We propose an enhanced log-likelihood ratio (E-LLR) technique for soft-decision forward error correction (SD-FEC) in intensity-modulation direct-detection (IMDD) systems with optical amplifiers and demonstrate its effectiveness in 7-core weakly coupled fiber transmission. In amplified IMDD, square-law photodetection and modulator nonlinearities cause amplified spontaneous emission noise to produce non-Gaussian signal statistics, invalidating conventional LLR assumptions and degrading SD-FEC performance. To address this issue, we abandon the classic additive white Gaussian noise (AWGN) model and introduce a channel model based on multi-parameter chi-square distributions that more accurately reflect optical channel noise. Unlike previous Gaussian and chi-square approaches, E-LLR uses a training sequence to estimate the variance and non-centrality parameters for each pulse-amplitude modulation (PAM) level, generating M distinct noncentral chi-square distributions for precise LLR computation in FEC decoding. Experimental results show that the proposed approach improves the receiver sensitivity by 1.13dB and reduces the decoding iterations of FEC compared with AWGN-model based schemes. These improvements contribute to reduced system complexity and decoding latency, while still achieving the required bit error rate threshold.

The escalation of symbol rates in next-generation coherent optical systems (e.g., 1.6 Tb/s and beyond) significantly exacerbates influence of transceiver frequency response impairments (FRIs). The system performance is degraded by FRIs of both transmitter (Tx) and receiver (Rx), being difficult to discriminate the individual contribution of each subsystem. Nevertheless, such discrimination is essential for the optimization of state-of-the-art transceiver designs, enabling effective mitigation of impairment propagation. To address this technical challenge, we propose a non-intrusive scheme for the simultaneous characterization and separation of Tx and Rx frequency responses (FRs). By leveraging tone-shift characteristics induced by frequency offset combined with multi-tone probe signals, the proposed method digitally separates the FRs of the Tx and Rx without requiring auxiliary components. Experimental results demonstrate the proposed scheme achieves high estimation accuracy, with deviations below 1 dB in amplitude and 0.25 rd in phase across a 20-dB bandwidth of 35 GHz. Furthermore, using the monitored responses for targeted pre- and post-compensation in a 69-GBaud DP-16QAM transmission system yields up to a 2-dB improvement in OSNR sensitivity at the 7% HD-FEC threshold. In addition, when combined with a reduced-tap MIMO equalizer, the proposed scheme still preserves a 1-dB OSNR sensitivity gain while reducing the equalization complexity by approximately 44%.. Therefore, the proposed scheme provides a practical solution for in-field transceiver characterization, paving the way for simplified DSP and enhanced performance in future ultra-high-speed coherent optical interfaces.

We propose a training sequence frequency-domain analysis scheme for simultaneous monitoring coherent Tx and Rx IQ skew in a single-shot measurement, achieving ±0.2ps accuracy over ≥ 16 symbol periods with IQ imbalance robustness and system compatibility.

We investigate multi-objective hyperparameter optimization of machine-learning-based nonlinear equalizers for coherent optical links. A sequentialized (S) MLP, also denominated time-delay neural network (TDNN) is adopted as a baseline model, and its hyperparameters (spanning discrete, categorical, and continuous variables) are jointly optimized using the Nondominated Sorting Genetic Algorithm II (NSGA-II). The objectives are the bit-error ratio (BER) after equalization and the computational cost expressed in floating-point operations. Results for a 400ZR-like DP-16QAM transmission show that NSGA-II efficiently constructs well-distributed Pareto fronts and identifies configurations that significantly reduce complexity while preserving near-optimal BER. Although the study focuses on MLPs due to their simplicity and interpretability, the proposed NSGA-II framework is directly extensible to more advanced architectures, such as CNNs and RNNs, where the hyperparameter space becomes even more complex. These results confirm the suitability of evolutionary multi-objective methods for designing practical nonlinear equalizers in coherent receivers.

We present a non-intrusive method to characterize coherent transceiver frequency responses without any hardware modification. Compensation using the characterization results improves 2 dB OSNR sensitivity at 7% FEC threshold in a 69-GBaud experimental system.

