Data-Based Design Method for Metamaterials Uses Artificial Intelligence

The process for designing metamaterials could be improved by using data-driven approaches based on deep learning, according to researchers at Pohang University of Science and Technology (POSTECH). The researchers used a deep-learning-assisted inverse design method to enable structural parameters and material for metamaterials to be designed simultaneously and with a greater degree of freedom.

The group led by professor Junsuk Rho taught an artificial neural network to recognize the correlation between the extinction spectra of electric and magnetic dipoles and core-shell nanoparticle designs that included material information and shell thicknesses. The AI system allowed greater freedom of design by categorizing types of materials and adding material types as a design factor, making it possible to design materials to meet specific optical requirements. An analysis of the metamaterials obtained through this design method showed that they exhibited optical properties that were identical to the properties that were input into the artificial neural network.

This is a schematic diagram of an artificial neural network that can design arbitrary photonic structures. Cross-section of structures is mapped as a two-dimensional cross-sectional bitmap so that an artificial neural network can design structures of metasurface antennas that cannot be designed with structural parameters. Courtesy of POSTECH.

The researchers demonstrated deep-learning-assisted inverse design of core-shell nanoparticles for spectral tuning electric dipole resonances, for finding spectrally isolated pure magnetic dipole resonances, and for finding spectrally overlapped electric dipole and magnetic dipole resonances.

The use of an artificial neural network significantly reduced the time needed to design photonic structures, the researchers said. Also, it allowed various designs of new metamaterials because the researchers were no longer limited by the need to develop designs intuitively, based on observations and trial-and-error.

These are schematics of an artificial neural network that can design structural parameters and material simultaneously. When desired optical properties (electric/magnetic dipole spectrum) are input, each level of thickness and type of materials of the three-layer core-shell nanoparticle is provided as output. Courtesy of POSTECH.

The team’s work could facilitate the practical utilization of deep-learning technology for nanophotonic inverse design. Metamaterials can be used in applications such as display, security, and military technology, but only within the limits of the currently available development methods. The introduction of AI to the design method could contribute to the technological development of metamaterials.

Regarding his team’s approach to metamaterials design, professor Rho said, “Our research was successful in bringing it to a higher degree of freedom of the design, but the new design still requires users to input certain problem settings in the beginning. It sometimes produced wrong designs and therefore made it impossible to produce desired metamaterials.

“So, I’d like to take our findings a step further by developing a complete design method of metamaterials utilizing AI. Also, I’d like to make innovative and practical metamaterials by training AI with reviews of the design constructed in consideration of final products.”

AI can learn designs of various metamaterials and the correlation between photonic structures and their optical properties. Using this training process, AI can provide a design approach to efficiently make a photonic structure with desired optical properties. Once trained, it can provide a desired design promptly and efficiently. AI-based approaches to metamaterials design have already been researched, but according to the POSTECH team, the previous studies have required material and structural parameters to be input beforehand and for photonic structures to be adjusted after the design is completed.

The research was published in ACS Applied Materials & Interfaces (https://pubs.acs.org/doi/10.1021/acsami.9b05857).