Artificial intelligence has reached a threshold. And physics can help it break new ground

For years, physicists have made great strides and breakthroughs in the field using their minds as their primary tool. But what if artificial intelligence could help with these discoveries?

Last month, researchers at Duke University showed that incorporating known physics into machine learning algorithms could result in new discoveries in material properties. according to a press release by the institution. They conducted a unique project where they constructed a machine learning algorithm to derive the properties of a class of engineering materials known as metamaterials and determine how they interact with electromagnetic fields.

Predicting Metamaterial Properties

The results turned out to be extraordinary. The new algorithm predicted the properties of the metamaterial more accurately than previous methods, while also yielding new insights.

“By incorporating known physics directly into machine learning, the algorithm can find solutions with less training data and in less time,” said Willie Padilla, a professor of electrical and computer engineering at Duke. “While this study was primarily a demonstration showing that the approach could recreate known solutions, it also revealed some insights into the inner workings of nonmetallic metamaterials that no one knew before.”

In their new work, the researchers focused on making discoveries that were accurate and logical.

“Neural networks try to find patterns in the data, but sometimes the patterns they find don’t conform to the laws of physics, making the model they create unreliable,” said Jordan Malof, assistant research professor of electrical and computer engineering at Duke. “By forcing the neural network to obey the laws of physics, we prevented it from finding relationships that might fit the data but aren’t actually true.”

They did that by imposing a physics on the neural network called a Lorentz model. This is a set of equations that describe how the intrinsic properties of a material resonate with an electromagnetic field. However, this was not easy to achieve.

“If you make a neural network more interpretable, which in a way is what we’ve done here, it can be more challenging to refine it,” said Omar Khatib, a postdoctoral researcher who works in Padilla’s lab. “We definitely struggled to optimize the training to learn the patterns.”

A significantly more efficient model

The researchers were pleasantly surprised to find that this model worked more efficient than previous neural networks the group had created for the same tasks by drastically reducing the number of parameters needed for the model to determine the properties of the metamaterial. The new model could even make discoveries on its own.

Now the researchers are getting ready to use their approach in uncharted territory.

“Now that we’ve shown this is possible, we want to apply this approach to systems whose physics is unknown,” Padilla said.

“Many people use neural networks to predict material propertiesbut getting enough training data from simulations is a huge pain,” added Malof. “This work also shows a path to building models that don’t require as much data, which is helpful across the board.”

The study has been published in the news Advanced optical materials.

Leave a Comment

Your email address will not be published.