Automated quasar continuum estimation using neural networks. A comparative study of deep-learning architectures

Ongoing and upcoming large spectroscopic surveys are drastically increasing the number of observed quasar spectra, making the development of fast and accurate automated methods to estimate spectral continua necessary. This study evaluates the performance of three neural networks (NNs) --- an autoencoder, a convolutional NN (CNN), and a U-Net --- in predicting quasar continua within the rest frame wavelength range of $1020 Å $ to $2000 Å $. The ability to generalize and predict galaxy continua within the range of $3500 Å $ to $5500 Å $ is also tested. We evaluated the performance of these architectures using the absolute fractional flux error (AFFE) on a library of mock quasar spectra for the WEAVE survey and on real data from the early data release observations of the Dark Energy Spectroscopic Instrument (DESI) and the VIMOS Public Extragalactic Redshift Survey (VIPERS). The autoencoder outperforms U-Net, achieving a median AFFE of 0.009 for quasars. The best model also effectively recovers the Lyα optical depth evolution in the DESI quasar spectra. With minimal optimization, the same architectures can be generalized to the galaxy case, with the autoencoder reaching a median AFFE of 0.014 and reproducing the D4000n break in DESI and VIPERS galaxies.