Researchers at Harvard and the Max Planck Institute demonstrated an AI-driven inverse-design approach that makes photonic chip components 500 times smaller than conventional equivalents.

Testing showed insertion losses around 2 dB at operating wavelengths, crosstalk below -10 dB for multiplexers, and mirrors achieving reflectivity up to 98.5%, enabling optical cavities for multiple light bounces.

This advance could allow billions of optical connections on a single chip, reducing heat generation and electrical consumption.

Potential applications span data centers, advanced AI models and supercomputers, and ultra-fast optical routing in telecommunications, with wide-reaching impacts for devices from smartphones to internet infrastructure.

Researchers produced three families of silicon nitride components—wavelength multiplexers, spatial mode multiplexers, and reflectors—with footprint reductions of up to 500x for some spatial mode multiplexers (roughly 8 × 8 square micrometers).

The AI-optimized material distribution yielded nanostructures unlike human-designed shapes, yet manufacturable and effective at controlling light at very small scales.