A brief introduction by Dr Ingo Waldmann, the co-PI of the Ariel data challenge, on the current demographics of exoplanets, recent findings and the use of AI in the field.

An introduction to the Ariel Space Mission by the PI Prof Giovanna Tinetti. She will discuss about the science of Ariel, its aim, capability and how the Ariel mission can push the field forward in the coming decades.

A brief introduction to this year's data challenge and a short summary of the results obtained from each track.

In this talk we present the solution to the Ariel Data Challenge of the “Gators” team from the Physics Department at the University of Florida. We train a model of interconnected neural networks to estimate a posterior distribution over possible exoplanetary atmospheric chemical compositions and surface temperatures from their transit spectrum and system auxiliary priors. A significant improvement of the model’s prediction was made by preprocessing the data using physically motivated feature engineering. The constructed model consists of several fully connected neural networks which use concatenations or products of the outputs of previous modules as inputs. To minimize the Wasserstein-2 distance while reducing the complexity of our model, we trained on a parameterization of the estimate of the posterior distribution. In cases when a concentration is too small to be detected, a functional term is added to reproduce the observed effect of the prior where the posterior distribution ends up being uniform across all compatible concentrations.

This talk will provide an overview of the second place solution for the Regular Track. The presentation will describe what method is used for modeling the distribution of planetary atmospheric properties, model architecture, how multi-task and semi-supervised learning are applied for the challenge solution.

Standard methods for inferring planetary characteristics from exoplanets’ atmospheric spectra are slow. We present a fast machine learning method: a deep ensemble of convolutional neural networks (CNNs) that outputs mixtures of normal distributions for planetary characteristics. The architecture of our CNN was inspired by VGG networks. We train each CNN with Kullback–Leibler divergence as its loss function on simulated exoplanet’ atmospheric spectra and their auxiliary data from Ariel Data Challenge. We expect that the performance of our deep ensemble would be worse on real data because machine learning methods assume that (in this case) both simulated and real are independent and identically distributed. However, it is highly probable that probability distributions of simulated and real data differ. We suggest using an active domain adaptation (ADA) method to mitigate the difference and thus improve performance on real data. In ADA, the deep ensemble (trained with simulated data) would query real spectra (that would improve its performance most) in rounds. In each round, queried real spectra might be annotated by a slow standard method or human annotator, and the deep ensemble would use them to improve its performance.

The challenge of predicting confidence intervals and distributions can be faced with several statistical techniques. The chosen method for my approach was a direct inference of a weighted combination of multivariate normal distributions along the 6 dimensions of the target variables, to provide information about mutual correlations, using a custom version of the Mixture Density Network, on top of a series of convolution layers for the spectra, and a dense layer for some engineered features from the auxiliary data, against an adapted version of the dataset.

I would like to introduce an overview of the idea which might be useful for someone.