gallifrey: JAX-based Gaussian process structure learning for astronomical time series

¹ Kapteyn Astronomical Institute, University of Groningen, Landleven 12 (Kapteynborg, 5419), 9747 AD Groningen, The Netherlands
² GELIFES Institute, University of Groningen, Nijenborgh 7, 9747 AG Groningen, The Netherlands

^★ Corresponding author: boettner@astro.rug.nl

Received: 13 March 2025
Accepted: 21 May 2025

Abstract

Context. Gaussian processes (GPs) have become a common tool in astronomy for analysing time series data, particularly in exoplanet science and stellar astrophysics. However, choosing the appropriate covariance structure for a GP model remains a challenge in many situations, limiting model flexibility and performance.

Aims. This work provides an introduction to recent advances in GP structure learning methods, which enable the automated discovery of optimal GP kernels directly from the data, with the aim of making these methods more accessible to the astronomical community.

Methods. We present gallifrey, a JAX-based Python package that implements a sequential Monte Carlo algorithm for Bayesian kernel structure learning. This approach defines a prior distribution over kernel structures and hyperparameters, and efficiently samples the GP posterior distribution using a novel involutive Markov chain Monte Carlo procedure.

Results. We applied gallifrey to common astronomical time series tasks, including stellar variability modelling, exoplanet transit modelling, and transmission spectroscopy. We show that this methodology can accurately interpolate and extrapolate stellar variability, recover transit parameters with robust uncertainties, and derive transmission spectra by effectively separating the background from the transit signal. When compared with traditional fixed-kernel approaches, we show that structure learning has advantages in terms of accuracy and uncertainty estimation.

Conclusions. Structure learning can enhance the performance of GP regression for astronomical time series modelling. We discuss a road map for algorithmic improvements in terms of scalability to larger datasets, so that the methods presented here can be applied to future stellar and exoplanet missions such as PLATO.