A team of researchers at Penn State have devised a new, streamlined approach to design metasurfaces, a class of engineered materials that can manipulate light and other forms of electromagnetic radiation with just their structures. This rapid optimization process could help manufacture advanced optical systems like camera lenses, virtual reality headsets, holographic imagers and more, the team said.
The method, which was featured in Nanophotonics, uses large language models (LLMs) to accurately predict how a metasurface will influence light. LLMs are a type of artificial intelligence (AI) model capable of learning and improving an action over time based on provided training data and repeated behavior. This approach bypasses the traditional metasurface simulation process that required extensive domain knowledge and time, making it possible for engineers to quickly design these nanoscopic materials and predict how they will influence light solely through prompts fed to AI.
According to Doug Werner, the John L. and Genevieve H. McCain Chair Professor of Electrical Engineering and corresponding author on the work, metasurfaces offer much more flexibility and capability than traditional materials in nanophotonic devices — systems that can manipulate light at a scale even smaller than a wavelength of visible light.
“You can only go so far using naturally occurring materials when trying to manipulate light or other types of electromagnetic waves,” Werner explains. “Through the structure of the subwavelength unit cells that make up the materials, metasurfaces can manipulate the way light behaves at a nanoscopic level, allowing us to slim down optical systems that are traditionally very bulky.”
Despite their usefulness, metasurfaces are challenging to develop, according to Haunshu Zhang, a third-year electrical engineering doctoral student and first author on the paper. Zhang says that although AI has been integrated into the development process for a few years in the form of deep-learning neural networks, which mimic the nonlinear way human brains can make connections, researchers would still have to go through the time-consuming and knowledge-intensive process of simulating potential designs and constructing a custom neural network for each metasurface. This problem inspired him to integrate LLMs into the process.
“The main limitation of current neural-network-based methods is that you must try many neural network configurations in order to find one that accurately predicts how a metasurface will interact with light,” Zhang says. “By training LLMs, we can accurately predict how a metasurface will interact with light in seconds compared to the hours, days or even months it previously took, without needing specialized AI expertise or countless trials.”
The team compared their LLM-generated predictions to computer-simulated metasurfaces to test their method. The LLMs would predict how light would react when exposed to a metasurface with designated control points that morphed the design into a desired shape. The team then trained and compared these predictions to a data set of over 45,000 randomly generated metasurface designs. The team found that their approach provided highly accurate predictions of how light would react with the metasurfaces, 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, called arbitrarily shaped metasurface elements. Lei explains how, compared to standardized shapes like cylinders or cubes, using highly specialized shapes in metasurface design can significantly impact performance and efficiency — but these free-form designs come with a substantial drawback.
“Arbitrary designs allow researchers to create application-specific metasurfaces that vastly outperform designs based on traditional shapes,” Lei says. “However, these designs couldn’t be optimized and tested effectively because traditional simulation methods would take an impractically long time to complete. By integrating LLM predictions, we can see how the metasurfaces will influence light at unprecedented speeds.”
The new method also makes engineering metasurfaces extremely approachable, according to Sawyer Campbell, associate professor of electrical engineering and co-author on the paper. The LLMs are very good at “inverse design,” or starting with the desired outcome and working backward to find the exact system, material, structure or combination of factors that produces it, he says. While inverse design of metasurfaces was possible before, the simulation process meant it could sometimes take multiple weeks or months to complete, according to Campbell.
Looking ahead, the team plans to continue developing and optimizing this new approach. According to Werner, the primary goal is to significantly reduce the design time and complexity for metasurface-enabled devices, accelerating their development and integration into commercial nanophotonic applications across the healthcare, defense, energy, and consumer electronics industries.
“We believe this approach could usher in a new standard for how industry engineers and researchers approach developing nanophotonic devices,” Werner says. “With this new method, researchers unfamiliar with the complex metasurface design process can approach the LLMs with an explanation of what they need and effectively generate it.”
This research was supported by the John L. and Genevieve H. McCain Endowed Chair Professorship at Penn State.
This article was written by Ty Tkacik, Penn State. For more information, contact Douglas Werner at
Overview
The paper "Chat to chip: large language model based design of arbitrarily shaped metasurfaces" by Zhang et al. presents a novel workflow leveraging large language models (LLMs) for rapid design and spectral prediction of arbitrarily shaped metasurfaces. Traditional metasurface design relies heavily on full-wave electromagnetic simulations (e.g., FDTD), which are computationally expensive and time-consuming, particularly for complex or aperiodic geometries. While deep neural networks (DNNs) have accelerated design by predicting optical responses quickly, they require task-specific architectures and extensive hyperparameter tuning for each new optical function, limiting accessibility.
This work demonstrates that LLMs, pre-trained on large text corpora and fine-tuned with relatively small photonics datasets, can effectively replace bespoke DNNs for metasurface design without architectural modifications. Arbitrarily shaped meta-atoms are parameterized via 4×4 control-point grids, converted into token sequences input to LLMs. The authors fine-tune the Meta-Llama-3.1-8B model using low-rank adaptation (LoRA) to map between geometric tokens and transmission spectra sampled over 31 wavelengths. This approach recasts forward optical prediction as a natural language sequence completion problem, enabling rapid spectral inference consistent with FDTD simulation results.
They further exploit LLM stochasticity for inverse design, addressing the ill-posed many-to-one mapping challenge by prompting the model with target spectra to generate candidate meta-atom control grids. Benchmarking across 11 open-weight LLMs reveals that mid-sized models (~7–9B parameters) achieve good accuracy-to-cost trade-offs, with architecture and tokenizer design often more impactful than mere parameter count. Reasoning-focused LLMs proved less effective for strict numerical prediction due to conversational detours.
Overall, the study advances a "chat-to-chip" paradigm linking natural language interaction with nanophotonic device modeling. It opens a low-barrier, flexible route to accelerate metasurface design, bypassing cumbersome DNN engineering. The work establishes LLMs as practical tools for both forward and inverse design of complex, freeform metasurfaces, promising widespread applicability and facilitating rapid prototyping with modest consumer-grade GPUs. Supplementary materials and detailed benchmarks provide valuable baselines for future LLM-driven photonics research and applications.
