ABSTRACT

The field of scientific machine learning investigates the symbiosis of differential equations with deep learning. Using tools like physics-informed neural networks, dynamicists are able to discover novel solutions to complex differential equations using sparse and imperfect information. These advances can offer benefits in speed, accuracy, compactness, and more. This seminar will explore how the field of scientific machine learning offers a compelling new solution to the gravity modeling problem. Rather than relying on analytic basis functions which come with their own unique limitations, physics informed neural networks can be used to learn novel and flexible basis sets capable of representing complex gravitational perturbations in a variety of environments. These advances offer exciting applications within the field of astrodynamics spanning problems within orbit discovery, estimation, and reinforcement learning.

BIOGRAPHY

John Martin is a NSF graduate research fellow and PhD candidate in the Aerospace Engineering Sciences department at the University of Colorado Boulder. He earned his Bachelor’s degree in Physics and Astronomy from the University of North Carolina at Chapel Hill and his Masters in Aerospace Engineering from CU Boulder. During his PhD, John worked with the Laboratory for Atmospheric and Space Physics as part of the attitude determination and control subsystem team for a Mars orbiter, and he is currently an active developer for the open-source high-fidelity astrodynamics simulation software Basilisk. His research focuses on the intersection between scientific machine learning and astrodynamics, exploring how tools like physics informed neural networks can be used to find novel solutions to complex differential equations. and characterizing their corresponding advantages over prior analytic methods.