UNIVERSITY PARK, Pa. — A research team at Penn State University has devised a new streamlined approach to designing metasurfaces. Metasurfaces are a type of artificial material that can manipulate light and other forms of electromagnetic radiation just by their structure. This rapid optimization process could be useful in manufacturing advanced optical systems such as camera lenses, virtual reality headsets, and holographic imagers, the researchers said.
The method, featured on the cover of the October issue of Nanophotonics, uses large-scale language models (LLMs) to accurately predict how metasurfaces will affect light. LLM is a type of artificial intelligence (AI) model that can learn and improve its actions over time based on provided training data and repeated actions. This approach bypasses traditional metasurface simulation processes that required extensive disciplinary knowledge and time, allowing engineers to quickly design these nanoscopic materials and predict how they will affect light solely through prompts fed to the AI.
According to electrical engineering professors Doug Warner, John L. McCain, and Genevieve H. McCain, and the study's corresponding author, metasurfaces offer much greater flexibility and functionality than traditional materials in nanophotonic devices, systems that can manipulate light on scales even smaller than the wavelength of visible light.
“When you try to manipulate light or other types of electromagnetic waves, you can only use naturally occurring materials,” Werner explained. “Through the structure of the subwavelength unit cells that make up the material, metasurfaces can manipulate the behavior of light at the nanoscale level, allowing optical systems that are traditionally very bulky to be slimmed down.”
Despite their usefulness, metasurfaces are difficult to develop, said Haunshu Zhang, a third-year electrical engineering doctoral student and lead author of the paper. Zhang said that while AI has been integrated into the development process in recent years in the form of deep learning neural networks that mimic the nonlinear way the human brain connects, researchers still need to go through the time- and knowledge-intensive process of simulating potential designs and building custom neural networks for each metasurface.
This problem inspired him to integrate an LLM into the process.
“The main limitation of current neural network-based methods is that they have to try many neural network configurations to find one that accurately predicts how the metasurface will interact with light,” Zhang said. “By training LLM, we can accurately predict how a metasurface will interact with light in seconds, compared to hours, days, or even months previously, without the need for specialized AI expertise or countless trials.”
The team tested their method by comparing LLM-generated predictions to computer-simulated metasurfaces. LLM predicts how light will react when exposed to a metasurface that has specified “control points” that cause the design to deform into the desired shape. The team then trained and compared these predictions to a dataset of more than 45,000 randomly generated metasurface designs. The researchers found that their approach was able to predict how light would interact with the metasurfaces with great accuracy, while effectively eliminating the time-consuming neural network design process.
The increased efficiency allows researchers to focus on developing what Lei Kang, associate research professor of electrical engineering and co-author of the paper, calls “arbitrarily shaped” metasurface elements. Lei explained that using highly specialized shapes in metasurface design can have a significant impact on performance and efficiency compared to standardized shapes like cylinders and cubes. However, these free-form designs have significant drawbacks.
“Arbitrary designs allow researchers to create application-specific metasurfaces that significantly exceed traditional shape-based designs,” Lei said. “However, we have not been able to effectively optimize and test these designs because traditional simulation methods take unrealistically long times to complete. By integrating LLM predictions, we will be able to see how metasurfaces affect light at unprecedented speeds.”
The new method also makes engineering metasurfaces much more accessible, said Sawyer Campbell, associate professor of electrical engineering and co-author of the paper. LLMs are very good at “reverse design,” meaning starting with a desired outcome and working backwards to find the exact system, material, structure, or combination of factors that will produce it, he said. Inverse design of metasurfaces has been possible before, Campbell said, but the simulation process could take weeks or months to complete.
