Characterizing exoplanetary atmospheres typically requires computationally intensive radiative transfer calculations to model their spectral signatures. This challenge is particularly true for ground-based high-resolution spectroscopy (R 50 000 - 100 000), which relies on line-by-line calculations across broad spectral ranges. As a result, retrieving atmospheric properties for a single target often requires days of computation on dedicated clusters, necessitating simplified assumptions such as isothermal structures and constant abundance profiles at the expense of physical realism. Model emulation offers a promising solution by training neural networks to reproduce radiative transfer outputs at a fraction of the computational cost. I will present our ongoing efforts to develop a neural-network-based emulator for the petitRADTRANS code, widely used in high-resolution spectroscopy studies, and discuss architectural choices and broader insights for surrogate modeling of complex physical systems.
Fast & Surrogateous : Speeding up Exoplanet Atmospheres Modeling with Neural-Network Emulators
Lundi 8 juin 2026
de
16:00 à
17:00
Salle de conférence, bâtiment 17 - Zoom
