Modern cameras throw away much of the light that enters them. Although the electronic sensor records the intensity of incoming light, it cannot determine its colour. To create a colour image, cameras therefore cover individual detector pixels with red, green or blue filters, each of which rejects most of the spectrum. The familiar Bayer filter has enabled decades of digital photography, but its reliance on absorption places a fundamental constraint on how efficiently a camera can use the available light.

One possible alternative is to replace these absorptive filters with a metasurface: a patterned layer of silicon-nitride nanopillars. Rather than blocking unwanted colours, such a device could steer red, green and blue light towards the appropriate detector pixels. Designing a pattern capable of controlling all three colours is, however, an enormous computational challenge.

The main bottleneck is finding the right arrangement of nanopillars among a huge number of possible patterns. Accurate electromagnetic simulations are also too computationally expensive to test millions of candidates directly.

This is the problem that Hyoseok Park and collaborators sought to address. The team trained a neural network on 140,000 simulated structures, then combined it with a genetic algorithm that progressively selected and modified the most promising designs.

The trained network evaluated optical fields around 230,000 times faster than the original simulations, enabling a search through 40 million candidates.

The resulting device transmitted 88.8% of incident light experimentally, compared with 24.7% for a conventional colour-filter array used in the researchers’ reconstruction – a factor of 3.6 improvement. Some residual colour crosstalk and a modest loss of spatial resolution do remain, so this isn’t a completely solved problem, just yet.

Importantly, because the neural network predicts the complete optical-field distribution, it can be reused to design devices for different detector geometries without being retrained.

This technology could be useful wherever compact cameras need to capture as much light as possible. Likely applications include smartphone cameras, low-light and night-vision imaging, security cameras, medical endoscopes, and microscopes.