2025 Theses Doctoral

Differentiable Models for High-Precision Time Series Observations of Planetary Systems

Cassese, Benjamin C.

Planetary astronomy has entered an era in which space-based observatories such as JWST now routinely send back observations with photometric and/or astrometric precisions orders of magnitude beyond what was possible just a few years ago. To overgeneralize, however, the complexity of the models used to explain these data, as well as the inference techniques used to apply those models, have not kept pace with these dramatic developments in data quality. Newer models, especially if they can be written in a differentiable framework that allows them to leverage recent advantages in artificial intelligence (AI)-related inference algorithms and computing hardware, offer two distinct promises in this landscape: the possibility of more accurate and precise inference, and the possibility of qualitatively new measurements that previously fell within the noise of older observations.

This thesis describes the development and application of several such models. These include a differentiable implementation of a high-precision N-body integrator; a framework for rapidly interpolating between stellar atmosphere models to compute limb darkening profiles during inference of transit parameters; a generalization of a transiting planet model to include non-spherical planets; and a novel likelihood-based framework for reducing raw JWST data. Each of these models are written in JAX, an autodifferentiable framework that powers much of Google's recent AI advances. The latter three of these models, which form the majority of this thesis, were combined together to attempt to measure the oblateness of a Jupiter-like transiting exoplanet.