Atmospheric retrieval of exo-atmospheres using machine learning
Thursday 5 July, 09:00
Atmospheric retrieval solves the inverse problem of obtaining chemical abundances and other properties given a measured spectrum of the atmosphere of an exoplanet. Traditionally, there is a compromise between physical/chemical realism and computational efficiency in order to sample millions (or more) of models and compute the posterior distributions of parameters. We report on a new approach to achieve significant gains in both aspects based on novel applications of machine learning. Since the machine learning procedure may be trained on pre-computed grids, we may obviate concerns of physical/chemical self-consistency and use grids from other research groups as well. Natural, added-value outcomes of machine-learning retrieval include information and sensitivity analyses. For example, one may estimate the percentage information content of a specific data point in a spectrum with regards to retrieving the mixing ratio of water. In this talk, I report on the application of machine-learning retrieval to both WFC3 transmission spectra, brown dwarf spectra and mock JWST spectra. The machine-learning procedure may be trained in minutes and each spectrum may be analyzed in milli-seconds. These novel results are the outcome of a long-term collaboration between astrophysicists and medical computer scientists at the University of Bern, Switzerland, and are adapted from methods that are used to control surgical tools and perform medical image analysis in cancer diagnosis. Machine-learning retrieval team: Chloe Fisher (astrophysics), Kevin Heng (astrophysics), Pablo Marquez Neila (medicine and biomedical engineering), Raphael Sznitman (medicine and biomedical engineering).