Engineers have developed an artificial-intelligence-designed surface that could make digital cameras considerably more efficient at gathering light, a problem that has long limited image quality in dim conditions. The approach replaces the colour filters used in most sensors, which absorb the majority of incoming light and waste much of the energy a camera receives.
The proposed alternative is a metasurface made of silicon-nitride nanopillars arranged to steer light toward the correct pixels rather than blocking unwanted wavelengths. The difficulty lies in finding the right arrangement. The number of possible nanopillar patterns is enormous, and conventional simulations are too slow to test millions of candidate designs directly. A team led by Park and collaborators addressed this by training a neural network on 140,000 simulated structures. The network was paired with a genetic algorithm that refined the most promising designs. The trained surrogate ran about 230,000 times faster than the original simulations, allowing the search to cover roughly 40 million candidate designs.
In an experimental device, the optimised surface transmitted 88.8 per cent of incident light. A comparable conventional colour-filter array transmitted 24.7 per cent in the team's reconstruction, an improvement factor of about 3.6. The researchers acknowledge that the design still shows residual colour crosstalk and a slight loss of resolution, and these limitations will need to be addressed before the technology can be used commercially.
The trained network can also be reused for different detector geometries without retraining, which could speed development of future designs. The researchers point to possible applications in smartphones, night vision and security cameras, as well as endoscopes and microscopes, where capturing more light from weak signals can matter most.
Credit. Based on reporting by Paul Mabey for Physics World. Is this your work? Write to us.
100
