Source-linked AI summary
VortexChat: An agentic framework for autonomous multi-objective integrated photonic design
Faqian Chong, Yulun Wu, Shilong Li, Andrew Forbes, Hongsheng Chen, Song Han
TL;DR
Integrated-photonic inverse design remains constrained by expert supervision and limited end-to-end automation, motivating autonomous workflows that reduce manual intervention. VortexChat combines an LLM decision agent with topology generation, gradient-based refinement, and full-wave simulation in a closed loop. It demonstrates autonomous multi-objective design through Vortex100 performance constraints and an experimentally confirmed broadband terahertz perfect-vortex multiplexer.
Problem
Integrated-photonic design workflows remain heavily dependent on expert intervention, while generated structures can lack clear physical interpretability and manufacturability.
Method
VortexChat uses an LLM agent to coordinate topology generation, gradient-based refinement, and full-wave simulation in an end-to-end closed loop.
Results
VortexChat demonstrates multi-objective photonic design through multiplexers evaluated under simultaneous efficiency, mode purity, and bandwidth constraints, with experimentally confirmed device operation.
Takeaways & Limitations
The framework supports autonomous inverse design of complex integrated photonic devices while maintaining physical fidelity and fabrication feasibility.
Takeaways & Limitations
Existing inverse-design approaches can produce structures lacking clear physical interpretability and manufacturability, and some physical optimization settings yield structures that cannot be directly fabricated using standard processes.
Abstract
from arXiv · showhide
The advancement of modern integrated photonics is frequently bottlenecked by device design workflows that rely heavily on manual simulation and expert intuition. While inverse design offers an alternative, it remains constrained by expert supervision and a lack of end-to-end automation. To address these issues, we present VortexChat, an agentic framework for the autonomous, end-to-end inverse design of integrated photonic devices directly from natural language specifications. VortexChat couples a large language model (LLM) decision agent with topology generation, gradient-based refinement, and full-wave electromagnetic simulation. This closed-loop architecture enables the system to iteratively decompose design objectives, orchestrate computational tools, and update strategies based on feedback with minimal human intervention. Constrained by the absolute metrics of the Vortex100 Benchmark, VortexChat autonomously generates devices that strictly meet all predefined performance thresholds without any human-in-the-loop. As an experimental demonstration, we fabricated a broadband terahertz perfect vortex beam multiplexer, autonomously designed by VortexChat, with measurements confirming high-efficiency operation, high mode purity and low inter-channel crosstalk in agreement with full-wave simulations. These results demonstrate that an LLM agent can assume key aspects of expert decision-making in photonic inverse design while maintaining physical fidelity and fabrication feasibility, providing a scalable route towards autonomous design of complex integrated photonic systems.
Introduction
Integrated-photonic inverse design remains dependent on expert intervention and tool-centered workflows, while existing automated approaches face limitations in efficiency, interpretability, manufacturability, or complexity. VortexChat addresses this gap through an agent-driven closed loop that coordinates topology generation, gradient refinement, and full-wave simulation for autonomous multi-objective design.
- Motivation: Existing inverse-design approaches trade off global search, optimization efficiency, data dependence, physical interpretability, and manufacturability.Heuristic methods require many simulations, gradient methods depend on initialization and constraints, and deep networks depend on training data while often producing structures lacking clear physical interpretability and manufacturability.
- Motivation: Inverse-design workflows still require experts to select tools, configure optimization, monitor convergence, and correct fabrication violations.This dependence persists even when the desired optical response is known, because practical device structures require substantial expert intervention.
- Related approaches: LLM-based photonic approaches have supported prediction and inverse design, but are often data-intensive, computationally costly, and limited for complex free-form structures.Prior efforts include spectral prediction for metasurfaces and inverse design of thin-film structures.
- VortexChat framework: VortexChat integrates an LLM agent with external solvers, simulation tools, and optimization modules to create a closed-loop autonomous design workflow.The framework interprets natural-language requirements, coordinates tool execution, evaluates full-wave metrics, and adjusts optimization strategies with minimal human intervention.
- VortexChat framework: The framework dynamically orchestrates OAMDiff for physically plausible topology initialization, GBR for fabrication-compatible binary structures, and StructureTester for full-wave feedback.These tools support rapid initialization, adjoint-based refinement, and autonomous simulation feedback.
Results
VortexChat integrates agentic decision-making with topology generation, physics-based refinement and full-wave validation for autonomous photonic design. It achieves strong benchmark performance and experimentally validated broadband vortex multiplexing.
- Framework: VortexChat dynamically integrates objectives, structural parameters, simulation feedback and optimization history to select tools, update parameters and generate fabrication-oriented layouts.Its closed-loop logic separates decision-making from task-specific physical tools and can transfer across photonic design settings.
- Physical refinement: Gradient-based Refinement converts continuous initial topologies into silicon-air binary structures while balancing optical performance and fabrication constraints.The optimization trajectory showed sharp figure-of-merit fluctuations, reflecting dynamic trade-offs during refinement.
- Benchmark evaluation: VortexChat achieved a 71% overall Vortex100 success rate under simultaneous efficiency, purity and bandwidth criteria.Individual pass rates were 78% for efficiency, 72% for purity and 90% for bandwidth, indicating that joint multi-objective optimization was the main bottleneck.