Space-division multiplexing (SDM) enhances system capacity by exploiting multicore or multimode fibers, but its multiple-input-multiple-output (MIMO) equalization often relies on matrix inversion. Newton-Schulz iteration (NSI) provides an efficient alternative, yet its convergence degrades under high dimensionality and strong mode-dependent loss (MDL). We propose a modified NSI (M-NSI) with cubic convergence, integrated into a training-sequence-aided frequency-domain MIMO (TSA-FD-MIMO) equalizer. To the best of our knowledge, this is the first experimental demonstration of M-NSI in a two-mode DP-16QAM system, achieving direct matrix inversion (DMI) comparable performance with only two iterations. Numerical simulations further confirm that M-NSI sustains reliable equalization with no more than three iterations even under 10-dB MDL.

We propose and experimentally demonstrate iterative receiver combining sMLSE and LDPC decoding in a 184-Gb/s PAM-4 signal transmission over 2-km SSMF, enabling a > 2 dB optical power sensitivity improvement compared with the non-iterative scheme.

An automatic grouping-enabled non-uniform quantization (AGNQ) approach is proposed for efficient neural network (NN) equalization in a 120 Gb/s 10 km directly modulated laser (DML)-based PAM-8 intensity-modulation/direct-detection (IM/DD) system. Compared with traditional uniform quantized feedforward neural network (UQ-FNN) and the additive power-of-two quantized FNN (APoT-FNN) approaches, AGNQ-FNN achieves comparable bit error rate (BER) performance below the 7% hard-decision forward error correction (HD-FEC) threshold using only 4-bit weight and bias quantization, whereas UQ-FNN and APoT-FNN require 10-bit and 8-bit quantization, respectively. Furthermore, when approaching the BER performance of an un-quantized FNN baseline, AGNQ-FNN reduces weight and bias memory usage by 47.9% and 35.5% compared with UQ-FNN and APoT-FNN, respectively. These results highlight the potential of AGNQ-FNN as a hardware-friendly equalizer for high-speed short-reach optical communication systems.

The feedforward equalizer (FFE), the decision feedback equalizer (DFE), and Volterra-based equalizer are widely adopted and shown effective in mitigating linear or nonlinear channel impairments in intensity-modulation direct-detection (IM/DD) systems. However, in practical hardware implementa tions, these equalizers often rely on floating-point (FP) representations for tap coefficients, resulting in high memory consumption and significant energy overhead. In this paper, we propose two novel equalization solutions with clustering-based non-uniform quantization schemes: automatic grouping-enabled non-uniform quantization (AGNQ) and k-means clustering-based non-uniform quantization (KCNQ), to address this challenge. In a 20 km C band pulse amplitude modulation (PAM)-8 IM/DD transmission system employing a directly modulated laser (DML), experiments are conducted to evaluate the performance of the proposed methods under various equalization schemes. At a transmission rate of 36 GBaud, KCNQ achieves a bit error rate (BER) below the 25% soft-decision forward error correction (SD-FEC) threshold in FFE using only 2-bit quantization, and meets the 7% hard-decision FEC (HD-FEC) threshold with 4-bit and 6-bit quantization in Volterra equalizer and Volterra-DFE (VDFE), respectively. Compared to the 32-bit FP baselines, KCNQ reduces coefficient storage by 83.0%, 81.3% and 76.4% for FFE, Volterra, and VDFE equalizers. In contrast, AGNQ achieves reductions of 80.2%, 74.5%, and 67.5%. Compared with AGNQ, KCNQ can further improve memory efficiency by 13.9%, 26.6%, and 27.4%. The proposed approach not only reduces storage requirements but also maintains reliable communication quality compared with original equalization approaches, offering a balanced trade-off between complexity efficiency and performance.

