Experts develop new AI technologies to aid research where traditional knowledge and methods fail to work. In a new announcement, astronomers revealed a new technology developed to eliminate noise in astronomical data.
Artificial intelligence continues to solve some of science’s most enduring problems. Experts develop new AI technologies to aid research where traditional knowledge and methods fail to work. In a new announcement, astronomers revealed a new technology developed to eliminate noise in astronomical data.
Once the new tool was tested on actual data from the Subaru telescope, the results were consistent with the standard models of the universe, thus, proving that this new method could serve a great purpose in the future.
Astronomers study the large-scale structure of the universe by observing and measuring gravitational reflection. When studying foreground objects, gravitational lensing distorts the overall image of any background objects. Dark matter is another problem that has been found to distort the shapes of distant objects.
In the past, it has proven difficult to distinguish naturally distorted objects in space from gravitationally distorted objects. Scientists call this shape noise and it has long been one of the major factors that prevent astronomers from studying the universe effectively.
For this new breakthrough study, astronomers generated 25,000 mock galaxy catalogs using the ATERUI II supercomputer. These catalogs were based on data collected by the Subaru Telescope over the years.
The next step was to add noise to these data sets and train the new artificial intelligence technology to isolate the mock data from lensing dark matter.
Here is an image describing the whole process in several sequential illustrations. An explanation of the image can be found below.
In the image above, you see the whole schematic of the work done with the new artificial intelligence tool for isolating noise in astronomical data.
Astronomers have called it an adversarial generative network (GAN). As you can see, they used two networks over the course of the experiments.
Image generator G (Network 1) was given noisy lens maps with a single particular task – to denoise them. Image generator D (Network 2) was tasked to compare the creations of Network 1 with noise-free lens maps in real observational data. The purpose of Network 2 was to identify the artificial intelligence’s maps as fake.
Over the course of the experiment, Network 1 was consequently trained to make more accurate maps while Network 2 was trained to find the fakes among the improved maps. As we mentioned earlier, the astronomers used 25,000 pairs of lens maps (noisy and noise-free).
After the training in artificial intelligence was complete, it was tested on real data from the Subaru telescope. The AI was able to isolate noise and reveal a distribution of foreground mass that was later confirmed as consistent with the standard models of the universe.
In other words, this new tool can be used for analyzing old and future data in studies of the large-scale structure of the Universe. Astronomer Masato Shirasaki noted that the success of this research proves how beneficial it is to combine different scientific methods, both old and new. The study involved observations, simulations, and artificial intelligence in this particular case.
Once again, we see proof that traditional science can be improved and it is time that the scientific world realizes that this is the most efficient way forward if we are to find answers to the mysteries of the universe.
Join the discussion and participate in awesome giveaways in our mobile Telegram group. Join Curiosmos on Telegram Today. t.me/Curiosmos
• Center for Computational Astrophysics. (n.d.). Observation, Simulation, and AI Join Forces to Reveal a Clear Universe.
• ScienceDaily. (2021, July 2). Observation, simulation, and AI join forces to reveal a clear universe.
• Shirasaki, M., Moriwaki, K., Oogi, T., Yoshida, N., Ikeda, S., & Nishimichi, T. (2021, April 9). Noise reduction for weak lensing mass mapping: an application of generative adversarial networks to Subaru Hyper Suprime-Cam first-year data.