Robust Training of Optical Neural Network with Practical Errors using Genetic Algorithm: A Case Study in Silicon-on-Insulator-Based Photonic Integrated Chips

Optical neural network (ONN) utilizes light to process a mass amount of information in parallel using photonic integrated chips. It has great potential to bypass the limitation of Moore’s law and overcome the inherent bandwidth bottleneck in electronics enabled by the >10 THz wide optical telecommunications band. One of the main challenges for the realization of ONNs is how to avoid practical errors, including various device parameter errors during fabrication and the limited phase shifter control precision. Characterization of each individual chip is possible but time-consuming. To address this issue, in this paper, we propose a robust method to train a series of ONN chips with practical errors using the genetic algorithm (GA). The effect of different parameter errors on the data classification accuracy is analyzed, including the errors in phase shifters, coupling coefficient or extinction ratio, optical absorption loss, and photodetection noise. As a proof-of-concept demonstration, a simulated feedforward ONN is implemented to identify a customized dataset with four classes and four uncorrelated features. The simulation results show that our proposed method could increase the average classification accuracy from 86% to 96% for 50 erroneous ONN chips, approaching the ideal ONN accuracy of 99.69% and demonstrating the effectiveness for significant enhancement in training robustness against practical errors.