With the advancement of artificial intelligence (AI) technologies such as large language models and autonomous driving, the data traffic via optical interconnects in data centers has surged significantly. The stability of the optical interconnects relies on intelligent operation and maintenance (O&M). Integrated sensing and communication (ISAC) over fibers enables vibration sensing utilizing existing communication fibers, providing critical support for intelligent O&M in data centers. Compared to sensing in the coherent systems, it is difficult to use phase and state of polarization (SOP) monitoring for vibration detection in intensity-modulation and direct-detection (IM-DD) systems. In this paper, we propose a joint phase-based and SOP-based sensing scheme integrated in IM-DD systems. In the proposed scheme, the received IM-DD communication signals are tapped for sensing with a power ratio of 10%. Then the tapped signals are split for vibration sensing based on SOP and phase, respectively. In the phase-based sensing arm, a circulator, a 3×3 coupler and two Faraday rotating mirrors (FRMs) are used to build an unbalanced Michelson interferometer without phase fading and polarization fading. For the purpose of SOP-based sensing, a polarizer is used to monitor the vibration-induced SOP variations. Experimental results demonstrate that the proposed scheme enables vibration sensing based on both phase and SOP across a frequency range of 200 Hz to 10 kHz. Regarding the communication performance, the integration of the sensing system only induces 0.8 dB received optical power penalty. This vibration-sensing scheme based on both phase and SOP can be integrated into pluggable optical modules, providing an efficient and reliable solution for intelligent optical network O&M.

With an increasing demand of data rates and power efficiency, optical self-homodyne coherent (SHC) communication system with training-aided frequency-domain equalizer (TAFDE) has been proposed for short-reach optical interconnects or metropolitan area networks, due to its great potential to reduce receiver DSP complexity and power. However, the system is severely restricted by optical receiver in-phase and quadrature (IQ) impairments, including skew, amplitude imbalance and phase deviation. This work presents a comprehensive closed-form IQ mixing model that accounts for the conjugate symmetric interference and frequency-dependent (FD) imbalances. Based on the model, we explore the influence mechanism of receiver IQ impairments on the zero-autocorrelation property of constant amplitude zero auto-correlation type (CAZAC-type) training sequences (TS). We show that Zadoff-Chu (ZC) sequences based TAFDE (ZCS-TAFDE) is vulnerable to these IQ impairments due to autocorrelation deterioration of the complex-valued CAZAC-type TS. To overcome this limitation, an m-sequences based TAFDE (MS-TAFDE) scheme is proposed and enables significantly improved immunity against IQ impairments. An experimental comparison is conducted in a 28-GBaud polarization multiplexed 16 quadrature amplitude modulation SHC system. Under identical IQ conditions, MS-TAFDE exhibits almost no optical signal-to-noise ratio (OSNR) penalty at a 7% forward error correction threshold, whereas ZCS-TAFDE incurs a 5-dB OSNR penalty at an IQ skew as low as 5 ps. Numerical stress tests further reveal that MS-TAFDE tolerates FD amplitude and phase deviations up to 15.58 dB and 83.6°, respectively, within a 1-dB Q-penalty boundary. When multiple IQ impairments are present simultaneously, the ZCS-TAFDE scheme undergoes over 5 dB penalty and may fail, while MS-TAFDE shows strong tolerance when all the three IQ impairments coexist. The experimental results agree well with the theoretical predictions using the complex valued IQ mixing impairment model. Therefore, the MS-TAFDE scheme delivers high-precision equalization with minimal complexity in the presence of multiple coexisting receiver IQ impairments, which facilitates SHC implementation for optical interconnects.

We proposed a neural architecture search (NAS) framework that automatically optimizes nonlinear equalizers. Experiments on an 800-Gb/s 960-km transmission show that the NAS-derived equalizers deliver a 0.3 dB higher peak Q-factor than manual baselines.

The increasing demand for high-speed, low-cost data transmission in short-reach optical interconnects has cemented intensity-modulated direct-detection (IM/DD) optical systems as a practical solution. However, nonlinear impairments stemming from square-law detection under chromatic dispersion channels, as well as the nonideal optoelectronic components, pose significant challenges to signal integrity. To address these issues, neural network (NN)-based equalization techniques have garnered considerable attention due to their inherent capability to model complex channel behaviors. While these approaches offer performance improvements, their deployment is often constrained by computational burden and hardware resource consumption. This chapter focuses on recent advances in hardware-friendly NN designs for IM/DD equalization, highlighting effective complexity reduction strategies such as network pruning, low-precision quantization, and multi-task learning. We examine a range of NN architectures and evaluate their adaptability to low-power, low-latency implementations. Furthermore, this chapter discusses practical trade-offs between accuracy and computational cost, with insights into deployment in real-time systems. We outline future directions toward scalable and resource-efficient NN equalization, paving the way for intelligent digital signal processing in next-generation optical communication links.

