Machine Learning Makes Scientific Research 42,000 Times Faster
Scientists can learn about advanced materials in minutes rather than years.
A research team at Sandia National Laboratories has successfully used machine learning—computer algorithms that improve themselves by learning patterns in data—to complete cumbersome materials science calculations more than 40,000 times faster than normal, according to a Sandia press release.
Their results, published in the January 4 issue of a journal called npj Computational Materials, could herald a dramatic acceleration in the creation of new technologies for optics, aerospace, energy storage and potentially medicine while simultaneously saving laboratories money on computing costs, according to the press release.
The research, funded by the U.S. Department of Energy’s Basic Energy Sciences program, was conducted at the Center for Integrated Nanotechnologies, a Department of Energy user research facility jointly operated by Sandia and Los Alamos national labs.
Sandia researchers used machine learning to accelerate a computer simulation that predicts how changing a design or fabrication process, such as tweaking the amounts of metals in an alloy, will affect a material. A project might require thousands of simulations, which can take weeks, months or even years to run.
The team clocked a single, unaided simulation on a high-performance computing cluster with 128 processing cores—more than 120 more processing cores than the average home computer—at 12 minutes. With machine learning, the same simulation took 60 milliseconds using only 36 cores, equivalent to 42,000 times faster on equal computers. This means researchers can now learn in under 15 minutes what would normally take a year.
Sandia’s new algorithm arrived at an answer that was 5 percent different from the standard simulation’s result, a very accurate prediction for the team’s purposes. Machine learning trades some accuracy for speed because it makes approximations to shortcut calculations.
The team will use the algorithm first to research ultrathin optical technologies for next-generation monitors and screens. Their research, though, could prove widely useful because the simulation they accelerated describes a common event: the change, or evolution, of a material’s microscopic building blocks over time.
Machine learning previously has been used to shortcut simulations that calculate how interactions between atoms and molecules change over time. The published results, however, demonstrate the first use of machine learning to accelerate simulations of materials at relatively large, microscopic scales, which the Sandia team expects will be of greater practical value to scientists and engineers.
For instance, scientists can now quickly simulate how miniscule droplets of melted metal will glob together when they cool and solidify, or conversely, how a mixture will separate into layers of its constituent parts when it melts. Many other natural phenomena, including the formation of proteins, follow similar patterns. And while the Sandia team has not tested the machine-learning algorithm on simulations of proteins, they are interested in exploring the possibility in the future.