Flow reconstruction is an important problem across the physical, biological, and engineering sciences. Given some sensor sensor measurements (e.g., orange points in the image below) the aim is it to reconstruct the corresponding high-dimensional flow field.
Traditionally, modal approximation techniques, such as the proper orthogonal decomposition (POD), are used for this task. Here, we propose a neural network-based methodology. Indeed, the reconstruction quality of the shallow decoder outperforms traditional methods. For instance, the following image shows the reconstructed flow field using only 5 sensors:
For comparison, the reconstructed flow field using the traditional POD technique:
Start by cloning the git repository:
git clone https://github.com/erichson/ShallowDecoder
Then, you can run a simple examples as follows (without plotting the results):
python flow_driver.py --data flow_cylinder --sensor wall --n_sensors 5 --epochs 4000
Or optional you can plot the results:
python flow_driver.py --data flow_cylinder --sensor wall --n_sensors 5 --epochs 4000 --plotting True
- PyTorch
- scikit-learn
- For plotting you need in addition the cmocean package:
pip install cmocean
Coming soon!