MT-OPMNet: Attention-Enhanced Multi-Task Deep Learning for Joint OSNR Estimation and Modulation Format Recognition in Elastic Optical Networks

Abstract

Optical performance monitoring (OPM) underpins quality of transmission in modern elastic optical networks. This paper introduces MT-OPMNet, a multi-task deep learning framework for joint optical signal-to-noise ratio (OSNR) estimation and modulation format identification (MFI) from asynchronous amplitude histograms extracted at coherent receivers. The model combines a shared one-dimensional convolutional backbone, a channel-aware attention module and two task-specific heads for parameter-efficient inference. Validation uses a corrected nine-channel wavelength-division-multiplexed simulation dataset covering distances from 500 to 3000 km, symbol rates of 28 GBd and 64 GBd, five launch-power levels and five modulation formats. At 28 GBd, MT-OPMNet achieves an OSNR root-mean-square error of approximately 2.4 dB; over the grouped mixed-rate test split, it achieves an average OSNR RMSE of 3.4 dB and an MFI accuracy of 97.5%, while reducing trainable parameters by 23.8% relative to two separate single-task models. The analytical channel model is further cross-validated using a split-step Fourier waveform simulator, achieving mean histogram cosine similarity of 0.976 and zero-shot transfer to SSFM-generated waveforms with 100% MFI accuracy and 2.1 dB OSNR RMSE over 14–22 dB OSNR. CPU inference latency remains suitable for low-latency monitoring studies.