Keywords
Atmospheric mapping; Brown dwarfs; Rotational light curves; Bayesian neural networks; Rapid atmospheric retrieval; Synthetic training data
Abstract
Mapping atmospheres using rotationally modulated light curves offers insights into cloud structures and dynamics. Current retrieval methods, primarily based on Markov Chain Monte Carlo (MCMC) techniques like Aeolus, can infer atmospheric features but are computationally prohibitive for large datasets. This project proposes a neural network (NN) framework for the rapid, variational inference of atmospheric structure from light curves, particularly those of brown dwarfs. The primary approach focuses on training a Bayesian NN (BNN) to perform regression, predicting the spot parameters that describe the object's surface brightness map. Given the scarcity of suitable observational training data, the BNN is trained on a large synthetic dataset generated using jaxoplanet. Since the problem has infinitely many possible solutions through spot parameters, the BNN's performance is validated by simulating its predicted parameters and comparing the resulting light curve against the original light curve it took as input. Independent BNNs are trained for spot counts of one through four, with the three-spot model used as a representative example when presenting detailed results. The BNN achieves a low mean squared error (MSE) loss on each light curve and performs inference thousands of times faster than MCMC. In addition, a convolutional classifier is trained on the same synthetic dataset to infer the number of spots directly from a light curve, achieving ~67% accuracy on a testing set and serving as a routing mechanism that directs each light curve to the BNN specialized for its predicted spot count.
Thesis Completion Year
2025
Thesis Completion Semester
Fall
Thesis Chair
Karalidi, Theodora
College
College of Sciences
Department
Department of Physics
Thesis Discipline
Physics - Astronomy
Language
English
Access Status
Open Access
Length of Campus Access
None
Campus Location
Orlando (Main) Campus
Notes
Dual degree:
B.S. Physics - Astronomy
B.S. Computer Science
This work is officially for astrophysics, but also involves computer science.
STARS Citation
Diaz, Eugenio A., "Rapid Inference of Atmospheric Feature Parameters from Light Curves Using Bayesian Neural Networks" (2025). Honors Undergraduate Theses. 421.
https://stars.library.ucf.edu/hut2024/421
Included in
Astrophysics and Astronomy Commons, Computer Sciences Commons, Numerical Analysis and Computation Commons
Rights Statement