- Experimental demonstration: Measurements confirmed mode purity above 84% throughout the operating band and inter-channel crosstalk of 5.2%, agreeing with the target vortex modes.Measured efficiencies for the devices remained within −5.0 to −4.6 dB across the operating band.
Conclusions and discussion
VortexChat provides an end-to-end, agentic framework for autonomous multi-objective photonic design, coordinating specialized tools to translate natural-language specifications into fabrication-compatible structures. Systematic evaluation and fabrication demonstrate broadband vortex multiplexers satisfying simultaneous performance constraints while reducing dependence on continuous expert supervision.
- Framework and workflow: VortexChat coordinates three specialized tools for automated integrated-photonic device design.The framework is presented as an autonomous agentic system for end-to-end design and tool orchestration.
- Framework and workflow: Natural-language specifications are translated into executable design objectives, enabling autonomous objective trade-off and design correction.The agent operates through a closed-loop workflow that revises design decisions during optimization and feedback.
- Evaluation and demonstration: The framework was systematically evaluated on complex optical-field manipulation tasks, including broadband terahertz perfect vortex beam multiplexers.The multiplexers were designed under simultaneous constraints on efficiency, mode purity, and bandwidth.
- Evaluation and demonstration: Fabrication and experimental characterization confirmed that the generated structures are practically realizable and reproduce the intended optical-field responses.The experimental results support agreement between fabricated-device behavior and the intended design responses.
- Implications: The results reduce dependence on continuous expert supervision while maintaining physical fidelity and fabrication feasibility.The reported outcome positions LLM-based tool orchestration as a scalable foundation for autonomous complex integrated-photonic system development.
- Implications: Natural-language interaction can provide non-expert users access to high-performance computing resources for complex integrated-photonic design.The framework can be integrated with multimodal language models through APIs and deployed on cloud servers.
- Limitations and outlook: Further validation across broader device classes and experimental settings remains necessary.The discussion identifies more efficient, specialized, and physically reliable tools as important for expanding autonomous-design capability boundaries.
Methods
VortexChat combines conditional topology generation, electromagnetic simulation, and benchmark evaluation into an automated photonic inverse-design workflow. Its data and model pipeline encode target performance conditions, generate candidate structures, refine them, and assess efficiency, purity, bandwidth, fabrication constraints, and computational cost.
- Training data: Training samples pair 121 × 121 binary silicon–air permittivity topologies with efficiency, mode-purity, and operating-wavelength conditions for perfect vortex beams.Simulation-derived efficiency and mode purity serve as training labels.
- Simulation: Generated structures use 200 μm thickness, 5000 μm width, and 5000 μm length, then undergo full-wave cross-validation in Ansys Lumerical FDTD.Perfectly matched layers, mode-port excitation, and a field monitor three wavelengths above the sample surface support output-field evaluation.
- Topology generation: OAMDiff denoises latent topology representations conditioned on target performance vectors and diffusion timesteps to generate designs.The performance vector is embedded through a multilayer perceptron and injected into the network backbone alongside timestep embeddings.
- Conditional architecture: Cross-attention injects global design objectives into spatial features while convolutional pathways preserve local geometric and fabrication-relevant details.Attention is concentrated in intermediate and low-resolution U-Net stages, whereas high-resolution layers rely mainly on convolution and timestep embeddings.
- Benchmark: Evaluation records task success, port efficiency, mode purity, bandwidth robustness, tool-calling rounds, runtime, and computational-resource consumption.GPU hours, CPU hours, four-task-parallel wall-clock runtime, and estimated FLOPs are measured on a specified hardware platform.
- Performance metrics: Simulation efficiency is obtained by normalizing monitor-plane output power to input power, while mode purity is evaluated by Fourier analysis of the angular field distribution.The monitor-plane power integral uses the time-averaged Poynting vector and the surface normal.
Conflict of Interest
VortexChat is presented as an autonomous framework that combines language-driven decision-making with photonic topology generation, refinement, simulation, and feedback. The framework is evaluated through benchmark tasks and a fabricated terahertz vortex-beam multiplexer.
- Framework: VortexChat converts natural-language design requirements into structured objectives and orchestrates generation, evaluation, and refinement in a closed loop.The decision engine selects tools, adjusts parameters, and regenerates or further optimizes designs when constraints are not satisfied.
- Framework: OAMDiff generates conditional continuous electromagnetic topologies, which are refined by gradient-based optimization and evaluated through full-wave simulation.OAMDiff uses latent diffusion conditioned on design requirements, while GBR refines the initial topology using an adjoint-gradient workflow.
- Evaluation: The Vortex100 evaluation measures overall, efficiency, purity, and bandwidth success rates across 100 autonomous design tasks.The reported evaluation also groups final structures by success or failure and records tool-calling rounds and computational resource consumption.
- Experimental demonstration: VortexChat autonomously designs a broadband terahertz perfect vortex beam multiplexer whose simulated response is examined across the target wavelength band, mode purity, and crosstalk.The device is evaluated within the 750–900 μm target range, with field distributions, broadband response, mode crosstalk, and OAM purity reported.
- Experimental demonstration: Experimental characterization reports measured vortex-beam field distributions, energy conversion efficiency, OAM mode purity, and inter-channel crosstalk.The measurements concern perfect vortex beams with different topological charges and span different wavelength ranges for purity and crosstalk.