In one of the most advanced efforts of its kind, researchers used machine learning algorithms to determine the characteristics of metamaterials, an engineered material class. They also predicted how they would interact with electromagnetic fields.

Since the algorithm first had to consider the metamaterial’s physical limitations, it was forced to demonstrate its abilities. This method allowed the algorithm to predict metamaterial properties with high accuracy. It also performed faster and provided additional insight than previous approaches.

This silicon metamaterial, which has rows of cylinders that extend into the distance, can manipulate the light depending on its features. Recent research has shown that machine learning algorithms can be improved by incorporating known physics.

Professor of electrical and computer engineering at Duke Willie Padilla said, “By incorporating existing physics directly into machine learning, the algorithm is able to find solutions with fewer training data and in a shorter time.” While this study demonstrated that the approach could recreate known solutions, it also provided insight into the inner workings of non-metallic metamaterials.

Metamaterials are synthetic materials that combine many engineered features to produce properties not found anywhere else in the world. The metamaterial is a silicon cylinder grid resembling a Lego baseplate.

The metamaterial interacts with electromagnetic radiation in various ways depending on how the cylinders are spaced. It can absorb, emit, or deflect specific wavelengths. The researchers created a neural network, a machine learning model that can be used to study how the different heights and widths within a single-cylinder influence these interactions. They also wanted the answers to make sense.

Jordan Malouf, an assistant professor of electrical and computer engineering at Duke, stated that neural networks attempt to find patterns in data but that sometimes the patterns they find don’t follow the laws of Physics, making the model it creates unpredictable. We made it impossible for the neural network not to find relationships that might fit the data but aren’t true by forcing it to follow the laws of Physics.”

The Lorentz model, a set of equations describing how an intrinsic property of material interacts with an electromagnetic field, was the physics that the research group imposed on the neural network. Instead of jumping to predicting the response of a cylinder, the model had first to learn how to predict the Lorentz parameters, which are used to calculate its response.

However, it is easier to include this extra step than it is to do.

Omar Khatib is a postdoctoral researcher in Padilla’s lab. He said, “when you make a neural net more interpretable, which was in some sense what they did here, it can be harder to fine-tune.” “We had difficulty optimizing the training to learn these patterns.

However, once the model worked, it was more efficient than the previous neural networks that the group had built for similar tasks. This approach, in particular, can drastically reduce the number of parameters required for the model to determine metamaterial properties.

They also discovered that the physics-based approach of artificial intelligence could make discoveries on its own.

An electromagnetic wave travelling through an object doesn’t always interact with it in the same way at both the beginning and end of its journey. This phenomenon is called spatial dispersion. The researchers had to adjust the spatial dispersion parameters to make the model work correctly. This gave them insights into the process’s physics that they didn’t know before.

“Now that we have shown that it is possible, we want this approach to systems in which the physics remains unknown.”