Summary: A new artificial intelligence technique can dramatically speed up complex physics simulations and enable the tailored design of multilayered nanoparticles, researchers report.
Source: MIT
MIT physicists have developed an AI-driven method that could enable the customized design of multilayered nanoparticles with specific optical responses—useful for applications ranging from displays and cloaking to biomedical sensors. The approach also promises to accelerate certain physics simulations by many orders of magnitude compared with conventional numerical methods.
The team’s method trains artificial neural networks to learn how a nanoparticle’s layered structure determines its light-scattering behavior. After exposure to thousands of examples, the network internalizes the relationship between structure and optical response. The trained model can then be used in reverse—specifying desired scattering characteristics and allowing the network to propose layer compositions and thicknesses that achieve them—an approach known as inverse design.
The work appears in the journal Science Advances and was authored by John Peurifoy, Yichen Shen, Li Jing, Marin Soljačić, and colleagues.
Although the technique may eventually enable practical devices, Soljačić emphasizes that its immediate value lies in providing fast, reliable predictions of physical properties for a wide range of nanoengineered materials—without relying on the computationally intensive simulations normally required.
“We wanted to explore whether modern neural-network techniques can be useful in our physics research,” Soljačić explains. “Specifically, can these algorithms perform intelligent tasks that help us understand and control complex physical systems?”
To evaluate the approach, the researchers focused on a well-defined nanophotonics problem: spherically concentric, multilayer nanoparticles—like onions where each layer is a different material with its own thickness. These particles are on the scale of visible wavelengths or smaller, and their scattering of different light colors depends sensitively on the properties and arrangement of the layers. As the number of layers increases, exact calculations become substantially more demanding.
The key question was whether a neural network could not only interpolate between examples it had seen, but also infer deeper patterns enabling accurate extrapolation to new particle configurations. “The exact simulations match experiments point by point, but they’re computationally heavy,” says Peurifoy. “We wanted to know if a network exposed to many examples could develop a sort of ‘intuition’ about how structure maps to optical response.”
In practice, the trained neural network produced scattering-versus-wavelength curves that closely matched the full numerical simulations, though not perfectly. Crucially, the AI predictions were generated far faster than the exact models. “Once trained, the neural network runs orders of magnitude quicker than the full simulations,” Jing notes. “That makes it useful in contexts where repeated or rapid predictions are needed.”
Training does require a large initial set of simulated examples, which represents the main upfront cost. But after that investment, every subsequent evaluation benefits from the speedup, making the approach particularly valuable for iterative design workflows and parametric studies.
The researchers then used the trained model to perform inverse design: feeding the network a desired scattering profile and asking it to infer the layer structure that would produce it. In many inverse-design problems, setting up and running standard optimization methods can demand specialized expertise and substantial effort. “Often you have to be an expert and spend weeks or months preparing an inverse-design problem,” Soljačić says.
By contrast, the team’s neural-network toolbox required no bespoke preparation for this problem. When run backward, the model produced solutions that compared favorably with more traditional inverse-design techniques and, in many cases, delivered results far more quickly. Shen describes the goal as building a general, user-friendly design toolbox that a broadly educated researcher—someone who is not a photonics specialist—could use effectively, and the results suggest that aim is achievable for this class of problems.
The speed advantages can be dramatic. While exact comparisons vary with problem specifics, Peurifoy estimates that the network can provide speedups on the order of hundreds of times in some cases—reducing calculations that formerly took days down to minutes.
Beyond this specific study, the researchers view their results as a proof of concept showing that neural networks can serve as efficient emulators for challenging physical simulations and as practical engines for inverse design. The approach could be adapted to other nanophotonic systems and to a broader set of physics problems where repeated, rapid evaluations are required.
Funding: Supported by the National Science Foundation, the Semiconductor Research Corporation, and the U.S. Army Research Office through the Institute for Soldier Nanotechnologies.
Contributors: John Peurifoy, Yichen Shen, Li Jing, Yi Yang, Fidel Cano-Renteria, Brendan G. DeLacy, John D. Joannopoulos, Max Tegmark, and Marin Soljačić.
Source: David L. Chandler, MIT
Publisher: NeuroscienceNews.com
Image credit: MIT
Original research: “Nanophotonic particle simulation and inverse design using artificial neural networks,” published in Science Advances on June 1, 2018. DOI: 10.1126/sciadv.aar4206
Nanophotonic particle simulation and inverse design using artificial neural networks
We propose using artificial neural networks to approximate light scattering by multilayer nanoparticles. The network requires only a relatively small sample of training data to approximate the simulation with high precision. Once trained, the neural network can simulate these optical processes orders of magnitude faster than conventional methods. Additionally, the trained network can solve inverse-design problems in nanophotonics via backpropagation, producing analytical gradients rather than relying on numerical approximations.