ExoMDN is a machine-learning-based exoplanet interior inference model using Mixture Density Networks. The model is trained on more than 5.6 million synthetic planet interior structures. Given mass, radius, and equilibrium temperature, ExoMDN is capable of providing a full inference of the interior structure of low-mass exoplanets in under a second without the need for a dedicated interior model.

This repository contains the trained models shown Baumeister & Tosi 2023 (accepted in A&A), as well as Python notebooks to load the models and run interior predictions of exoplanets. Interactive widgets are included to simplify loading an MDN model and running a prediction. We also make available the training routines in more_examples/model_training_demo.ipynb.

Create a new conda environment named exomdn from the environment.yml file, which installs all the required packages:

conda env create -f environment.yml

Activate the environment:

conda activate exomdn

Then install the exomdn package by running

pip install -e .

This project requires Python 3.7 or higher.

Install the required packages and dependencies using

pip install -r requirements.txt

Then install the exomdn package by running

pip install -e .

• python>=3.7
• tensorflow=2.11
• tensorflow-probability=0.15
• scikit-learn=1.1.1
• numpy
• pandas
• scipy
• matplotlib
• seaborn
• joblib
• ipywidgets
• jupyter

To get started check out introduction.ipynb. More in-depth examples can be found in the more_examples directory and more will be added over time. This directory also contains the training routines used to train the ExoMDN models, with some example trianing data to be found in the data directory.

We are using the MDN layer for Keras by https://github.com/cpmpercussion/keras-mdn-layer