Keywords
Brown Dwarfs, Neural Networks, Atmospheric Retrieval
Abstract
In the era of the James Webb Space Telescope, there is a pressing need for atmospheric retrievals to be fast, accurate, and physically motivated. The computational cost of self-consistent atmosphere models has thus far made them unfeasible for use in retrieval. Machine learning techniques have greatly accelerated computational tasks in a variety of fields, and have been applied to retrieval before. However, previous application to brown dwarfs has focused on inferring parameters from spectra tailored to the resolution and wavelength range of a specific instrument.
We train a neural network on cloud-free model atmospheres computed self-consistently with PICASO, and use it as a surrogate forward model in Bayesian retrieval. Our neural network produces a model spectrum from 0.5 to 30 µm with R = 5000. We demonstrate the application of this model in re- trievals of simulated JWST observations, and of the benchmark T dwarf Gliese 570 D. We find that the neural network can be used to complete a retrieval in an amount of time comparable to a single PICASO model computation.
Completion Date
2025
Semester
Fall
Committee Chair
Karalidi, Theodora
Degree
Master of Science (M.S.)
College
College of Sciences
Department
Physics
Format
Identifier
DP0029811
Document Type
Thesis
Campus Location
Orlando (Main) Campus
STARS Citation
Braunschweig, Peter T., "Neural Network Accelerated Retrieval of Clear Brown Dwarf Atmospheres" (2025). Graduate Thesis and Dissertation post-2024. 517.
https://stars.library.ucf.edu/etd2024/517