BP neural network enables 2bits precision improvement for microring weight bank

We experimentally demonstrate a method for predicting and compensating the system error of a 3×3 on-chip microring resonator (MRR) weight bank using a pre-designed back propagation (BP) neural network. The system error can be quickly predicted and well compensated to improve the computation precision of optical matrix-vector multiplication (MVM), without bulky experimental devices. The results show that the computation precision can be increased from 5- bit to 7-bit. This work provides a weight accuracy improvement method for MRR-based photonic integrated chips for applications such as optical computing and optical communication.