I wonder. Could it be an idea to calculate with raytraced renders of object instead of just the boarder and nothing else.
I mean. With machine leaning you could just input the raytraced object as data and do the calculation.
Machine learning does not care if the object is more complex than it should be.
If the render looks useful to the human eye than it could very well be useful complexity to the algorithm.
So the idea is to use renders or mesh objects of the objects you normally just use a boarder definition of.
I think this could be used for calculations of physical transfer of information by light between objects which are ?useful for physical networks.