Will Machine Learning Unlock the Cosmos?
Our planet is just one of 400 billion planets that reside in our galaxy with another 150 billion to 250 billion stars and a diameter of 100,000 lightyears. One light year is about 6 trillion miles and is the distance light is able to travel in one earth year. Our galaxy, the Milky Way, is just one of about 100 billion galaxies that our best telescopes can even see, as our technology advances there could be as many as 200 billion galaxies.
Or as Carl Sagan put it, the legendary American Astronomer remarked on our place in the universe: "Who are we? We find that we live on an insignificant planet of a humdrum star lost in a galaxy tucked away in some forgotten corner of a universe in which there are far more galaxies than people."
Our Cosmos seems so large, that it practically never ends. To chart it seems like a fantastically impossible journey, but now with Machine learning, a journey we may be able to take one day.
Machine learning (ML), is especially useful in analyzing the patterns of large data sets and developing algorithms.
In the field of astrophysics, it can be used to describe complex relationships, identify data clusters and outliers, generate simulated results, classify observations and explore data sets to understand the underlying physics. This powerful tool has helped astronomers and physicists make important discoveries.
Estimating cosmological parameters:
This constant arises in Albert Einstein's field equation of general relativity. It is proposed to balance the effect of gravity to reach the static universe. This notion was abandoned by Hubble's discovery of the expanding universe. In the 1990s, studies showed that this value could account for the dark energy in the vacuum that causes the universe to have accelerated expansion.
To study this value, the group from both the Astrophysics and the Computational Science divisions of the Physics Department at Carnegie Mellon University (CMU) used the distribution of dark matter as the indicator.
They first created a vast amount of training sets with different input parameters, meanwhile using different algorithms to calculate the accelerations in this N-body system due to gravity.
Eventually, the ML techniques provided fairly accurate results that are comparable to those from the traditional method. They hope to further work on the robustness of their model before applying it to real data from our universe.
It is widely known that we live in a galaxy as we watch the Milky Way up in the sky. However, structure on a larger scale exists -- galaxy clusters that consist of hundreds to thousands of galaxies are bound together by gravity. This is a challenge to the measurement of their mass.
A group of scientists from Harvard and CMU tried to use Convolutional Neural Network (CNN) to find good approximations of cluster mass. They first made CNNs using a catalogue of mock observations. Then, they used these CNNs on those unseen observations and compared the results with evaluations from traditional methods.
This novel technique has successfully reduced the scatters in estimations for the cluster mass compared to the conventional methods.
A prominent astrophysics group using ML
In New York City, a group of engineers are working on different fields of advanced science through the computational method. They initiated the Flatiron Institute under the Simons Foundation. Prof. David Spergel from Princeton led one of the subdivisions, the Center for Computational Astrophysics. Their aim is to find computational methods for astrophysicists to analyze large data sets from measurements and eventually understand the physical implications.
One of their groups, Cosmology X Data Science, led by Prof. Shirley Ho, seeks to answer fundamental questions of our cosmos by looking into datasets obtained from astronomical surveys. Their recent topics cover the galaxy distribution, galaxy structure formation and universe models using AI and ML. This following video is a demonstration of how they apply ML in their study.
The future applications and prospects of ML for mapping the Cosmos are very exciting. Seeing the influence of ML, many leading physicists have shared their perspectives towards the future application of machine learning on the study of cosmology4. In one of their collective papers, they wrote, “the next decade will bring new opportunities for data-driven cosmological discovery, but will also present new challenges for adopting ML methodologies and understanding the results.”
While enjoying its success, they are also cautious about the potential failure from the black box process of ML. Along with the challenges, they also listed many awesome opportunities in the next decade. For instance, Big Data Opportunities in Radio Astronomy and Archival Data from Hubble intrigue them.
The upcoming missions like The Hydrogen Epoch of Reionization Array (HERA) and eROSITA, the Square Kilometre Array (SKA), will yield large volumes of data, where ML has the potential to facilitate analysis and make progress. As those scientists claim, “ML has the potential to be transformative to our field.”
“Imagination will often carry us to worlds that never were, but without it we go nowhere.”- Carl Sagan
Written by Saiyang Zhang & Edited by Alexander Fleiss