Abstract

Contributed Talk - Splinter EScience

Tuesday, 16 September 2025, 15:33

Back to the feature: Reconstructing Spectra from Photometry

Johanna Riedel, Nikolaos Gianniotis, and Kai Lars Polsterer
Astroinformatics - HITS gGmbH

The reconstruction of a full spectrum from photometry is an ill posed problem. Nevertheless, this is a task worth undertaking since photometry is more abundant and easier to obtain than spectroscopic data. In our approach, we train an ensemble of neural network decoders that learn how to represent high-dimensional observed spectra in terms of a low-dimensional latent space. Once trained, we can query the model to generate a spectrum that, when passed through a given set of photometric filters, reproduces the observed photometry. Our model can learn from spectra that contain missing values, it is scale-invariant to the amplitude of the spectra and can be used with any set of filters without the need to retrain.