Date: Wednesday 13-Sep-2023
Time: 11:30-12:30 pm
Location: PSC 1136
Speaker: Arjun Savel (University of Maryland)
Title: Knowing when to know: bridging data-driven and physics-driven modeling for exoplanet atmospheres
The study of exoplanet atmospheres has long grappled with the difficulties inherent to low signal-to-noise data. Even as observations have become more constraining, ambitions for what constitutes "signal" have become proportionally larger. Recent high-resolution cross-correlation spectroscopy (HRCCS) studies exemplify this trend by aiming to precisely measure the abundance of multiple gasses in exoplanet atmospheres. Given the opaque nature of HRCCS data processing, it is unclear where biases in these inferred abundances may emerge. In this talk, I present my work done at the Flatiron Institute last fall in the Pre-Doctoral Program. While there, I investigated techniques for empirically diagnosing and empirically improving HRCCS. By forward-modeling the complex and non-linear data-processing of HRCCS, I demonstrated that two observing regimes — high photon noise and high telluric variability — can yield biased inferences of atmospheric properties. Beyond these results, I applied data-driven methods to improve atmospheric parameter inference, which is rapidly becoming a computationally intractable problem. Taking this approach produced an order of magnitude increase in memory efficiency and an order of magnitude decrease in compute hours. While describing this work, I’ll also provide an overview of useful data-driven approaches to Bayesian sampling, empirical modeling, and simulation optimization methods. By appropriately choosing when to model systems empirically in place of physically, it is possible to both comfortably push into lower signal-to-noise regimes and to infer more complex physics.
Guidelines and format for the CTC Lunch seminars.
Past CTC Lunch seminars.
For other questions, contact the CTC Lunch host: Matt Nixon
SPECIAL ACCOMMODATIONS:
Special accommodations for individuals with disabilities can be made by calling (301) 405-3001. It would be appreciated if we are notified at least one week in advance.