Lilin Yi

Fiber nonlinear distortion remains a fundamental bottleneck that constrains the capacity of optical transmission systems. By exploiting the temporal correlation of nonlinear effects, bidirectional long short-term memory (Bi-LSTM) networks have demonstrated notable potential for impairment mitigation. However, their multiple gating mechanisms incur substantial computational overhead, limiting practical deployment. Although bidirectional gated recurrent units (Bi-GRUs) reduce complexity through gate simplification, the presence of two gating operations still imposes a considerable burden. Here, inspired by first-order perturbation theory, we establish for the first time a direct connection between the GRU architecture and perturbative nonlinear compensation. Theoretical analysis reveals that the reset gate can be pruned without compromising the preservation of essential historical nonlinear information, thereby reducing the internal complexity of the GRU by 33% without performance degradation. Furthermore, by adopting a unidirectional propagation structure, the model eliminates the dual-pass iterations inherent to bidirectional networks, achieving an additional 50% reduction in computational complexity. We experimentally validate the proposed approach in a 21-channel, 60-GBaud polarization-division-multiplexed 16-QAM wavelength-division-multiplexing system over 1600 km of fiber transmission. The resulting unidirectional pruned GRU (Uni-PGRU) delivers a 0.60 dB Q-factor improvement over linear compensation, matching the performance of Bi-GRU and surpassing digital backpropagation with four steps per span. Crucially, at equivalent compensation performance, the Uni-PGRU requires only 1.9% of the computational complexity of Bi-GRU, 33.6% of Co-GRU, 9.1% of digital backpropagation, and 36.4% of learned digital backpropagation, establishing it as a highly efficient and practically viable solution for fiber nonlinearity mitigation.

In this paper, we propose a subcarrier-merged digital back-propagation (SM-DBP) scheme to compensate for fiber nonlinearity in subcarrier-multiplexing (SCM) systems. By merging the intensities of interfering subcarriers, SM-DBP significantly reduces the complexity of nonlinear compensation. For high-symbol-rate SCM systems that require more subcarriers to mitigate fiber nonlinearity and equalization-enhanced phase noise, SM-DBP offers a more prominent advantage in complexity, as its complexity is independent of the number of subcarriers. The proposed method is numerically and experimentally evaluated in single-channel dual-polarization 16-ary quadrature amplitude modulation (16-QAM) SCM transmission systems over a distance of 480 km. For 8-subcarrier and 16-subcarrier 240-GBaud simulation systems, 6-step SM-DBP achieves a complexity reduction of 78% and 85%, respectively, while maintaining performance comparable to that of SCM-DBP. In 16-subcarrier systems, 6-step SM-DBP also reduces the complexity by 50% while achieving an additional compensation gain of 0.42 dB, compared with CB-ESSFM. We have also verified the performance of SM-DBP in a 4-subcarrier SCM experimental system. The results indicate that SM-DBP reduces complexity by 74% while achieving the same performance as SCM-DBP. Compared with CB-ESSFM, SM-DBP achieves an additional compensation gain of 0.12 dB while reducing the number of CMs by 20%. The proposed algorithm is also validated in distributed Raman amplifier (DRA)-amplified systems through simulations and experiments. In a 320-km 8-subcarrier 240-GBaud DRA-amplified simulation system and a 185-km 4-subcarrier 40-GBaud DRA-amplified experimental system, SM-DBP reduces complexity by 80% and 61%, respectively, while maintaining performance comparable to that of SCM-DBP.

Industrial automation and the Internet of Things (IoT) are driving demand for optical access networks capable of supporting massive connectivity with deterministic low latency. Conventional passive optical networks (PONs) face scalability and cost limitations in ultra-dense deployment scenarios. Here we propose and experimentally demonstrate a polarization-carrier dual-reuse coherent orthogonal frequency division multiplexing PON (CO-OFDM-PON) architecture enabled by optical frequency comb (OFC) sources. The design maps data and carrier signals onto orthogonal polarization states. This enables carrier reuse for both downstream coherent detection and upstream transmission through injection-locked laser (ILL)-based carrier regeneration at the optical network unit (ONU). We comprehensively characterize the key subsystems. These include the OFC, ILL, and erbium-doped fiber amplifier (EDFA). This ensures stable multi-wavelength generation, carrier regeneration, and enhanced receiver sensitivity under high split ratios. Through simulation and experimental analysis of fiber nonlinearities in industrial PON scenarios, we identify an optimal per-channel launch power of 4 dBm. This power balances sensitivity and link budget requirements. Scalability analysis for standard PON reach is also provided. The system demonstrates 16-channel 100 Gb/s per wavelength downstream 16-QAM OFDM transmission. The link budget exceeds 34 dB with bit error rates (BERs) below the forward error correction (FEC) threshold of 1 × 10⁻². While the coherent ONU architecture offers superior spectral efficiency, it entails higher component costs than direct-detection alternatives due to the required coherent receiver and polarization management components.

