicted velocity and pressure fields on topologies it had not seen during training. The results guided the next changes to the architecture and training process.
The project is still ongoing. Our recent work focuses on loss functions and physics-informed neural networks (PINNs), so that the model follows the relevant physical laws rather than only fitting the training data.
Technologies used
- Python
- PyTorch
- NumPy
- Matplotlib
- Jupyter Notebooks
- Hydra (for configuration management)
- MLflow (for experiment tracking)