Using machine learning to increase energy efficiency in nanosystems

Getting something for free doesn’t work in physics. But it turns out that if you think like a strategic gamer and enlist the help of the devil, you might be able to improve the energy efficiency of complex systems like data centers.

Steven Whiterum of the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) used computer simulations to use neural networks, a type of machine learning model that mimics the processes of the human brain. trained the Nanosystem, a small machine in Work more energy efficiently.

Furthermore, simulations showed that the learned protocol could draw heat from the system by constantly measuring the protocol to find the most energy-efficient behavior.

“You can take energy out of the system, or you can store work in the system,” Whiterum said.

This is an insight that can prove valuable when operating very large systems, such as computer data centers. Computer banks generate an enormous amount of heat, and even more energy must be used to extract the heat to prevent damage to sensitive electronics.

Whiterum conducted his research at the Molecular Foundry, a DOE Office of Science user facility at Berkeley Lab. His research is described in a paper published in Physical Review X.

Inspiration from Pac-Man and Maxwell’s Demons

When asked about the origin of the idea, Whiterum said, “People were using techniques found in the machine learning literature to play the Atari video game, which seems to be a natural fit for materials science.” I was.

He explained that in video games like Pac-Man, the goal of machine learning is to choose specific times to perform actions such as up, down, left, or right. Over time, machine learning algorithms “learn” when to do the best moves to achieve a high score. The same algorithm works for nanoscale systems as well.

Mr. Whiterum’s simulation is also something of an answer to an old thought experiment in physics called Maxwell’s demon. Briefly, in 1867 physicist James Clark He Maxwell proposed a box filled with gas. In the center of the box was a massless “devil” controlling a trap door. The devil opens the door, allowing the faster molecules of the gas to move to one side of the box and the slower molecules to move to the other side.

Eventually, when all the molecules are separated in this way, the “slow” side of the box will be cooler and the “faster” side hotter, consistent with the energy of the molecules.

check the fridge

Whiterum said the system constitutes a heat engine. Importantly, however, information is equivalent to energy, so Maxwell’s demons do not violate the laws of thermodynamics, the law of getting something for nothing. Measuring the position and velocity of molecules inside the box takes more energy than is available from the resulting heat engine.

And a heat engine could come in handy. Whiterum said the refrigerator is a good analogy. When the system works, the food inside stays cold even though the back of the refrigerator is hot due to the refrigerator motor. This is the desired result.

In Whiterum’s simulations, machine learning protocols can think like the devil. In the optimization process, the information extracted from the modeled system is converted into energy as heat.

Unleash the Devil with Nanoscale Systems

In one simulation, Whiterum optimized the process of dragging nanoscale beads into water. He modeled a so-called optical trap in which a laser beam that acts like optical tweezers can hold beads and move them around.

“The name of the game is to get from here to there with as little work on the system as possible,” Whiterum said. When water molecules collide, the beads undergo a natural motion known as Brownian motion, causing them to sway. Whiterum showed that if we could measure these fluctuations, we could perform bead movement at the most energy-efficient moments.

“Here we train a neural network demon to show that we can do something similar to Maxwell’s thought experiment with optical traps,” he said.

computer cooling

Whiterum extended this idea to microelectronics and computing. He used machine learning protocols to simulate flipping the state of nanomagnetic bits between 0 and 1, a fundamental information erase/information copy operation in computing.

“We do this over and over. Eventually the devil ‘learns’ how to flip the bits to absorb heat from its surroundings,” he said. He returns to the refrigerator analogy. “You can also create a computer that transfers heat to another location in the data center to cool it while it is running.”

Whiterum said simulations are like testbeds for understanding concepts and ideas. “And the idea here is that these protocols can be run using measurements that are applicable to real experiments, expending little energy, or sucking energy at the cost of going somewhere else. It just shows,” he said.