AI-designed metasurface could make cameras far more light-efficient

Modern cameras lose a significant amount of incoming light due to color filters, which reject most of the spectrum. The Bayer filter, widely used for decades, relies on absorption to create color images, limiting the efficiency with which cameras can utilize available light. A promising alternative involves using a metasurface—a patterned layer of silicon-nitride nanopillars—to steer red, green, and blue light towards appropriate detector pixels, rather than blocking unwanted colors.

Designing such a pattern, however, is a complex computational challenge due to the vast number of possible arrangements.

Researchers Hyoseok Park and collaborators sought to tackle this problem by training a neural network on 140,000 simulated structures and integrating it with a genetic algorithm. This approach allowed them to evaluate optical fields around 230,000 times faster than original simulations. By leveraging the trained network, they could search through 40 million candidates, ultimately designing a metasurface that transmitted 88.8% of incident light, compared to just 24.7% for a conventional color-filter array used in the researchers' reconstruction—a remarkable 3.6-fold improvement.

While some residual color crosstalk and spatial resolution loss remain, this neural network-based approach can be reused to design devices for various detector geometries without retraining. The potential applications for this light-efficient technology are vast, including smartphone cameras, low-light and night-vision imaging, security cameras, medical endoscopes, and microscopes.

The full study, titled "Metasurface color routers inverse-designed with a deep-learning-accelerated genetic algorithm for high-efficiency imaging," was published in Rep. Prog. Phys. 89 097901.