Machine Learning Finds New Metamaterial Designs for Energy Harvesting

Estimated reading time: 4 minutes

Electrical engineers at Duke University have harnessed the power of machine learning to design dielectric (non-metal) metamaterials that absorb and emit specific frequencies of terahertz radiation. The design technique changed what could have been more than 2000 years of calculation into 23 hours, clearing the way for the design of new, sustainable types of thermal energy harvesters and lighting.

Metamaterials are synthetic materials composed of many individual engineered features, which together produce properties not found in nature through their structure rather than their chemistry. In this case, the terahertz metamaterial is built up from a two-by-two grid of silicon cylinders resembling a short, square Lego.

Adjusting the height, radius and spacing of each of the four cylinders changes the frequencies of light the metamaterial interacts with.

Calculating these interactions for an identical set of cylinders is a straightforward process that can be done by commercial software. But working out the inverse problem of which geometries will produce a desired set of properties is a much more difficult proposition.

Because each cylinder creates an electromagnetic field that extends beyond its physical boundaries, they interact with one another in an unpredictable, nonlinear way.

“If you try to build a desired response by combining the properties of each individual cylinder, you’re going to get a forest of peaks that is not simply a sum of their parts,” said Willie Padilla, professor of electrical and computer engineering at Duke. “It’s a huge geometrical parameter space and you’re completely blind—there’s no indication of which way to go.”

When the frequency responses of dielectric metamaterial setups consisting of four small cylinders (blue) and four large cylinders (orange) are combined into a setup consisting of three small cylinders and one large cylinder (red), the resulting response looks nothing like a straightforward combination of the original two. 

One way to find the correct combination would be to simulate every possible geometry and choose the best result. But even for a simple dielectric metamaterial where each of the four cylinders can have only 13 different radii and heights, there are 815.7 million possible geometries. Even on the best computers available to the researchers, it would take more than 2,000 years to simulate them all.

To speed up the process, Padilla and his graduate student Christian Nadell turned to machine learning expert Jordan Malof, assistant research professor of electrical and computer engineering at Duke, and Ph.D. student Bohao Huang.

Malof and Huang created a type of machine learning model called a neural network that can effectively perform simulations orders of magnitude faster than the original simulation software. The network takes 24 inputs—the height, radius and radius-to-height ratio of each cylinder—assigns random weights and biases throughout its calculations, and spits out a prediction of what the metamaterial’s frequency response spectrum will look like.

First, however, the neural network must be “trained” to make accurate predictions.

Page 1 of 2

• Next Page >