Github repo for "Neural ODEs as a discovery tool to characterize the structure of the hot galactic wind of M82," NeurIPS 2023 Machine Learning and the Physical Sciences Workshop.
The paper can be found here.
Additional details on neural galactic winds can be found in my first paper.
Plot of un-trained and trained neural ODE model applied towards mock data:
Plot of un-trained and trained neural ODE model applied towards M82 data for two velocity guesses.
This repository contains the source julia code alongside the Jupyter notebooks used to make many of the plots in the paper. There are two repositories: one for the mock data test and another for the comparison to M82 data. Each repository contains a .jl file. This file contains the code for defining the ODE system, incorporation of a neural network, and the optimization loop.
Please note, that the uniform Glorot inialization of the neural network has some randomness and there is not a specific key that is used for exact reproducability when running the code. To exactly reproduce the initialization I used in the paper, you can extract the saved neural networks parameters from the first line of the p_ADAM.jld2 file, which represents the untrained model in the paper. In most cases, the default Glorot initialization should give results that look identical to mine.
Below is an animation of the optimization over different epochs for the mock test: