|
|
|
“Diffractive Optical Elements for Space-Variant Interconnections in Three-Dimensional Computation Structures”
by Ching-Chu Huang
August 1997
Optical architectures for fully connected and limited-fanout space-variant weighted interconnections based on diffractive optical elements for fixed-connection multilayer neural networks are investigated and compared in terms of propagation lengths, system volumes, connection densities, and interconnection crosstalk. The architectures are designed to be physically compact, and to be usable for optical interconnections between optoelectronic smart-pixel arrays for 3-D optically interconnected multichip modules. For a small overall system volume the limited-fanout architecture can accommodate a much larger number of input and output nodes. However, the interconnection crosstalk of the limited-fanout space-variant architecture is relatively high due to noise from the diffractive optical element reconstruction. Therefore, a crosstalk reduction technique based on a modified design for diffractive optical elements is proposed. It rearranges the reconstruction pattern of the diffractive optical elements such that less noise lands on each detector region. This technique is verified by simulating one layer of an interconnection system with 128 x128 input nodes, 128 x128 output nodes, and 5 x5 nearest neighbor weighted connections from each input node to the output nodes.
We have also analyzed and simulated the effect of incorrect illumination wavelength and etch depth error on DOE reconstruction. The DOE-independent properties found in the simulations for binary-phase-level DOEs are theoretically verified. A way of cancelling the effect of etch depth error on DOE reconstruction by adjusting the illumination wavelength is described. The effect of incorrect illumination wavelength and etch depth error on DOE reconstruction with more than two phase levels has founded to be DOE-dependent.
Two design algorithms, Gerchberg-Saxton and simulated annealing, are studied and modified for better DOE performance. A dual-cost-function simulated annealing algorithm that produces better DOE performance than the original simulated annealing algorithm has been proposed. Finally, a new design approach that increases design freedom without increasing fabrication complexity is developed. It allows the design algorithm to find the best non-uniform quantization of phase levels and can generally obtain DOEs with substantially better performance.