How we taught AI to recognize clusters of galaxies
Recently, together with a team of astrophysicist friends, I completed a project whose goal was to search for distant galaxies and their clusters hidden by the fabric of outer space. Now I will share with you what we have done as a result of this difficult work.
Galaxies and their clusters are large-scale objects of the visible part of the Universe. The results of the studies devoted to them provide valuable information for expanding the field of knowledge about various large-scale structures and allow us to reveal the peculiarities of the formation of the modern form of the Universe. I will tell you more about this in the following articles (if you are interested).
To analyze a huge amount of information from telescopes, at least for the presence of galaxies, an automatic mechanism (or more astronomers) is required. You can write a program that performs this task. But how to teach her to distinguish galaxies and their clusters from other objects of space?
We were lucky, there was a place in space for “magic,” specifically for the Sunyaev-Zeldovich effect, discovered in the last century.
The effect is as follows: initially relic radiation photons are not energetic, like a sloth on a eucalyptus branch, but after interacting with electrons with a large amount of energy inside a gas, their energy increases due to the temperature of the gas in the cluster, which is heated under adiabatic compression or under the action of forces gravity, or in the collision of galaxies and clouds of intergalactic matter.
Fig. 1. The effect of Sunyaev - Zeldovich.
By increasing the energy, the photon increases its frequency and goes from millimeter to submillimeter range. At this moment, in the direction of galaxy clusters of relic radiation photons with a given temperature in the millimeter range is not enough, therefore, in the direction of the cluster of galaxies there is a dip in relation to the average background. And in the submillimeter range, on the contrary, an excess of photons and a local peak.
This manifests itself as follows: the effect of a cosmic microwave background (i.e., uniform thermal radiation filling the Universe, hereinafter CMB), observed along the cluster line of galaxies, looks weaker at low frequencies and brighter at higher frequencies.
Thus, under the influence of the effect, the background is converted into a negative signal for frequencies below the threshold (Fig. 2, image on the left) and a positive signal for frequencies above the threshold with no signal at zero frequency 217 GHz (Fig. 2, the image on the right). This feature of the effect allows astronomers to find clusters of galaxies and superclusters in the microwave region of the spectrum.
What is not magic?
Fig. 2. Influence of the Sunyaev-Zeldovich effect on the visible properties of clusters of galaxies
Experimental evidence for the existence of the effect was obtained quite recently, when astrophysicists conducted a study of the electromagnetic spectrum on the Planck telescope and noticed that at some frequencies the observed sky region appears to be “empty”, and others on it appear whole clusters of galaxies.
Fig. 3. This is the first supercluster discovered using the Sunyaev-Zeldovich effect. At the left - the image received by "Plank". The right panel shows the image obtained using the observatory "XMM-Newton."
It's all great, but what have we done?
You know, there are often situations when you make a decision to do something simply because you like it, although you assume that it will not be needed in the future. It was the same situation.
When the text for the main part of the work was written and there was quite a bit of time for processing the results, and there was just under a week before the deadline, I sat in front of the monitor and did not know what to do. I sometimes even like such situations, because only they have to solve the problem on the optimal strategy. I understood that I couldn’t physically recognize a large amount of data (about 10,000 images), and I only had three courses I took, one of which I was just rescued. The course is dedicated to working with Inception, a convolutional neural network of Google, which I once passed “for self-development” (link at the end of the article).
Anaconda 2 software, Python 2.7 programming language, Keras library for working with machine learning and big data, and Theano for working with numerical data were used to work with the neural network.
Of course, without the advice of people who are engaged in machine learning for two years, it wasn’t. Therefore, after four days we had a program for working with neural networks of deep learning.
A network consists of sequences of convolutional layers (CL) and layers of a union (PL). Convolutional layers allow you to extract several feature maps from the input images, and the union layers perform the specified subsampling on function maps.
These sequences of layers correspond to the stage of feature extraction. To classify images, the output level is a fully bound layer with a number of units equal to the number of classes. The network is built according to the basic architecture with two stages of convolution (a special type of integral transformation) and subsample connected to the classifier, which is shown in the figure.
Fig. 4. Neural network architecture
The network was trained with a teacher. Photo catalogs for network learning and further recognizing clusters of galaxies are compiled using the GLESP, a pixelization scheme for cosmic microwave background maps, which creates a strict orthogonal decomposition of the display. To create a neural network training catalog, data from the Planck telescope mission was used, the purpose of which was to search for galaxies and their clusters using the Sunyaev-Zeldovich effect. The data from the mission is presented in the form of 6,135 images taken at frequencies of 100, 143, 217, 353 and 545 GHz.
Some of the results of the network are presented in Figure 5. We obtained two coefficients (0.35 and 0.87). And if the coefficient is greater than 0.5, then there is a cluster of galaxies in the image.
And, lo and behold, we found a flock!
Fig. 5. Network performance
The program has been applied to a catalog of images of different parts of the sky and is currently analyzing them for the presence of galaxies and their clusters.
In the perspective of the project, we will study in more detail the principle of the effect of the Sunyaev-Zeldovich effect on the visible properties of large-scale objects of the Universe and create a universal analytical algorithm for a more detailed study of space objects.
I really hope that this small article took you to the wonderful world of space for a minute. See you in the next articles!
- Course Inception
- O. V. Verkhodanov, N. V. Verkhodanova, O. S. Ulakhovich et al., Astrophysical Bulletin, vol 73, 1, 2018
- Ostriker, Jeremiah P., Ethan T., Nature, 322 (6082): 804, 1986
- Passmoor S., Cress C., MNRAS, 397 (1), 2009
- Planck Collaboration, Astron. Astrophys.571, A29, 2014