Use of Neural Network to speed up getTemDen and getcrossTemDen

[Christophe Morisset]

Dear PyNeb users,

With the increase of IFUs and the MonteCarlo techniques to follow the uncertainties, it may happen that one needs to compute Te and Ne for a lot of data points.

The getTemDen and getcrossTemDen are fast, but not that much when it has to deal with thousands of diagnostics.

The idea is to use Machine Learning tools to speed up the computation: a Neural Network is trained with examples of line ratios and Te/Ne values, and is then used to predict physical parameters from observations. The training is fast (10 tp 20 seconds), and may even be stored for next use (and reproductibility). 

To use this option in the methods, you first need to install the AI4Neb library:

pip install -U git+https://github.com/morisset/AI4neb.git
which in turn requires the installation of :

• keras : https://keras.io/getting_started/
• tensorflow: https://www.tensorflow.org/
• scikit-learn: https://scikit-learn.org/stable/

The notebook describing the use of this tool is there:

https://github.com/Morisset/PyNeb_devel/blob/master/docs/Notebooks/MonteCarlo_and_ANN.ipynb

Comments and feedbacks welcome,

Christophe