Microring-Based Dual-Drop Weight Bank With Improved Linearity

Abstract

In this work, we demonstrated a microring-based dual-drop weight bank (DD-WB) for photonics neural networks. By tuning the resonant wavelengths of a microring pair, a large and tunable linear region can be realized for weight matrix mapping. In addition, a specialized configuration method is presented to efficiently employ weight values upon the DD-WB. To verify the performance of our scheme, we designed and fabricated a silicon photonics chip with 1-channel and 2-channel DD-WB on-chip structures. Experiments show improved linearity and good robustness of the fabricated chip by performing the weight encoding. Furthermore, an on-chip photonic neural network for Iris classification task is implemented, which achieved an accuracy of 93.33% with the testing set.