Machine Learning in Short-Reach Optical Systems: A Comprehensive Survey

Analysis

Why This Paper Matters

This survey addresses a critical gap in the application of machine learning to short-reach optical communication systems, particularly passive optical networks (PONs). As data demand surges, PONs must evolve to support higher bandwidths, but traditional digital signal processing (DSP) equalizers like feed-forward/decision-feedback equalizers (FFE/DFE) and Volterra-based nonlinear equalizers face performance limitations and prohibitive complexity. The paper systematically reviews ML techniques that can handle stochastic phenomena in optical networks, offering a timely and comprehensive resource for researchers and engineers.

The introduction of a novel taxonomy for time-series ML methods is a key contribution. By categorizing methods into traditional, Fourier convolution-based, transformer-based, and time-series convolutional networks, the survey provides a structured framework that helps practitioners navigate the growing landscape of ML approaches. This taxonomy is particularly valuable because time-series models excel at capturing temporal dependencies and nonlinear patterns, which are prevalent in optical signal transmission. The paper's focus on complexity reduction strategies is also crucial, as cost-sensitive PONs demand practical solutions that can be implemented in hardware.

Technical Contributions

• Comprehensive review: The paper covers a wide range of ML applications in short-reach optical systems, including bandwidth request prediction, signal quality monitoring, fault detection, traffic prediction, and DSP-based equalization.
• Novel taxonomy: Introduces a structured classification of time-series ML methods into four groups: traditional methods (e.g., ARIMA, RNNs), Fourier convolution-based methods, transformer-based models, and time-series convolutional networks (e.g., TCNs). This taxonomy aids in understanding the strengths and weaknesses of each approach.
• Analysis of DSP equalizers: Provides a detailed comparison of conventional equalizers (FFE/DFE/Volterra) with ML-based alternatives, highlighting the marginal performance gains and high complexity of the former in cost-sensitive scenarios.
• Complexity mitigation: Discusses specific solutions to reduce hardware implementation complexity, such as model compression, pruning, and efficient architectures, paving the way for practical deployment.
• Future directions: Outlines prospective research areas, including hybrid models, online learning, and hardware-aware algorithm design.

Results

As a survey, the paper does not present new experimental results. Instead, it synthesizes findings from existing literature, noting that time-series ML models offer distinct advantages over frequency-domain models in capturing temporal dependencies and handling irregular or nonlinear patterns. The paper emphasizes that while ML can address stochastic phenomena, the performance improvements over conventional equalizers are often marginal, and complexity remains a significant barrier. The proposed taxonomy provides a qualitative framework for selecting appropriate methods, but quantitative comparisons are not provided.

Significance

This survey has significant implications for the optical communications community. By providing a structured taxonomy and highlighting complexity challenges, it guides researchers toward more practical ML solutions for short-reach systems. The emphasis on time-series methods aligns with the growing interest in transformer-based and convolutional architectures, which have shown promise in other domains. The paper's focus on hardware implementation is particularly relevant for industry adoption, as cost-sensitive PONs require efficient algorithms. Overall, this survey serves as a valuable reference for advancing ML in optical networks, potentially accelerating the deployment of intelligent, adaptive systems in next-generation broadband infrastructure.