This repository is a set a function implementing the method described in the paper: "Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations" by Julien Brajard, Alberto Carrassi, Marc Bocquet, and Laurent Bertino.
A example script is provided to run the "reference setup" described in the paper.
Tested on Linux/MacOS1
- Prerequisiste:
python3.8(suggest setting it up with anaconda). - Create a directory to save results:
mkdir example_data - Install the required python modules:
pip install -r requirements.txt - Run the example file example.py(you can modify the file to speed up the run):
python example.py
The code example.py will run the algorithm described in the paper for the standard setup (to run the other setups, you can modify the example.py code). The standard experiment run can take several hours.
The output of the code are saved on the example_data directory:
weights_init.h5: initial weights of the neural networkweights_nn.h5: weights of the neural network after optimizationsimulation.png: figure showing one simulation of 5 unit time steps (about 8 Lyapunov time steps).
If the algorithm has run at least once, and you have already produced weights saved in example_data/weights_nn.h5, you can run the code plot_simu.py to load the weights, make a simulation and a plot without the long optimization process: python plot_simu.py
In the file simulation.png, you should obtained the following figure:
1: For MacOS, the pythonw was used after installation through conda install python.app