Machine learning is used to make complex systems work efficiently the learned patterns can draw heat from the system by constantly comparing them to find the most energy-efficient operation.
Getting things for free doesn’t work in physics. But it turns out that by thinking like a gamer and with the help of the devil, the power is good for complex systems like data centers that can happen.
In a computer demonstration, Stephen Whitelam of the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) used neural networks (a type of machine learning model that mimics the human brain) to train nanosystems, which are small molecular-sized machines, to job and great performance.
In addition, the simulations showed that the learned patterns can draw heat from the system by constantly comparing them to find the most energy-efficient operation.
“We can take energy out of the system or save jobs in the system,” Whitelam said.
This is a concept that can prove useful, for example, in the operation of large systems such as computer data centers. Computer desks produce a lot of heat that must be dissipated – using more energy – to avoid damaging electronic equipment.
Whitelam conducted the research at the Molecular Foundry, a DOE facility staffed by scientists at Berkeley Lab. His work is described in an article published in Physical Review X.
Inspired by Pac Man and Maxwell’s Demon
Asked where his idea came from, Whitelam said, “People have used machine learning systems to play Atari video games that seem to fit into material science.”
In a video game like Pac Man, he explained, the goal of machine learning would be to choose the same time for an action – up, down, left, right, etc. – to do. Over time, machine learning algorithms will learn to “learn” the best moves to perform and when to score high. The same algorithm can work for nanoscale systems.
Whitelam’s demonstration was also a response to an old physics thought experiment called Maxwell’s Demon. In short, in 1867, the physicist James Clerk Maxwell proposed to make a box filled with gas, and in the middle of the box, there would be an infinite “demon” controlling the trap door. The demon will open the door to allow the fast particles of the gas to go to one side of the box and the slow particles to the other side.
Finally, with all the molecules thus separated, the “slow” side of the box will cool and the “fast side” will be hot, corresponding to the strength of the molecule.
Check the refrigerator
The system will work like a generator, Whitelam said. Importantly, Maxwell’s demon does not violate the laws of thermodynamics – gain anything – because information is like energy. Measuring the position and speed of the particles in the box costs more energy than is obtained from the heat engine.
And the generator can be useful. The refrigerator provides a good comparison, Whitelam said. When the system is running, the food inside is cool – what is required – even if the back of the refrigerator is warm due to the work of the refrigerator.
In Whitelam’s presentation, machine learning can be seen as a demon. In the best method, it converts the information learned from the model process into energy in the form of heat.
Leave daemons in nanoscale systems
In simulations, Whitelam improved the process of dragging nanoscale beads into water. He created a device called optical trap that laser light, acting like tweezers of light, can capture and move the pearl.
“The name of the game is: get from here to there with as little system work as possible,” Whitelam said. The pearl oscillates under a natural phenomenon called Brownian motion as it is pushed by the water droplets. Whitelam showed that if these changes could be measured, then the bead could be reshaped at the most powerful moment.
“Here, we show that we can train a neural network daemon to do something similar to Maxwell’s thought experiment but in a blink of an eye,” he said.
Computer freezes
Whitelam extended the concept to microelectronics and computers. It uses machine learning techniques to change the state of the nanomagnetic bit between 0 and 1, which is the basic information erase/print function in the computer.
“Do it over and over again. “Eventually, your demon will ‘learn’ to turn a little to get heat from the environment,” he said. It goes back to the refrigerator analogy. “You can create a computer that cools down while it’s running, and heats it up and sends it somewhere else in your data center.”
Whitelam said the simulations are like a test bed for understanding ideas and concepts. “And here the idea just shows that you can run these lead shots, but with less energy expenditure or with the energy sucked to go elsewhere, using the size that has can put in a real experiment. “, he said. The research was supported by the Department of Energy’s Office of Science.
Source: Berkeley Lab
