Tunable Photoinduced Liquid Crystal Retarders for All-Optical Diffractive Deep Neural Networks

Abstract

An all-optical diffractive deep neural network (D²NN) consists of deep-learning-based design of passive diffractive layers and uses light to perform massive computations at the speed of light with zero extra power consumption, exhibiting advantages of large bandwidth, high interconnection, and parallel processing capability. In this paper, we introduce a novel approach utilizing a 5-layer all-optical D²NN constructed with photoinduced liquid crystal (LC) alignment technology to create LC-based tunable phase retarders as artificial neural layers. The D²NN architecture leverages microscale multidomain LC retarders as optical neurons to manipulate the geometric phase of incident light. We systematically simulate pixel-level displacements to enhance alignment tolerance during experiments, achieving robust resilience against misalignment interference with a 2-pixel tolerance in the x and y directions. By actively tuning the LC retarders with external voltage, we optimize the alignment strategy for all network layers, incorporating specially designed concave or convex lenses at each LC retarder for precise alignment in the x, y, and z directions. Through training with a handwritten dataset from MNIST, the D²NN demonstrates a simulated accuracy of 94.17% with a 2 pixel misalignment tolerance. Experimental validation achieves a classification accuracy of 89% with 500 random digits from the test dataset. This research showcases the potential for network miniaturization, integration, and compatibility with visible light, underscoring the practical applicability of an all-optical D²NN for diverse real-world applications.