Computers have the ability to simulate stellar explosions, outsmart chess champions and predict global climate. Researchers are even training them to be error-free problem-solvers and quick learners.
At present, physicists from the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) and their colleagues have shown that computers can now solve the greatest mysteries of the universe. In order to train computer networks to recognize significant features, the researchers fed thousands of images from simulated high-energy particle collisions.
The team programmed powerful arrays called neural networks to function similar to a hivelike digital brain in investigating and elucidating the images of the simulated particle debris remaining after the collisions. At the time of the test run, the team discovered that the success rate of the neural networks in identifying significant features in a sampling of around 18,000 images was nearly 95%.
The research has been reported in the Nature Communications journal on January 15, 2018.
The next move would be to implement the same machine learning procedure to more realistic experimental data.
Robust machine learning algorithms enable these networks to enhance their investigation when they process a higher number of images. The underlying technique is adopted in facial recognition and other kinds of image-based object recognition applications.
The images used by the team in this research, pertinent to particle-collider nuclear physics experiments performed at Brookhaven National Laboratory’s Relativistic Heavy Ion Collider and CERN’s Large Hadron Collider, simulate the conditions of a subatomic particle “soup,” a superhot fluid state called the quark-gluon plasma, considered to have existed just millionths of a second subsequent to the genesis of the universe. Physicists from the Berkeley Lab take part in experiments at both the sites.
“We are trying to learn about the most important properties of the quark-gluon plasma,” stated Xin-Nian Wang, a nuclear physicist in the Nuclear Science Division at Berkeley Lab, who is one of the members of the team. A few of the characteristics are so short-lived and exist at very tiny scales that they are surrounded by mystery.
During the experiments, nuclear physicists make use of particle colliders for making heavy nuclei, such as lead or gold atoms stripped off electrons, to collide with each other. Such collisions are considered to release particles inside the nuclei of the atoms, resulting in a fleeting, subatomic-scale fireball that disintegrates even neutrons and protons into a free-floating form of their typically linked building blocks—gluons and quarks.
Scientists believe that by understanding the exact conditions in which the quark-gluon plasma is formed, for example, the amount of energy packed in, as well as its pressure and temperature as it transforms into a fluid state, they can acquire innovative knowledge of its constituent particles of matter and their characteristics, and regarding the formative stages of the universe.
However, precise evaluations of these characteristics, or the well-known “equation of state” involved when matter transitions from one phase to the other during these collisions, have been highly difficult. As the result of the experiments might be impacted by the initial conditions, it is difficult to extract equation-of-state evaluations that do not rely on these conditions.
In the nuclear physics community, the holy grail is to see phase transitions in these high-energy interactions, and then determine the equation of state from the experimental data. This is the most important property of the quark-gluon plasma we have yet to learn from experiments.
Xin-Nian Wang, Nuclear Physicist - Nuclear Science Division, Berkeley Lab
Scientists also wish to understand the basic forces controlling the interactions between gluons and quarks, referred by physicists as quantum chromodynamics.
Long-Gang Pang, a Berkeley Lab-affiliated postdoctoral researcher at UC Berkeley and the lead author of the research, stated that in 2016 when he was a postdoctoral fellow at the Frankfurt Institute for Advanced Studies, he was intrigued by the ability of artificial intelligence (AI) to assist in solving difficult problems in science.
He noticed that deep convolutional neural network—one form of AI with architecture similar to the image-handling processes in animal brains—seemed to be optimal for investigating science-related images.
These networks can recognize patterns and evaluate board positions and selected movements in the game of Go. We thought, ‘If we have some visual scientific data, maybe we can get an abstract concept or valuable physical information from this’.
Long-Gang Pang, Lead Author
Wang further stated that “With this type of machine learning, we are trying to identify a certain pattern or correlation of patterns that is a unique signature of the equation of state.” Therefore, subsequent to training, the network will be able to spot the portions of as well as the correlations in an image most correlated to the problem researchers are attempting to solve, on its own.
According to Pang, gathering data required for the investigation can be computationally highly rigorous, and in specific instances, it nearly required one whole day of computing time to develop just a single image. By using an array of GPUs that operate in parallel, where a GPU is a graphics processing unit first developed to improve video game effects and have since been adopted for a range of applications, they required time was reduced to nearly 20 minutes per image.
Computing resources at Berkeley Lab’s National Energy Research Scientific Computing Center (NERSC) were used for the study, with the majority of the computing work concentrating on GPU clusters at GSI in Germany and Central China Normal University in China.
The scientists said that an advantage of adopting sophisticated neural networks is that they can recognize features that were not even attempted in the first experiment, such as searching for a needle in a haystack when one was not even looking for it. Moreover, they can derive useful data even from fuzzy images.
“Even if you have low resolution, you can still get some important information,” stated Pang.
Talks are already on to use the machine learning tools to analyze data extracted from realistic heavy-ion collision experiments, and the simulated outcomes must be helpful in training neural networks to interpret the real-time data.
“There will be many applications for this in high-energy particle physics,” stated Wang, apart from particle-collider experiments.
Kai Zhou, Nan Su, Hannah Petersen, and Horst Stocker, respectively from Frankfurt Institute for Advanced Studies, Goethe University, GSI Helmholtzzentrum für Schwerionenforschung (GSI), and Central China Normal University, were the other researchers who were part of this study. The U.S Department of Energy’s Office of Science, the National Science Foundation, the Helmholtz Association, GSI, SAMSON AG, Goethe University, the National Natural Science Foundation of China, the Major State Basic Research Development Program in China, and the Helmholtz International Center for the Facility for Antiproton and Ion Research supported this study.
NERSC is a user facility of the DOE Office of Science.