Generalized Design of Diffractive Optical Elements Using Neural Networks

Diffractive optical elements (DOE) utilize diffraction to manipulate light in optical systems. These elements have a wide range of applications including optical interconnects, coherent beam addition, laser beam shaping and refractive optics aberration correction. Due to the wide range of applications, optimal design of DOE has become an important research problem. In the design of the DOEs, existing techniques utilize the Fresnel diffraction theory to compute the phase at the desired location at the output plane. This process involves solving nonlinear integral equations for which various numerical methods along with robust optimization algorithms exist in literature. However all the algorithms proposed so far assume that the size and the spacing of the elements as independent variables in the design of optimal diffractive gratings. Therefore search algorithms need to be called every time the required geometry of the elements changes, resulting in a computationally expensive design procedure for systems utilizing a large number of DOEs. In this work we have developed a novel algorithm that uses neural networks with possibly multiple hidden layers to overcome this limitation and arrives at a general solution for the design of the DOEs for a given application. Inputs to this network are the spacing between the elements and the input/output planes. The network outputs the phase gratings that are required to obtain the desired intensity at the specified location in the output plane. The network was trained using the back-propagation technique. The training set was generated by using GS algorithm approach as described in literature. The mean square error obtained is comparable to conventional techniques but with much lower computational costs