New technologies developed by MIT physicists will one day offer a way to custom design multi-layered nanoparticles with the required properties for use in displays, stealth systems or biomedical devices. It may also help physicists solve a variety of difficult research problems in ways that may be orders of magnitude faster than existing methods.

The innovation uses computational neural networks as a form of artificial intelligence to "learn" how the structure of a nanoparticle affects its behavior, in which case it scatters light of different colors based on thousands of training instances. Then, after understanding this relationship, the program can run essentially backwards to design particles with the desired light scattering properties - a process called inverse design.

The Science Progress Journal reported these findings in a paper by Professor John Peurifoy of MIT, researcher Yishen Shen, graduate student Li Jing, and physics professor Marin Soljacic.

Although this approach may lead to practical applications, Soljacic says the work is primarily a method of predicting the physical properties of various nanoengineered materials, without the need for computationally intensive simulation processes that are often used to solve These issues are mainly related to scientific research.

Soljacic said that the goal is to look at neural networks. This field has made a lot of progress and excitement in recent years. See if we can use some of these technologies to help us conduct physics research. Basically, computers are "smart enough" so that they can Can you do some smarter tasks to help us understand and use some physical systems?"

To test this idea, they used a relatively simple physical system, Shen explained. “To understand which technology is appropriate, and to understand these limitations and how to use them well, we [ use a neural network on a particular nanophotonic system ], nanophotonics is a spherical concentric nanoparticle System .” Nanoparticles are layered like onions, but each layer is made of a different material and has a different thickness.

The size of the nanoparticles is comparable to or smaller than the wavelength of visible light, and the manner in which the different colors of light are scattered from these particles depends on the details of the layers and the wavelength of the incident beam. Calculating all of these effects for multi-layered nanoparticles can be a computationally intensive task for nanoparticles, and the complexity becomes worse as the number of layers increases.

Researchers want to see if neural networks can predict how new particles scatter light, not just by interpolating between known examples, but by actually identifying potential patterns that allow neural network extrapolation.

"The simulations are very accurate, so when you compare them to the experiment, they will reproduce each other little by little," said Peurifoy, who will become a Ph.D. student at the Massachusetts Institute of Technology next year. "But they are quite dense in value, so it takes quite a long time. What we want to see here is if we show the neural network a bunch of examples of these particles, many examples of different particles, whether the neural network is Can develop 'intuition'."

Sure enough, the neural network can reasonably predict the exact pattern of the light scatter plot and the wavelength map, but it is very close and less time consuming. Jing said that neural network simulation "is much faster than simulation now. "So now you can use a neural network instead of real simulation , it will give you a fairly accurate prediction, but it has a price, and the price is us The neural network must be trained first, and then we have to generate a lot of examples. "

However, once the network is trained, any future simulation will get the full benefits of acceleration, so it may be a useful tool for situations where repetitive simulations are needed. But the real goal of the project is to understand the method, not just the specific application. "One of the main reasons we are interested in this particular system is to let us understand these technologies, not just to simulate nanoparticles," Soljacic said.

The next step is to basically run the program in reverse, using a set of required scattering properties as a starting point, and then see if the neural network is able to calculate the exact combination of nanoparticle layers needed to achieve that output.

“In engineering, many different reverse design techniques have been developed and it is a huge area of research,” says Soljacic. "But in order to create a given inverse design problem, it often takes a long time, so in many cases, you have to be an expert in the field, and sometimes even take a few months to set it up to fix it. ”

But with the team's well-trained neural network , "We didn't make any special preparations, we said, 'Okay, let's try to run backwards. Surprisingly, when we compare it to others, This is an effective method when comparing standard reverse design methods," he said. “It's actually much faster than the traditional counter design.”

Co-author Shen said: "What we do is to build a universal toolbox. Any ordinary well-educated person is not a photonics expert. It can be used... This is our original motivation. For this special case. Obviously very effective."

Acceleration in some types of inverse design simulations can be very important. Peurifoy said: "The exact comparison between Apple and Apple is difficult, but you can effectively say that you have hundreds of times the benefits, so the benefits are very impressive - in some cases, it will drop from a few days to a few minutes."

For more information: J. Peurifoy et al., "Nanophotonic particle simulation and inverse design using artificial neural networks," Science Advances (2018). Advances.sciencemag.org/content/4/6/eaar4206  

Read more at: https://phys.org/news/2018-06-ai-based-method-specialized-nanoparticles.html#jCp