A multi-agentic framework for real-time, autonomous freeform metasurface design

TL;DR

MetaChat is introduced, a multi-agentic design framework that can translate semantically described photonic design goals into high-performance, freeform device layouts in an automated, nearly real-time manner, and the design of multiobjective, multiwavelength metasurfaces orders of magnitude faster than conventional methods is demonstrated.

Abstract

Innovation in nanophotonics currently relies on human experts who synergize specialized knowledge in photonics and coding with simulation and optimization algorithms, entailing design cycles that are time-consuming, computationally demanding, and frequently suboptimal. We introduce MetaChat, a multi-agentic design framework that can translate semantically described photonic design goals into high-performance, freeform device layouts in an automated, nearly real-time manner. Multi-step reasoning is enabled by our Agentic Iterative Monologue (AIM) paradigm, which coherently interfaces agents with code-based tools, other specialized agents, and human designers. Design acceleration is facilitated by Feature-wise Linear Modulation-conditioned Maxwell surrogate solvers that support the generalized evaluation of metasurface structures. We use freeform dielectric metasurfaces as a model system and demonstrate with MetaChat the design of multi-objective, multi-wavelength metasurfaces orders of magnitude faster than conventional methods. These concepts present a scientific computing blueprint for utilizing specialist design agents, surrogate solvers, and human interactions to drive multi-physics innovation and discovery.

Figures (8)

• Figure 1: MetaChat agentic framework for autonomous metasurface design. (A) Standard paradigms for creating LLM assistants. (Left) Standard prompting instructs the LLM to assume a certain role for each response, which is generated directly from a given input query. (Right) In chain-of-thought prompting, the LLM is further instructed to explicitly break down the input request into multiple steps before responding, which structures its internal reasoning to help maintain coherence for complex queries. External tool calls can be invoked in a non-iterative, single shot manner. (B) The AIM prompting strategy creates agents that refine their intermediate responses through self-dialogue, autonomously leveraging feedback from external tools, agents, and the human designer at multiple stages to optimize the final answer. (C) Overview of the MetaChat framework. AIM Design (ii) and Materials Expert (iii) Agents iteratively interface with high speed metasurface optimization and design algorithms, Python computing packages, and the human user to process semantic user inputs (i) into outputted metasurface design files (iv) in nearly real time.
• Figure 2: FiLM WaveY-Net fullwave surrogate solver, with support for variable sources and structures. (A) FiLM WaveY-Net is composed of a UNet backbone that uses learned affine transformations on the decoding blocks to assist with conditioning on variable source angles and wavelengths. The UNet input channels include the dielectric distribution and the real and imaginary components of the source. The network output is the magnetic near-field profile, from which electric near-field profiles and far-field scattering profiles via the Stratton-Chu formalism are computed. (B) The input to FiLM WaveY-Net in the context of the simulation domain. The red dashed line indicates the location of the PML boundary condition. The black dashed line indicates the center section of the domain that is inputted into the CNN. The solid red line indicates the location of forward sources, and the solid blue line indicates the location of backward sources. (C) The ranges of the variable quantities in the FiLM-Conditioned WaveY-Net training set. (D) Overview of the differentiable neuroparameterization framework for defining superpixel geometry. The input spatial coordinates $x$ are mapped into a higher-dimensional space using rotary positional encoding. The encoded coordinates are then passed through a multilayer perceptron (MLP), which learns an implicit signed distance function (SDF) representation of the design. (E,F) Two examples of FiLM WaveY-Net surrogate simulation results (left) and complex percentage error between the predicted and ground truth magnetic field (right).
• Figure 3: (continued) (G) Test loss as a function of the variable source parameters: incidence angle (blue) and wavelength (red). The test loss remains stable across the entire angle range but decreases with increasing wavelength due to the reduced complexity of the field profile at lower frequencies. (H) Violin plot distributions of test loss across the full range of structure-related input parameters: grayscale temperature (blue), maximum permittivity (green), height (orange), length (purple), and number of ridges (red). Each parameter range is divided into six bins for direct comparison. The bottom quartile, median, and top quartile are indicated with black horizontal lines. The test loss is consistent across all binned parameter ranges, with at least 90% of the test samples exhibiting an MAE below 0.10 in each distribution. (I) Loss curves for the ablation study of the FiLM-Conditioned WaveY-Net. The "Baseline" model inputs the structure only; the "Source-only" model inputs the source along with the structure; the "Source+FiLM" model uses the FiLM conditioning technique in conjunction with the source inputted via separate CNN channels alongside the structure to achieve a low loss that generalizes to the test set.
• Figure 4: Automated multi-objective metalens design via MetaChat. (A) Design strategy of metasurfaces using superpixel arrays. The FoM captures the electric field magnitudes at the desired focal point locations, which is maximized by the constructive interference of complex scattering fields from all superpixels. (B) Expected minimum FoM with respect to sample size based on Monte Carlo sampling from 1,000 optimization runs. We observe a logarithmic improvement in FoM with increased superpixel batch size, and that batches of 60 structures balances performance with compute time. The inset depicts the best superpixel far-field scattering profile (target amplitude: red line) from a batch size of 60, showing minimal side lobes. (C) Salient speech excerpts from the design conversation between the User, Design Agent, and Materials Expert Agent for the design of a dual-wavelength, multi-functional metasurface. The Design Agent asks for clarification before engaging in self-thought and autonomous conversation with the Materials Expert Agent to gather the information needed to design a custom dual-objective metalens spanning 360 wavelengths. (D) The final metalens design, which comprises 100 superpixels that each are $\sim\! 3.5$ wavelengths wide.
• Figure 5: (continued) (E) The spectral color mapped power profile of the optimized metalens, plotted using the MetaChat visualization tool. The horizontal dashed line indicates the focal plane and the vertical dashed lines indicate the red and blue focal spots. (F) Normalized intensity in the metalens focal plane as a function of lateral position, featuring a $480\, \text{nm}$ peak at $20\,\upmu\text{m}$ and a $680\, \text{nm}$ peak at $0\,\upmu \text{m}$. (G) Execution time profiling of the design process. Design Agent iterations are colored in dark blue and tool calls are in light blue. The API call compute is in green and plotting visualization is in purple. This entire MetaChat design process takes 11 minutes, with the 300,000 FiLM WaveY-Net simulations accounting for 10 minutes. (H) The 60 parallel optimization trajectories for one of the superpixels comprising the metalens. The start of minimum feature size enforcement is indicated via the vertical dashed line and is accompanied by a temporary bounce in the FoM. The trajectory that leads to the best-performing structure that is ultimately included in the metalens design is the red line. The mean FoM is the blue line.