Skip to content Skip to main navigation Report an accessibility issue
Steven Johnston and student

Machine Learning Can Speed Up Calculations

Machine learning algorithms allow a computer to learn and make predictions based on existing data. Machine learning is already in use in the everyday lives of many. Netflix, for example, uses machine learning algorithms to make recommendations based on the viewing habits of its approximately 125 million subscribers. Machine learning is becoming more widespread in a variety of arenas, including health care and finance. Steven Johnston, assistant professor of physics and astronomy, thinks it can also be applied to physics.

Johnston works with quantum Monte Carlo simulations. These simulations work by taking a configuration or set of parameters and proposing a random change over and over again in the search for the best arrangement.

“The difficulty is that deciding whether or not you accept that proposed change is incredibly expensive computationally,” he said. “We need to find a way to do this cheaper and faster.”

Johnston’s JDRD team is working with machine learning to train a computer using a neural network to guess whether a proposed change is going to be accepted. His team has run some benchmark tests that show the machine learning algorithms capable of completing calculations at a significantly faster rate than other methods.

“The neural network approach can do in about six hours what used to take five or six days. The idea is now that we can do it faster, we can make the problem bigger and use the same computing time as before,” said Johnston.

He hopes to have completely benchmarked an algorithm by the end of the funding year, determining how well it performs against a conventional algorithm. He also plans to carry out at least one comparative study to confirm that both conventional and machine learning algorithms produce the same results.

Johnston’s ORNL collaborator, Markus Eisenbach of the Center for Computational Sciences, is using the same machine learning techniques to tackle a different set of algorithms. The work of both teams could contribute meaningfully to the future use of machine learning in physics.