Neural network-based model approximation (nnbma)

Neural network-based model approximation nnbma is a Python package that handle the creation and the training of neural networks to approximate numerical models. In [1], it was designed and used to derive an approximation of the Meudon PDR code, a complex astrophysical numerical code.

Installation

To build your own neural network for your numerical model, we recommend installing the package. The package can be installed with pip:

pip install nnbma

To reproduce the results from [1], clone the repo with

git clone git@github.com:einigl/ism-model-nn-approximation.git

Alternatively, you can also download a zip file.

This package relies on PyTorch to build neural networks. It enables to evaluate any neural network, its gradient, and its Hessian matrix efficiently.

If you do not have a Python environment compatible with the above dependencies, we advise you to create a specific conda environment to use this code (https://conda.io/projects/conda/en/latest/user-guide/).

References

[1] Palud, P. & Einig, L. & Le Petit, F. & Bron, E. & Chainais, P. & Chanussot, J. & Pety, J. & Thouvenin, P.-A. & Languignon, D. & Beslić, I. & G. Santa-Maria, M. & Orkisz, J.H. & Ségal, L. & Zakardjian, A. & Bardeau, S. & Gerin, M. & Goicoechea, J.R. & Gratier, P. & Guzman, V. (2023). Neural network-based emulation of interstellar medium models. Astronomy & Astrophysics. 10.1051/0004-6361/202347074.