Even if you somehow solve the angle issue (say you create a new camera system where all the lenses have the exit pupil at a fixed distance and you manage to design a micro-lens array that properly collimates the light at every pixel location), there are more aspects to consider.
- Luther condition: ideally the sensitivity curves for each pixel colour should be linear combinations of the so called colour matching functions (more or less the colour response of a human observer). This ensures that sources with different spectra but the same colour result in the same RGB values. There is
some room for compromise usually but just cutting the spectrum into three parts is far from ideal.
- crosstalk: in theory, any linear combination will allow you to calculate the correct RGB values. In practice, if there is too much overlap in the spectral response, the inversion becomes very sensitive to small changes. The result is that colour noise is amplified significantly. AFAIK this has been one of the problems with the Foveon design.
- efficiency: with gratings, light is diffracted into different diffraction orders (at different angles) but you can generally only use one. It is possible to concentrate 80%-90% into a single order at the design wavelength with blazed or VPH gratings. But at the edges of the visible spectrum that might well drop below 50%.
- cost, obviously
It would seem using diffraction instead of a bayer mask would be way more efficient and lead to less noise, As well as allowing easier adaption for different wavelength ranges.
In theory yes, its an emerging technology with early demonstrations:
https://doi.org/10.1186/s40580-023-00372-8
"Recent advancements of metalenses for functional imaging"
Functional metalenses are a very interesting concept but as far as I can tell, we are still several scientific breakthroughs away from those being viable in practice.
Looking specifically at the colour router in section 4.1.2, all the results are with laser sources. That's no accident, the design is tailored to plane waves, normal incidence, circular polarisation and three specific wavelengths (and nothing in between). Making broadband metalenses that work across the whole visible spectrum is hard and currently comes at the cost of massively reduced efficiency.
I don't want too sound too negative, I believe there is a lot of potential there. But I think we might see the first real world applications in optical sensing or display technology where we have a lot more control over the light sources than in imaging.
In the medium term, the best chance to potentially improve on Bayer sensors is probably still Foveon. But unless someone invests a lot of money to really make a state-of-the-art version, we might never know.