In this work, a systematic numerical investigation is undertaken of side-polished mode-selective couplers (SP-MSCs) and fused biconical taper mode-selective couplers (FBT-MSCs) based on an identical standard single-mode fiber (SMF) and few-mode fiber (FMF) platform. The influence of key geometric parameters, including the residual cladding thickness (RCT), taper ratio (TR), and core-to-core distance, on phase-matching mechanisms is analyzed in detail. The numerical results show that SP-MSCs can achieve superior mode selectivity at specific operating wavelengths, with intermodal crosstalk for LP11 and LP21 modes suppressed to values between −30dB to −40dB. While generally exhibiting higher levels of crosstalk, FBT-MSCs are numerically predicted to exhibit lower sensitivity to core-to-core distance deviations than SP-MSCs. In the simulated structural-sensitivity analysis, their coupling efficiencies remain above 80% within the investigated lateral-offset range of up to ±0.8μm. In addition, the wavelength-dependent coupling efficiencies of both SP-MSCs and FBT-MSCs remain mode-dependent within the investigated spectral range.

Space division multiplexing systems face critical challenges in monitoring coupled channel impairments. To address these challenges, we propose a low-complexity, training-aided framework for the joint and independent monitoring of chromatic dispersion (CD), mode-dependent loss (MDL), and differential mode group delay (DMGD). Leveraging the ideal correlation properties of constant amplitude zero auto-correlation (CAZAC) sequences, the proposed scheme decouples these impairments via frequency-domain analysis: CD is extracted from the quadratic phase of the channel determinant, DMGD from the frequency-dependent eigenvalue phase difference, and MDL from the singular value spread. Furthermore, an adaptive bandwidth selection strategy is introduced to resolve the trade-off between estimation resolution and aliasing limits. Experimental validation in a few-mode fiber system demonstrates high-precision monitoring results. For CD estimation, the scheme achieves an absolute error below 0.3 ps/nm in the low-dispersion regime (0-15,300 ps/nm) and maintains a relative error within 0.2% in the high-dispersion regime (15,300-34,000 ps/nm). Additionally, the monitoring errors are less than 0.1 ps for DMGD (0-100 ps) and 0.3 dB for MDL (0-15 dB). The results also confirm robust operation under low optical signal-to-noise ratios (down to 14 dB). By eliminating the need for ultra-wideband oscilloscopes and ensuring seamless integration with digital signal processing workflows, this method offers a scalable and costeffective solution for on-line impairment diagnosis in next-generation multi-dimensional transmission systems.

We propose a novel hardware-efficient optical frequency comb enabled self-homodyne joint WDM-SDM reception without using a wavelength de-multiplexer (De-MUX) at the receiver side. In this architecture, multiple wavelength signals can be simultaneously recovered by a single integrated coherent receiver (ICR), followed by multiple input multiple output digital signal processing (MIMO-DSP). To realize such joint reception, two different (de-)multiplexing schemes are further investigated and compared. The first scheme employs symmetric optical delay lines at both the transmitter and the receiver, whereas the second adopts a hybrid configuration by incorporating a wavelength De-MUX at the transmitter and optical delay lines at the receiver, respectively. These two schemes are referred to as the symmetric scheme and the hybrid scheme, respectively, throughout this paper. Extensive numerical investigation and comparison are undertaken to evaluate various transceiver impacts such as delay mismatch on the transmission performance for both schemes. Results show that the hybrid scheme significantly outperforms the symmetric scheme on aspect of resilience to transmitter imperfection. The superior performance of the hybrid scheme is further validated by experimental results of a 25 km four-core weakly coupled multicore fiber (MCF) proof-of-concept transmission system. Experimental results reveal that the hybrid scheme successfully demodulates dual-wavelength, dual-polarization 16QAM signals using a single ICR and MIMO-DSP. An optical signal-to-noise ratio (OSNR) of 21.5 dB is required for a 18 GBaud multidimensional 16QAM signal to reach the hard-decision forward-error-correction (HD-FEC) threshold, while maintaining robust operation under differential group delay (DGD) of up to 1700 ps and fiber-core perturbations. These results demonstrate that the hybrid scheme provides a hardware-efficient and physically robust solution for multidimensional signal joint recovery for the next-generation high-density, high-capacity optical communication systems.

