A new technique for custom design of multilayered nanoparticles could bring desired properties to displays, cloaking systems and biomedical devices. It may also help physicists tackle a variety of research problems, in ways that would be orders of magnitude faster than existing methods.

Developed by MIT physicists, the approach relies on computational neural networks, a form of artificial intelligence, to "learn" how a nanoparticle's structure affects its behavior. After learning the relationship through thousands of training examples, the program can essentially be run backward to design a particle with desired properties -- a process known as inverse design.

According to physics professor Marin Soljacic, the new method can predict the physical properties of a variety of nanoengineered materials -- without requiring computationally intensive simulation processes. The goal, he says, was to look at neural networks to determine “whether we can use some of those techniques in order to help us in our physics research. So basically, are computers 'intelligent' enough so that they can do some more intelligent tasks in helping us understand and work with some physical systems?"

To test the idea, the researchers used a relatively simple physical system – the neural network on one particular system for nanophotonics. The system is comprised of spherically concentric nanoparticles, with sizes comparable to the wavelengths of visible light or smaller. The particles are layered like an onion, with each layer made of a different material and possessing a different thickness. As a result, light of different colors scatters off of these particles in different ways, depending on the details of these layers and on the wavelength of the incoming beam.

Calculating all these effects for nanoparticles with many layers can be an intensive computational task, and the complexity gets worse as the number of layers grows. The researchers wanted to see if the neural network could predict the way a new particle would scatter colors of light -- not just by interpolating between known examples, but by actually figuring out some underlying pattern.

“What we want to see here is, if we show a bunch of examples of these particles – many, many different particles -- to a neural network, whether the neural network can develop 'intuition' for it," explained John Peurifoy, an MIT senior who will be a doctoral student next year.

The research proved successful: The neural network was able to predict reasonably well the exact pattern of a graph of light-scattering versus wavelength -- not perfectly, but very close, and in much less time. The catch is that a large number of examples still need to be developed in order for the neural network to be trained. Once trained, however, future simulations get the full benefit of the speedup. Running the program in reverse also worked much more quickly than traditional inverse design approaches, which often require quite a bit of time and expertise to set up.

The researchers, whose work appears in the journal Science Advances, pointed out that their overarching goal was achieving an understanding of the methodology, as opposed to success in only this particular application. "The initial motivation we had to do this was to set up a general toolbox,” said study coauthor Yichen Shen.

The speedup for certain kinds of inverse design simulations can be quite significant, according to Peurifoy. "It's difficult to have apples-to-apples exact comparisons,” he said, “but you can effectively say that you have gains on the order of hundreds of times. So the gain is very, very substantial -- in some cases it goes from days down to minutes."