Abstract

Classification of spectra (1) and anomaly detection (2) are fundamental steps to guarantee the highest accuracy in redshift measurements (3) in modern all-sky spectroscopic surveys. We introduce a new Galaxy Spectra Neural Network (GaSNet-III) model that utilizes neural networks to perform these three tasks simultaneously with high efficiency. Two different reconstruction networks – an autoencoder-like network and a U-Net – are used to reconstruct the rest-frame spectrum, which is then compared with the observed spectrum via a χ2 metric across the entire type and redshift spaces to find the best-fitting solution. SDSS DR16 spectra are used as a reference data set to provide a fully self-consistent science test to show that our model achieves accuracy comparable to that of classical principal component analysis-based methods, and even better in some specific metrics, while maintaining significantly higher efficiency. In particular, the model achieves an average of > 98 per cent classification accuracy across all classes, and redshift accuracies of over 99 per cent for stars, over 98 per cent for galaxies with errors of the order of O(10−4), and over 93 per cent for quasars with errors of the order of O(10−3). Tests on DESI spectra demonstrate that the model can generalize well to other surveys without retraining, with only a small degradation in performance. Furthermore, by comparing different peaks of χ2 curves, we define a robustness measure that enables the identification of anomalous spectra. The GaSNet-III provides accurate and high-efficiency spectrum modelling to perform accurate redshift estimates and anomaly detection in vast data volumes from future spectroscopic sky surveys. © The Author(s) 2025. Published by Oxford University Press on behalf of Royal Astronomical Society.