Researchers develop model to predict properties of stars in seconds
Scientists are now a step closer to understanding the evolution of galaxies around us and the formation of stars. While traditionally, it may take millions of minutes to decipher the properties of a survey of star or galaxies, a study by scientists, including researchers from Tata Institute Fundamental Research (TIFR), has used artificial intelligence to teach a machine to predict these properties in a few seconds.
In a first-of-its-kind project, scientists claimed to use deep learning — a specific kind of machine learning that can learn mappings in the data — to predict properties of stars and galaxies in seconds. The project is an industry-academia collaboration between scientists from the National Centre for Radio Astrophysics (NCRA) of TIFR, software consultancy firm ThoughtWorks, and Centre for Modeling and Simulation of Savitribai Phule Pune University.
Their paper ‘Predicting star formation properties of galaxies using deep learning’ was accepted for publication on February 7 in the Monthly Notices of the Royal Astronomical Society, the UK.
“Astrophysicists have always been intrigued by the formation of stars to understand the evolution of the galaxy. We want to know whether stars are formed separately or together, how long they live, what their properties are and what kind of light they emit,” said Yogesh Wadadekar, a faculty member from NCRA and a co-author of the paper. He said the traditional method of studying stars comprise of complex physical models that are time consuming.
Shraddha Surana from ThoughtWorks, the lead author, said they used deep learning to predict three important star formation properties – stellar mass, star formation rate, and dust luminosity.
Surana said the model has been trained to emulate existing MAGPHYS model that predicts stellar properties, but faster. The model takes around 30 minutes on an average to be trained to predict the three stellar properties, following which it can predict the same in negligible time.
“For instance, to estimate the three star-formation parameters using the MAGPHYS code for 10,000 galaxies would take around 1,00,000 minutes. However, to predict the same number of galaxies using a deep learning model, it takes less than seconds. The only time taken up by the model is for training,” said Surana.
Wadadekar said the model will help enable a better understanding of galaxies. “The idea was to see if we can teach a machine to predict properties of stars by giving it large volumes of data. It is much faster than the traditional method,” he said.
Surana said the model can be developed further to survey and predict other properties too. “We don’t believe in patenting research work. One idea that we are discussing is to develop the model so that it can give a confidence value with each prediction. This means we can also quantify the uncertainty of the predictions,” she said.
The model will remain an open source material for other scientists to make use of.
Scientists who were not part of the study said the model will allow them to have a better understanding of the galaxy.
“The results are good and the model seems to have functioned well. It is an important first step in using artificial intelligence for getting to understand the galaxy better. Astronomy, as a field, is now moving into a mode where large data need to be studied and we need to look for statistical properties of the galaxy. Projects like this are the right step towards the future,” said Mayank Vahia, a senior astrophysicist from TIFR, Mumbai.