Deep Learning Approach Streamlines the Design of Metamaterials

A technique for streamlining the process of designing and characterizing nanophotonic metamaterials, based on deep learning, could make the design, fabrication, and characterization of these elements easier. A Tel Aviv University (TAU) team has shown that a deep learning network can be trained to design nanophotonic metamaterial elements efficiently.

Researchers from TAU show that a deep learning network can be trained to design nanophotonic metamaterial elements efficiently.

Nanophotonic design has remained challenging because it is largely an iterative process in which the designer guesses a structure and solves the Maxwell’s equations for it. The inverse problem, that is, obtaining a geometry for a desired electromagnetic response, also continues to be a challenging task.

“The process of designing metamaterials consists of carving nanoscale elements with a precise electromagnetic response,” said researcher Michael Mrejen. “But because of the complexity of the physics involved, the design, fabrication, and characterization processes of these elements require a huge amount of trial and error, dramatically limiting their applications.”

To streamline this process, the researchers fed a deep neural network (DNN) with 15,000 artificial experiments to teach the network the complex relationship between the shapes of nanoelements and their electromagnetic responses. The DNN, once trained, was able to retrieve subwavelength dimensions from far-field measurements and obtain a geometry for a desired electromagnetic response.

The researchers’ approach to training a bidirectional network went from the optical response spectrum to the nanoparticle geometry and back, making it effective for both the design and characterization tasks. The DL-based approach was able to predict the spectral response of nanostructures with high accuracy. It could also address the inverse problem and provide a single nanostructure’s design, geometry, and dimension, to achieve a targeted optical response for both polarizations.

“Once a shape is fabricated, it usually takes expensive equipment and time to determine the precise shape that has actually been fabricated,” said researcher Haim Suchowski. “Our computer-based solution does that in a split second based on a simple transmission measurement.”

The TAU team’s approach could provide a way for direct on-demand engineering of plasmonic structures and metasurfaces for applications in sensing, targeted therapy, and more. The predictive capability of the DL model holds promise for multivariate characterization of nanostructures beyond the diffraction limit.

“For the first time, a novel Deep Neural Network, trained with thousands of synthetic experiments, was not only able to determine the dimensions of nanosized objects but was also capable of allowing the rapid design and characterization of metasurface-based optical elements for targeted chemicals and biomolecules,” Suchowski said. “These results are broadly applicable to so many fields, including spectroscopy and targeted therapy, i.e., the efficient and quick design of nanoparticles capable of targeting malicious proteins.”

The researchers have written a patent on their new method and are currently expanding their deep learning algorithms to include the chemical characterization of nanoparticles.

The research was published in Light: Science and Applications (https://doi.org/10.1038/s41377-018-0060-7).