With the recent advancements in large language models, high-speed optical interconnects are required between multiple servers in data centers to support model training. The stability of optical interconnects is critical to the model training. Integrated sensing and communication over fiber enable the effective utilization of existing communication fibers for vibration sensing, thereby achieving intelligent operation and maintenance of optical networks. Compared to coherent systems, vibration sensing and event recognition integrated into intensity-modulation and direct-detection (IM-DD) systems are still in the early stages, and the recognition accuracy is limited by the singular sensing source. In this paper, we propose a sensing hardware system that simultaneously monitors vibration-induced state of polarization (SOP) and phase variations for IM-DD systems. By fusing the phase-based and SOP-based sensing features using a software convolutional neural network (CNN) model, accuracy enhancement in vibration event recognition is achieved. In the hardware system, 10% of the received optical signal is tapped for sensing. This tapped portion is then split, with 90% allocated to phase-based sensing using an unbalanced Mach-Zehnder interferometer with a 3 x 3 coupler for fading-free phase demodulation and the remaining 10% to SOP-based sensing utilizing a polarization beam splitter (PBS) for SOP variation monitoring. Within the software CNN model for vibration recognition, the SOP-based and phase-based sensing features are concatenated along the channel dimension and adaptively fused using a channel attention mechanism. Experimental results demonstrate that the proposed scheme achieves vibration sensing within a frequency range of 100 Hz to 5 kHz, even at an extremely low received optical power (ROP) of -20 dBm. Furthermore, the model with the fusion of SOP-based and phase-based features improves the average event recognition accuracy by 3% compared to using the SOP data alone and by 14.8% compared to relying solely on the phase data. The recognition accuracy improvement arises from the complementarity of phase-based and SOP-based sensing features under different vibration events. This accurate vibration recognition can effectively improve link maintenance efficiency and promises to enhance the reliability of optical networks.

A low-complexity non-integer oversampling timing recovery scheme is proposed and specially designed for real-time FPGA implementation, which shows superior hardware efficiency and reliability under various roll-off and oversampling conditions. © 2026 The Author(s)

This work presents a novel hardware-efficient FPGA implementation of a real-time parallel loop-unrolled decision feedback equalizer (DFE) for PAM-4 IM/DD optical links. The proposed DSP architecture integrates timing recovery, a 5-tap FFE, and a novel single-tap parallel DFE leveraging an $8\times 8$ loop-unrolled structure and a hardware-friendly error-based tap update mechanism thus effectively reduces processing complexity and clock delay. The implementation occupies only 4.2% LUTs, 2.75% FFs, and 6.8% DSP48E2 resources of a Xilinx XCVU13P chip. A successful 29.4912 Gb/s real-time PAM-4 signal transmission over a 10 km SSMF is demonstrated using a cost-effective directly modulated laser (DML) and achieving a BER below the 7% HD-FEC threshold, indicating a scalable and efficient solution for high-speed short reach optical data links.

Integrated phase-based vibration sensing and communication faces challenges in achieving high-sensitivity detection under cost-effective commercial external-cavity lasers (ECLs), primarily limited by laser phase noise and frequency offset drift. To address this, we propose a machine-learning-enhanced phase-based vibration sensing scheme for coherent digital subcarrier multiplexing (DSCM) systems. In this scheme, a generative adversarial network (GAN)-based digital twin (DT) system is developed to synthesize massive forward-phase sensing datasets, overcoming experimental data scarcity. Principal component analysis (PCA) confirms high consistency between generated and experimental noise characteristics. The generated datasets are used to train a multi-layer perceptron (MLP) and an attention-enhanced MLP (AMLP). The trained artificial neural networks (ANNs) are deployed in receiver-side digital signal processing (DSP) for vibration-induced phase extraction. In experimental validation under 10 kHz vibration, the MLP-based scheme achieves a 13 dB sensing signal-to-noise ratio (SSNR) enhancement over the traditional band-pass filter (BPF) scheme. The AMLP-based scheme further delivers a 7 dB additional SSNR gain by leveraging attention mechanisms to emphasize critical phase features. Furthermore, the ANNs-based scheme successfully extracts multi-type vibrations without prior vibration information, demonstrating the flexibility of the ANNs-based schemes. This work establishes a machine-learning-driven framework that enhances phase-based sensing in commercial ECLs-based DSCM systems, providing a viable pathway for intelligent optical network operation and maintenance.