Reading time ( words)
“The initial predictions won’t look anything like the actual correct answer,” said Malof. “But like a human, the network can gradually learn to make correct predictions by simply observing the commercial simulator. The network adjusts its weights and biases each time it makes a mistake and does this repeatedly until it produces the correct answer every time.”
To maximize the accuracy of the machine learning algorithm, the researchers trained it with 18,000 individual simulations of the metamaterial’s geometry. While this may sound like a large number, it actually represents just 0.0022 percent of all the possible configurations. After training, the neural network can produce highly accurate predictions in just a fraction of a second.
Even with this success in hand, however, it still only solved the forward problem of producing the frequency response of a given geometry, which they could already do. To solve the inverse problem of matching a geometry to a given frequency response, the researchers returned to brute strength.
Because the machine learning algorithm is nearly a million times faster than the modeling software used to train it, the researchers simply let it solve every single one of the 815.7 million possible permutations. The machine learning algorithm did it in only 23 hours rather than thousands of years.
After that, a search algorithm could match any given desired frequency response to the library of possibilities created by the neural network.
“We’re not necessarily experts on that, but Google does it every day,” said Padilla. “A simple search tree algorithm can go through 40 million graphs per second.”
The researchers then tested their new system to make sure it worked. Nadell hand drew several frequency response graphs and asked the algorithm to pick the metamaterial setup that would best produce each one. He then ran the answers produced through the commercial simulation software to see if they matched up well.
The researchers chose arbitrary frequency responses for their machine learning system to find metamaterials to create (circles). The resulting solutions (blue) fit well with both the desired frequency responses and those simulated by commercial software (grey).
With the ability to design dielectric metamaterials in this way, Padilla and Nadell are working to engineer a new type of thermophotovoltaic device, which creates electricity from heat sources. Such devices work much like solar panels, except they absorb specific frequencies of infrared light instead of visible light.
Current technologies radiate infrared light in a much wider frequency range than can be absorbed by the infrared solar cell, which wastes energy. A carefully engineered metamaterial tuned to that specific frequency, however, can emit infrared light in a much narrower band.
“Metal-based metamaterials are much easier to tune to these frequencies, but when metal heats up to the temperatures required in these types of devices, they tend to melt,” said Padilla. “You need a dielectric metamaterial that can withstand the heat. And now that we have the machine learning piece, it looks like this is indeed achievable.”