From Star to Sensors: Differentiable Methods for Precision Astronomy

Abstract

This thesis introduces novel methods in astronomical data analysis based on differentiable forward modelling, enabling physical processes and statistical inference to be treated within a unified computational framework. By embedding optical propagation and detector physics into differentiable software, gradients can be propagated from photon arrival to detector readout, supporting principled calibration and inference. This approach is formalised in the "Pixels to Planets" philosophy: an endto- end framework that replaces traditional modular pipelines with continuous, interpretable models grounded in physics yet flexible enough for real data. A key outcome is dLux, an open-source library for differentiable optical modelling built in Jax. dLux supports efficient gradient computation through physically accurate models of optics and detectors, enabling uncertainty propagation across the instrument lifecycle. These tools are first applied to the Toliman mission, a low-cost astrometric cubesat targeting Earth-mass planets in the alpha Centauri