Baseline Model

Baseline Solution

The Baseline Solution can be found on our GitHub here

Inside this repo you will find the baseline solution for the Ariel Data Challenge. To run the script you will need access to the training and test data, both of which can be found on the website (after registration). There are two ways to run the baseline:

  • via jupyter notebook (also contains the evaluation metric), ADC2023-baseline.ipynb
  • via command line:


We trained a neural network to perform a supervised multi-target regression task. The architecture of the network is modified from the CNN network as described in Yip et al.

Preprocessing Steps

  • We used the first 5000 data instances to train the model
  • We augmented the data with the observation noise
  • Used stellar and planetary radii as additional features
  • Standardised both inputs and output

At test time we performed Monte Carlo Dropout to provide a mutlivariate distribution for each test example.


We have inlcuded the algorithm we used to compute the metric score in the jupyter notebook. Note that in order to compute the final score, you must have TauREx3 installed in your python environment, with the appropriate linelist, which can be found here. For installation procedure of TauREx3, please see here

For more information about the metric please see here.

Things to improve

There are different direction to take from here on, let us summarise the shortcoming of this model:

The data preprocessing is quite simplitic, we list the limitations below:

  • We have only used 5000 data points, instead of the full dataset.
  • We didnt train the model with results from the retrieval (Tracedata.hdf5), which are the GT for this competition.
  • The conditional distribution from MC-Dropout is very restricted and Gaussian-like.
  • So far we havent considered the atmospheric targets as a joint distribution.
  • We have only used stellar radius and planet radius from the auxillary information.
  • We have not done any hyperparameter tuning.