Machine learning and classification could help astronomers better understand and classify the Universe’s most explosive events.
Supernovae, massive cosmic explosions that represent the final death throes of stars, offer astronomers and cosmologists a vital tool for understanding the Universe. In particular, one type of these massive explosions — Type Ia supernovae — can be used to measure distances in the depths of space. Aside from this, learning more about supernovae can tell us how stars live and die and how elements are dispersed throughout galaxies.
Currently, supernovae are studied by using their observed spectra — the set of colors into which light from these objects can be split — which contains characteristic ‘gaps’ that tell astronomers what light is being emitted and absorbed, and thus, which elements are present in the explosion’s remains.
That method of study could soon be joined by another; however, one which relies on a platform of artificial intelligence (AI) and could allow more of these events to be classified. A team of astronomers from the Center for Astrophysics (CfA), Harvard & Smithsonian, is pioneering the use of AI to classify actual supernovae without the use of their spectra.
The team trained a machine learning model to categorize supernovae based on their light curves — their light intensity and brightness over a period of time represented as a graph — classifying real data from Panoramic Survey Telescope And Rapid Response System 1 (Pan-STARRS1, PS1) Medium Deep Survey for over 2000 supernovae. The CfA astronomers achieved an accuracy rate of 82 percent without the use of spectra.
The team’s research is published across two papers in the latest edition of the Astrophysical Journal. Not only this, but the data sets that the team created using their new method and the classifications complied with its use are freely available for open use.
Nothing Better than the Real Thing for AI Training
In order to train their AI, the team used around 2,500 light curves from data collected by Pan-STARRS1 Medium Deep Survey — a multi-wavelength observation program that operated between 2010 and 2014. From this data, the astronomers were able to select 500 supernovae with spectra that could be used for classification.
“We trained the classifier using those 500 supernovae to classify the remaining supernovae where we were not able to observe the spectrum,” explains Griffin Hosseinzadeh, a postdoctoral researcher at the CfA and lead author on one of the aforementioned papers.
Hosseinzadeh’s colleague at CfA, Edo Berger, explains that the team improved the accuracy of their results by asking the AI to answer specific questions.
The machine learning looks for a correlation with the original 500 spectroscopic labels. We ask it to compare the supernovae in different categories: color, rate of evolution, or brightness. By feeding it real existing knowledge, it leads to the highest accuracy, between 80- and 90-percent.
Edo Berger, Professor of Astronomy, CfA
The project is not the first time that AI and machine learning has been employed in the classification of supernovae, but it does represent the first time that researchers have had access to a real data set that is sufficiently large enough to train an AI as a supernovae classifying system. The consequence of this is the possible creation of machine learning algorithms that don’t depend upon the use of simulations.
Nature is Full of Surprises — Even for Supernovae
The benefit of training an AI with data collected from real supernovae is that that it equips the system with the ability to handle the unexpected surprises that nature can sometimes throw-up without classification errors.
If you make a simulated light curve, it means you are making an assumption about what supernovae will look like, and your classifier will then learn those assumptions as well. Nature will always throw some additional complications in that you did not account for, meaning that your classifier will not do as well on real data as it did on simulated data.
Griffin Hosseinzadeh, Postdoctoral Researcher, CfA
Building a system that can handle the quirks that a real data set can occasionally present means that the high degree of accuracy the AI classifier delivered is likely to be replicated when it is presented with another real cosmic survey.
“We will be able to study them both in retrospect and in real-time to pick out the most interesting events for detailed follow up,” adds Berger. “We will use the algorithm to help us pick out the needles and also to look at the haystack.”
The team will now unleash their AI on the mountain of archival data collected by previous cosmic surveys, picking out supernovae and classifying them. But the system will not just prove handy in sorting existing data.
An Open Eye on the Future
The discovery rate of supernovae is expected to increase substantially in 2023 when the Vera C. Rubin Observatory begins operations. The Rubin Observatory’s wide-field reflecting telescope with an 8.4 meter primary mirror will photograph the entire available sky every few nights. Astronomers expect this to result in the discovery of millions of new supernovae every year. This is a substantial increase upon the 10,000 supernovae currently discovered annually.
Whilst there is no doubt this presents an amazing opportunity for astronomers and astrophysicists alike, limited operating time on the telescope means the need for an effective and speedy classification system is great. The fact that 90 percent of these objects currently identified currently escape identification confirms this.
Should this situation continue, only 0.1 percent of supernovae discovered by the Rubin Observatory each year will get classified with a spectroscopic label, with the remaining 99.9 percent of data left unusable. Clearly, the team’s machine learning classification system is sorely needed.
Also, as things currently stand data sets and classifications have been available to only a limited number of astronomers. Results from the team’s AI classification system will be publicly available. In addition to releasing the Pan-STARRS1 Medium Deep Survey data and their classifications, the team has created an easy-to-use and accessible software platform that can be used by astronomers.
It was really important to us that these projects be useful for the entire supernova community, not just for our group. There are so many projects that can be done with these data that we could never do them all ourselves. These projects are open data for open science.
Edo Berger, Professor of Astronomy, CfA
Clearly, thanks to machine learning and AI the future is bright for supernovae research and classification.
Villar. V. A., Hosseinzadeh. G., Berger. E., et al, , ‘SuperRAENN: A Semi-supervised Supernova Photometric Classification Pipeline Trained on Pan-STARRS1 Medium Deep Survey Supernovae,’ The Astrophysical Journal.
Hosseinzadeh. G., Dauphin. F., Villar. V.A., et al, , ‘Photometric Classification of 2315 Pan-STARRS1 Supernovae with Superphot,’ The Astrophysical Journal