Hidden Feature Extraction Learning and End-to-End Joint Equalization With LDPC Decoding Method for Optical Interconnect

Abstract

The vertical-cavity surface-emitting lasers and multimode fiber (VCSEL-MMF) solution has successfully emerged in optical interconnects (OIs), which meets the challenges of signal impairments in ultra-high-speed scenarios. Deep learning (DL) techniques, which can approximate any nonlinear function, enable the design of communication systems by carrying out the optimization in a single end-to-end (E2E) process including the transceivers as well as communication channels. In this paper, we propose a hidden feature extraction learning method for neural network equalization to improve training efficiency without increasing computational burden. Superior bit error rate (BER) is demonstrated in achieving 288 Gb/s 100 m VCSEL-MMF interconnect compared with black-box training strategy. Furthermore, an E2E joint equalization and low-density parity-check (LDPC) decoding method is proposed to improve the overall performance. Based on the autoencoder (AE) architecture, the E2E network involves a digital pre-distorter (DPD), a digital optical link model, a feed forward equalizer (FFE) and a deep learning based Normalized Offset Min-Sum (DL-NOMS) LDPC decoder. Experimental results demonstrate that the BER performance of the proposed E2E scheme is two-magnitude lower than the E2E equalization and FFE+DL-NOMS decoding method in back-to-back (BTB) and 100 m VCSEL-MMF links.