Tackling Sparse, Multiwavelength Light Curves in Time-Domain Astronomy with Neural Processes

Because of current surveys like the Zwicky Transient Facility (ZTF), Transiting Exoplanet Survey Satellite (TESS), and Gaia, as well as future surveys like the Nancy Grace Roman Space Telescope (Roman) and the Rubin Observatory Legacy Survey of Space and Time (LSST), time-domain astronomy is in a period of revolution. These surveys are producing an unprecedented wealth of data, critical for time-domain science across all wavelengths. A significant challenge for science, however, is the inherent nature of this data: these multi-band light curves, while many in number, are often sparse, asynchronous, with different cadences, resolutions, and observational gaps across filters.

While Gaussian Process (GP) regression is a common method for interpolating light curves, it requires specifying a covariance kernel a priori. This makes a strong assumption about the object's temporal and chromatic nature, an assumption that may be inappropriate for diverse transient populations and could fail to represent "true novelties".

The sheer volume of survey data, however, enables a more flexible, data-driven approach. Neural Processes (NPs) are a family of models that learn distributions over functions, combining the probabilistic framework of GPs with the scalability and power of neural networks. Crucially, NPs allow the covariance kernel to be learned directly from the data via an ANN, rather than being imposed by the user.

In this talk, I will discuss my ongoing work on the neural process family of models for light curve interpolation. I will demonstrate how this more flexible representation naturally incorporates multi-wavelength data, and can aid key downstream astronomical tasks from feature extraction and parameter estimation to classification and anomaly searches.