Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C13]

Optimizing Astrochemistry Model Parameters Through Deep Learning Methods

Presenter: Zachary Mark Stomski

Faculty Sponsor: Andrew Burkhardt

School: Worcester State University

Research Area: Astronomy, Cosmology, and Astrophysics

ABSTRACT

Determining optimal physical conditions for astrochemical modeling is challenging due to the complexity and interdependence of the parameter space that governs them. Many of these parameters influence molecular abundances in complex, nonlinear ways, making manual tuning difficult and often subjective. Traditionally, researchers have addressed this challenge by adjusting only a subset of parameters deemed most influential, which could lead to biased or suboptimal results. To address this, we developed and trained a deep learning neural network to optimize physical parameters based on observed molecular abundances of astrophysical sources. The model creates predicted results using the three-phase NAUTILUS astrochemical model, which can produce accurate results of an astronomical source's molecular abundances over time given input parameters. By running NAUTILUS simulations with the model's predictions and comparing the results with observational data, the neural network iteratively refines parameter estimates to achieve the best possible fit. This approach streamlines the process for producing parameters, reduces human bias in parameter selection, enables more consistent and reliable simulations, and ultimately promotes comparability across publications. While our primary focus was fitting parameters for TMC-1, the model is adaptable to any astronomical source given the appropriate data. Using this method, we were able to half the average loss from outputted abundances compared to previous research's best parameter sets, and found major shifts in the parameters space that were required in order to do so. The model was unable to optimize to a physical parameter space without the use of restricting normalization techniques.

RELATED ABSTRACTS