Shedding light on low-surface-brightness galaxies in dark energy surveys with transformer models

Context. Low-surface-brightness galaxies (LSBGs), which are defined as galaxies that are fainter than the night sky, play a crucial role in our understanding of galaxy evolution and in cosmological models. Upcoming large-scale surveys, such as Rubin Observatory Legacy Survey of Space and Time and Euclid, are expected to observe billions of astronomical objects. In this context, using semiautomatic methods to identify LSBGs would be a highly challenging and time-consuming process, and automated or machine learning-based methods are needed to overcome this challenge. Aims. We study the use of transformer models in separating LSBGs from artefacts in the data from the Dark Energy Survey (DES) Data Release 1. Using the transformer models, we then search for new LSBGs from the DES that the previous searches may have missed. Properties of the newly found LSBGs are investigated, along with an analysis of the properties of the total LSBG sample in DES. Methods. We created eight different transformer models and used an ensemble of these eight models to identify LSBGs. This was followed by a single-component Sersic model fit and a final visual inspection to filter out false positives. Results. Transformer models achieved an accuracy of similar to 94% in separating the LSBGs from artefacts. In addition, we identified 4083 new LSBGs in DES, adding an additional similar to 17% to the LSBGs already known in DES. This also increased the number density of LSBGs in DES to 5.5 deg-2. The new LSBG sample consists of mainly blue and compact galaxies. We performed a clustering analysis of the LSBGs in DES using an angular two-point auto-correlation function and found that LSBGs cluster more strongly than their high-surface-brightness counterparts. This effect is driven by the red LSBG. We associated 1310 LSBGs with galaxy clusters and identified 317 ultradiffuse galaxies among them. We found that these cluster LSBGs are getting bluer and larger in size towards the edge of the clusters when compared with those in the centre. Conclusions. Transformer models have the potential to be equivalent to convolutional neural networks as state-of-the-art algorithms in analysing astronomical data. The significant number of LSBGs identified from the same dataset using a different algorithm highlights the substantial impact of our methodology on our capacity to discover LSBGs. The reported number density of LSBGs is only a lower estimate and can be expected to increase with the advent of surveys with better image quality and more advanced methodologies.