Understanding our Universe with the help of machine learning
Machine learning provides a new way of revealing the physical quantities behind the images observed by telescopes.
Machine learning (ML) is a novel method which uses artificial intelligence (AI) to make predictions with data. Artificial intelligence means using computers to do complex tasks, such as recognizing objects from pictures and playing chess. A lot of applications of ML appeared in many different fields of industry and research in recent years. It not only speeds up the process by efficiently dealing with a great amount of data, but also leads to new methods and new findings.
An international group including experts in astronomy research from the University of Edinburgh, Universidad Autonoma de Madrid (Spain), Sapienza University (Italy) and experts in machine learning models from the EURA NOVA company (Belgium), have started a collaboration to provide new and simple ways to infer important physical quantities of clusters of galaxies from multiwavelength images. Recently, they focused on accurate estimates of the mass of galaxy clusters from the Planck satellite microwave images. This study is published in the latest issue of Nature Astronomy.
Measuring galaxy mass
Galaxy clusters are the most massive object that ever formed in our universe, so precisely measuring their masses has very important meanings for many different astronomy studies. In order to obtain its mass from observed images, astronomers have to first process the image by excluding fore/background objects and removing the noises, then different assumptions have to be made to derive the mass from binned image quantities, such as profiles. These assumptions normally oversimplify the state of the real cluster, therefore the mass derived with such methods doesn’t agree with the true mass. This difference is referred to as bias. Furthermore, the whole process is very time-consuming and the clusters have been handled one by one.
The machine learning method proposed by this group overcomes all these problems and can directly get the cluster mass from observed images. And it is very fast - in just a few seconds, over 1000 observed cluster masses are provided. The machine learning model is based on a type of deep learning algorithm known as the Convolutional Neural Networks (CNN), which can get the most important features of an image to connect with a defined quantity. A perfect tool for this task. However, this involves training the model with over 10 thousand images from numerical simulations of galaxy clusters, of which the group used the results from The THREEHUNDRED project, hosted in Universidad Autónoma de Madrid (UAM). They verified the cluster mass from the machine learning model has no bias and has a very small scatter around the true cluster mass. Applying this model to the images observed by the Planck satellite, they provided bias-free cluster masses of over 1000 galaxy clusters.
PhD student Daniel de Andres from UAM has completed most of the work on this project. The paper’s co-leading author, Dr Weiguang Cui, from the School’s Institute of Astronomy commented:
These results as very exciting, and machine learning is proving to be a useful tool which will help us understand our Universe.