Tsinghua team uses diffusion AI to design photonic structures in seconds

By AI, Created 12:43 PM UTC, May 16, 2026, /AGP/ – Researchers at Tsinghua University have built AIGP, a diffusion-based framework that turns optical requirements into fabrication-ready metasurfaces without iterative optimization. The work, published in Light: Advanced Manufacturing, could speed design for devices such as metalenses, filters and beam splitters.

Why it matters: - AIGP cuts a major bottleneck in photonic inverse design by replacing slow, simulation-heavy optimization with one-shot generation. - The method could speed development of AI-designed optical devices for computing, imaging, structural color and light control. - The framework maps optical targets directly to manufacturable subwavelength structures, which matters because non-unique design spaces have made photonic design difficult to automate.

What happened: - A research team led by Professor Kaiyu Cui at Tsinghua University developed AIGP, a latent diffusion model for inverse photonic design. - The work appeared in Light: Advanced Manufacturing. - AIGP uses transmission, phase and polarization specifications as prompts to generate metasurface structures. - The system produces high-fidelity designs in seconds and removes the need for iterative optimization.

The details: - The team built a new encoding scheme for optical properties and a prompt encoder network to address the non-uniqueness problem in photonic design. - A fast forward prediction network speeds simulation and supports end-to-end training. - The training dataset includes freeform shapes and filters out geometries that cannot be manufactured. - The researchers say AIGP can convert full-band transmission spectra, phase profiles and polarization responses into metasurface structures ready for fabrication. - The framework supports polarization-insensitive design through C4 symmetry. - The system also supports band-specific masking to adapt to different design goals. - The researchers describe a fuzzy search mode that can approximate performance from partial specifications, such as a single cutoff wavelength. - The source text says AIGP can work without precise forward models for those abstract requirements.

Between the lines: - The main shift is from optimization to generation, which changes photonic design from an engineering search problem into a prompt-driven AI workflow. - That matters because iterative inverse design has been limited by convergence issues, computational cost and the challenge of finding global optima. - The approach also suggests a broader move toward generative tools that can handle fabrication constraints during design, not after the fact.

What’s next: - The team says the framework could accelerate development of optical computing devices, metalenses, hyperspectral imaging chips, structural colors and beam splitters. - Further work will likely focus on broader device classes, more complex targets and continued experimental validation across fabrication platforms. - The paper identifies funding from China’s national research programs and quantum information institutions, which suggests continued support for follow-on work.

The bottom line: - AIGP shows that generative AI can design physically manufacturable photonic structures directly from optical goals, with a path from simulation to chip in seconds.