Metasurfaces are constructed from millions of sub-wavelength nano-structures whose geometry must be precisely engineered to produce a desired optical effect. A single device may contain millions of individual nano-pillars, and exploring the resulting design space through conventional simulation or human intuition has been a persistent bottleneck for the field. The review, published in Opto-Electronic Advances, examines how AI is removing that constraint across three interconnected areas: inverse design, optical characterization, and fully autonomous end-to-end optical systems.
In the design phase, AI-powered surrogate models can predict how light will interact with a nanostructure in milliseconds, compared with the weeks a conventional simulation might require. More significantly, inverse design turns the process around: rather than proposing a geometry and checking whether it meets the target optical property, an engineer specifies the desired output - a particular focal length or spectral response - and the AI generates the required structure. The approach also allows fabrication tolerances to be embedded as constraints, reducing the gap between simulated performance and manufacturable reality.
The review also addresses how AI is being coupled directly to the optical systems metasurfaces feed. Because metasurfaces produce complex, multidimensional datasets, neural networks are being integrated at the sensor level to extract information that would otherwise be inaccessible. The resulting systems have demonstrated the ability to analyze hyperspectral data for disease detection in blood samples, identify atmospheric gases, and reconstruct high-resolution three-dimensional images for augmented reality displays, all in real time.
A further category reviewed is the end-to-end paradigm, in which the physical hardware and the controlling algorithm are optimized together rather than sequentially. This co-design approach enables cameras that compensate for their own optical aberrations and computational optical systems that perform processing at the speed of light. The authors also discuss programmable metasurfaces - surfaces whose optical behavior can be reconfigured dynamically by an AI controller - with proposed applications including adaptive camouflage and smart antenna arrays for 6G communications that adjust signal paths in real time.
The review addresses the broader resource question facing AI development. Current large-scale AI workloads run on power-intensive server infrastructure. The authors argue that optical computing, in which metasurfaces process data using light rather than electrical signals, could substantially reduce the energy cost of AI inference by performing calculations at the speed of light with minimal power consumption.
The lead authors, Dr. Trevon Badloe and Dr. Sunae So, are both Assistant Professors in the Department of Electronics and Information Engineering and the Department of Electro-Mechanical Systems Engineering, respectively, at Korea University's Sejong Campus. Both completed doctoral work at POSTECH in South Korea and held postdoctoral positions at POSTECH's Graduate School of Artificial Intelligence before joining Korea University.
The review is described by the authors as a roadmap for engineers and scientists working at the intersection of computer science and photonics, with stated application areas spanning non-invasive medical diagnostics, quantum computing, smart city sensor networks, and the Internet of Things.
Research Report: AI-assisted metaphotonics
| The content herein, unless otherwise known to be public domain, is Copyright 1995-2026 Space Media Network. All websites are published in Australia and are solely subject to Australian law and governed by Fair Use principles for news reporting and research purposes. |
