Quite curious that the backside of the lens is modeled as a mesh of quads with varying (x, y). I could imagine that a fixed grid of points with only varying height would be easier to model, am I missing something crucial?
You can probably build an end-to-end model of a grid of heights (constrained to be h=0 at the edges), a simulated ray exiting the slab (surface normals modified by whatever Snell's law tells you), and the eventual light intensity on the target plane... and immediately optimize the entire thing with backprop?
I'm probably massively oversimplifying this and ignoring half of the physics
You can probably build an end-to-end model of a grid of heights (constrained to be h=0 at the edges), a simulated ray exiting the slab (surface normals modified by whatever Snell's law tells you), and the eventual light intensity on the target plane... and immediately optimize the entire thing with backprop?
I'm probably massively oversimplifying this and ignoring half of the physics