Optical analog computing devices designed by deep neural network

• Published in: Optics communications Vol. 458; p. 124674
• Main Authors: Zhou, Yi, Chen, Rui, Chen, Wenjie, Chen, Rui-Pin, Ma, Yungui
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2020
• Subjects: Analog optical signal processing; Multilayer design; Neural networks; Optics in computing; Multilayer design; Analog optical signal processing; Neural networks; Optics in computing
• ISSN: 0030-4018; 1873-0310
• DOI: 10.1016/j.optcom.2019.124674

We proposed a multilayered spatial optical differentiator designing method by use of the deep neural network (DNN). After trained for approximately 30 h, the DNN is able to predict the reflection coefficient of a 12-layer multilayer film with high fidelity (validation mean squared error < 2.4×10−4). As a useful example, a second-order spatial optical differentiator was then designed. Compared with the general optimization method, the machine learning could help to quickly generate a wavefront computing device at an about 6-times faster speed. The performance of the designed device is confirmed from the comparison with the theoretical ideal operation output. Another first-order spatial optical differentiator was also designed to validate the generality of the method. The results indicate that the DNN may have a bright future in designing devices capable of all kinds of complex time-space wavefront mathematical operation, in particular based on the multilayer material systems.