Harnessing the power of generative AI, researchers at Tsinghua University have developed AIGP—a diffusion-based generative framework that enables instant translation of optical properties into fabrication-ready metasurfaces. By using transmission, phase, and polarization as “prompts,” AIGP directly maps optical properties to subwavelength, fabricable structures, generating high-fidelity metasurface designs in seconds. This breakthrough overcomes critical bottlenecks in photonic inverse design.
How It Works: From Prompts to Structures
The research team developed a novel encoding scheme for optical properties and a dedicated prompt encoder network to facilitate direct mapping from optical specifications to subwavelength photonic structures. This system leverages a latent diffusion model, interpreting optical requirements as 'prompts' to generate desired photonic structures with high precision, eliminating the need for iterative optimization. A fast forward prediction network enhances simulation speed and supports seamless end-to-end training. A comprehensive training dataset, including freeform shapes, was constructed to maximize the design space while adhering to fabrication constraints, ensuring that only manufacturable geometries are considered. The AIGP framework offers three key advantages: high-precision mapping of full-band transmission spectra, phase profiles, and polarization responses into fabrication-ready metasurface structures within seconds; flexible design constraints, such as enabling polarization-insensitive device generation and band-specific masking; and a 'fuzzy search' capability, allowing the AI to approximate ideal performance even with abstract requirements like a single cutoff wavelength without relying on precise forward models.
Experimental Validation: From Simulation to Chip
The power of the AIGP framework was experimentally validated on a silicon-on-sapphire platform. In one demonstration, sixty-four structural-color meta-atoms were directly generated and fabricated on a 230 nm silicon layer, successfully encoding a sunflower image onto a chip. This confirmed the system's readiness for a 'generate-and-fabricate' workflow. For challenging design targets, such as an ideal long-pass filter response that is theoretically unattainable, AIGP quickly produced near-optimal solutions, with measured transmission spectra closely matching the design targets. The method's ability to generalize was further proven across various applications, including bandpass filters, polarization beam splitters, and multi-wavelength phase modulators, showcasing its versatility and robustness in diverse photonic design scenarios.
A New Paradigm for Photonic Innovation
The Artificial Intelligence-Generated Photonic (AIGP) framework represents a significant advancement by moving beyond conventional methods that rely on iterative optimization. It uniquely addresses and overcomes critical challenges historically inherent in photonic inverse design, such as non-uniqueness, robustness to unseen inputs, and the computational inefficiency of iterative procedures. By completely eliminating these persistent obstacles, AIGP introduces a transformative paradigm for AI-driven generative photonic design. The technology demonstrates end-to-end reliability, enabling one-shot mapping from simulation to physical device fabrication without iteration. This breakthrough is poised to significantly accelerate the development of next-generation photonic devices and applications, including optical computing, metalenses, hyperspectral imaging chips, structural colors, and beam splitters, thereby ushering in a new era of large-scale, AI-driven innovation in photonics.
