An article published in “The Astrophysical Journal” reports the identifications of 335 new gravitational lens candidates discovered using machine learning software trained for this task. A team of astrophysicists led by Xiaosheng Huang of the University of San Francisco submitted images from the DECaLS investigation obtaining 335 possible gravitational lenses so far unknown. The verification will be carried out by humans, and 60 candidates have been included in the group that has the most chances of being confirmed. Gravitational lenses help astronomers in observing very far objects behind them, so the more are known the more likely they can be useful in some research.
To obtain a strong gravitational lens effect, the object’s gravity force must be so strong as to distort the light that passes close to it in a remarkable way. Even a galaxy may be too small to obtain an effect sufficient to be useful, and typically this type of lens is made up of a galaxy cluster. That effect is discovered in a very practical way, trying to examine other objects that are behind them from the observer’s point of view on Earth. It’s a method that takes a long time because it requires specific tests, so modern technologies come to the aid of astronomers, in particular machine learning software that runs on a supercomputer.
The Cori supercomputer at the Lawrence Berkeley National Laboratory (Berkeley Lab) was used for the training phase and the subsequent search for gravitational lenses. The researchers adopted a model adapted from the one created by François Lanusse and colleagues and described in an article published in January 2018 in the journal “Monthly Notices of the Royal Astronomical Society”. In the past years there were various developments in the software field and the development of a library used by the original software was abandoned, consequently, the researchers reimplemented the model in TensorFlow, the free / open source machine learning system created by Google.
The training took place by submitting 423 gravitational lenses and 9451 other objects to the software. When the software learned to recognize the gravitational lenses, the researchers made it analyze images of the Dark Energy Camera Legacy Survey (DECaLS). The result was the identification of 335 gravitational lens candidates.
The image (Dark Energy Camera Legacy Survey, Hubble Space Telescope) shows some gravitational lens candidates in the DECaLS color photos and black and white photos captured by the Hubble Space Telescope.
To be confirmed, the candidates must have certain characteristics and, according to how much they match them, they were divided into three groups: 60 with a high probability of being approved in group A, 105 with less pronounced characteristics in group B, and the 176 that match the least those characteristics in group C.
The machine learning software can be further perfected over time to help discover more gravitational lenses. The results will be useful for new observations of objects even billions of light-years away. More candidates might be sought in the future thanks to new investigations: for example, the Dark Energy Spectroscopic Instrument (DESI) survey, a step after DECaLS, is in its final stages of preparation.