Statistics and Data Challenges

Atmospheric retrievals are statistical methods requiring the exploration of distance functions (likelihood) over large dimentional parameter space. This task is very computational, demanding advanced statistical methods and/or machine learning approach. At ExoAIM, we explore these aspects by collaborating with experts of these fields, and by conduting open data challenges to search for novel methods.

With the advent of JWST, the data information content has increased drastically, requiring more complex forward models to be fitted to the data. This implies that larger parameter spaces needs to be explored, with models that are computationally more demanding. To overcome these challenges, modern retrieval techniques have used alternative techniques to the traditional Nested Sampling approach.

We are involved in developing novel methods, including Variational Inference and Surrogate Models. Variational inference leverages differentiable forward models to transform the sampling problem into an optmization problem. It is a promising approach to accelerate Bayesian atmospheric retrievals.

We also participate in the definition and organization of data challenges. These events open exoplanet problems to alternative communities. We promote the participation of experts in machine learning, statistics, and other methods to find novel techniques. In the 2022 Ariel Data Challenge, participants were tasked to solve the atmospheric retrieval problem using alternative AI-powered approaches.

Check out the Ariel data challenge website: Official Website