Coherent passive optical networks (PONs) enabled by digital subcarrier multiplexing (DSCM) provide a highly flexible and low-latency access solution for optical network units (ONUs). However, in-phase and quadrature (IQ) impairments induced by transceiver imperfections (e.g., IQ skew, IQ amplitude imbalance, and IQ phase imbalance) cause severe image crosstalk on symmetric subcarriers. In the downstream of time- and frequency-division multiplexing (TFDM) PON, the estimation of IQ impairments becomes particularly challenging because partial reception results in the loss of symmetric subcarrier information. In this paper, we propose a time-domain-interleaved preamble (TDIP) scheme to estimate and pre-calibrate transmitter (Tx) IQ impairments. In addition, receiver (Rx) IQ skew is estimated using a modified frequency-domain phase detector (FDPD), while Rx IQ amplitude and phase imbalance are estimated using an improved Gram–Schmidt orthogonalization procedure (GSOP). Numerical simulations and experimental demonstrations are conducted to validate the excellent robustness of the proposed scheme under challenging conditions, including edge-subcarrier transmission, varying degrees of impairments, multiple coexisting impairments, and low received optical power (ROP). Specifically, for preset transceiver IQ skew ranging from −15 to 15 ps and IQ amplitude imbalance from −3 to 3 dB, the proposed scheme achieves estimation errors within ±0.6ps and ±0.4dB, respectively. Consequently, BER performance is significantly improved after the calibration of transceiver IQ impairments, particularly for the edge subcarrier. For 256 Gb/s downstream coherent TFDM PON over 20 km standard single-mode fiber (SSMF) transmission, the compensated receiver sensitivity reaches −22.2dBm, corresponding to a 30.2 dB power budget. This work offers a promising solution for precise transceiver IQ impairment monitoring in future flexible coherent TFDM PON.

Submarine optical fiber systems underpin global information infrastructure and continuous development is required, making accurate channel modeling essential for the evolution of submarine optical fiber systems and digital signal processing (DSP) evaluation. Traditional numerical approaches, such as the split-step Fourier method (SSFM) is hugely time-consuming. Deep learning (DL)-based models offer both rapid simulation and prediction capabilities as well as the advantage of providing waveform information; however, in long-haul fiber-optic link simulation scenarios spanning thousands of kilometers, the accuracy of cascaded DL models degrades sharply. To overcome these limitations, we propose a Cascaded Training Neural Network (CTNN) method designed to significantly improve the modeling accuracy of DL-based optical fiber models in ultra-long-haul scenarios. In this architecture, the models are cascaded during the training phase, allowing the long-distance loss to be used for updating the network parameters. Compared with the high-accuracy SSFM, the normalized mean square error (NMSE) of the waveform simulated by the CTNN-based model remains below 0.007 even after 10,000 km of propagation in long-haul submarine fiber wavelength-division multiplexing (WDM) systems, whereas the NMSE of models without CTNN is 0.08. Moreover, the model's prediction errors for both the signal-to-noise ratio (SNR) and Q-factor remain below only 0.15 dB after applying nonlinear or linear compensation algorithms at the receiver. Meanwhile, the simulation time was reduced to 1% of that of the SSFM method. The CTNN model enables fast and accurate modeling of submarine optical fiber WDM systems with transmission distances ranging from several thousand to over ten thousand kilometers based on DL methods while providing waveform information.

Waveform-level modeling of optical transmission systems serves as a pivotal tool for characterizing system properties, enabling cost-effective system design and optimization. However, conventional physics-based methods, such as the split step Fourier method (SSFM), suffer from high computational complexity and exhibit performance discrepancies when compared with experimental systems due to parameter mismatches. Meanwhile, although purely data-driven neural network (NN) approaches are widely employed in simulations, they often lack rigorous validation within real-world experimental setups. To address these challenges, this paper proposes a diffusion model enabled data-physics hybrid framework to achieve accurate and generalizable modeling of optical transmission systems. By leveraging the inherent iterative denoising process of diffusion models, the proposed method effectively characterizes system noise distributions. Compared to traditional generative adversarial networks (GANs), it demonstrates superior training stability and enhanced generalization capabilities. The effectiveness of the proposed model is experimentally validated over a 40-channel, 60 GBaud, 1600 km full C-band transmission system. Experimental results indicate that the diffusion model exhibits consistently higher stability and accuracy across multiple training trials than the GAN-based baseline. In generalization tests across varying launch powers and transmission distances, the diffusion model achieves average Q-factor errors of merely 0.06 dB and 0.09 dB. These results correspond to error reductions of 50% and 78%, respectively, compared to the GAN baseline. Furthermore, the model demonstrates exceptional cross-modulation format generalization, enabling it to be trained on a single modulation format and directly applied to others. Leveraging this frame work, we successfully achieve efficient transmission performance prediction, modulation format selection, and parameter optimization for nonlinear compensation (NLC) algorithms within our experimental platform. Ultimately, this proposed framework holds promise for lowering the barrier to experimental validation and expediting the intelligent evolution of next-generation optical networks.

Artificial intelligence (AI) agents powered by large language models (LLMs) are regarded as a prospective solution for autonomous optical power optimization in optical networks. In this paper, we develop an AI agent that incorporates human expertise, including physical knowledge, optimization knowledge, and optimization workflow. Leveraging provided toolsets, the AI agent explores the configuration space with a digital twin (DT) and deploys selected configurations on the actual network. Following an expertise-guided workflow comprising DT exploration, online evaluation, and fine-tuning, the AI agent performs efficient and reliable optical power optimization in a field-deployed network. Field-trial results show that the expertise-guided AI agent reaches the target quality of transmission (QoT) with an average of only 5.9 adjustments. Under the same scenario, Bayesian optimization (BO) and a genetic algorithm (GA) require 5.8× and 27.9× as many adjustments, respectively. The AI agent also ensures a reliable optimization process by maintaining a Q-factor margin of greater than 4 dB at each adjustment, whereas BO and GA fall below the forward error correction (FEC) threshold. Additionally, the expertise-guided AI agent delivers superior optimization across different wavelength-loading scenarios, demonstrating its capability to generalize across varying channel loads. Consequently, the expertise-guided AI agent provides a promising approach for autonomous, efficient, and reliable optical power optimization in future autonomous optical networks.