Anshuraj Sedai

Anshuraj Sedai

Biography

Anshuraj is a senior at Caldwell University, majoring in Computer Science and Mathematics with a minor in Physics. Fascinated by the majestic dark skies in his hometown, the southeast plains of Nepal, he developed a strong passion for astronomy and science communication. In his home institution, he served as the Founding President of Caldwell Astronomy Club which organizes various activities to promote astronomy on campus. Previously, he has conducted research on analyzing large datasets to trace the metal abundance during galaxy evolution. Similarly, he also worked on modeling the atmospheric desiccation and water loss function of M-dwarf exoplanets.At MIT, Anshuraj is working with Dr. Sai Ravela on exoplanet atmospheric retrieval to createan interpretability for machine learning surrogate models. He aspires to pursue a career as an astrophysicist in observatories or major research facilities. Besides academics, he enjoys watching and playing soccer, doing astrophotography, solving puzzles, and traveling.

Abstract

The Need for Speed (and Accuracy): Surrogate Models vs. Bayesian Methods in Exoplanet Atmospheric Characterization

The atmospheric characterization of exoplanets is crucial for assessing habitability beyond Earth. Current retrieval methods based on spectroscopy and photometry generate massive datasets, creating computational challenges that traditional Bayesian methods struggle to manage efficiently. Conversely, machine learning surrogate models offer significant computational speed but lack interpretability and may not fully capture complex atmospheric physics. This research addresses this trade-off by benchmarking surrogate model performance against comprehensive Bayesian retrieval methods. Synthetic atmospheric spectra were generated using a Markov Chain Monte Carlo (MCMC) forward-modeling approach coupled with the petitRADTRANS radiative transfer code. Surrogate models, built using neural network architectures and Gaussian processes via the MARGE library, were trained on these synthetic spectra. To evaluate surrogate model accuracy, predicted atmospheric characteristics were compared directly with Bayesian retrieval results derived from observational data of the well-studied exoplanet HD189733b. Preliminary analyses suggest that surrogate models closely match Bayesian retrieval for data similar to their training sets but struggle when conditions diverge from training conditions. Future work will explicitly identify and incorporate these missing physical processes into surrogate models, enhancing their interpretability and accuracy. The improved surrogate models promise rapid, reliable atmospheric characterization, benefiting future exoplanet exploration and advancing our understanding of planetary habitability.