optical performance modeling

To address the challenge of real-time synchronization between digital twins (DTs) and physical optical networks—limiting dynamic service adaptability throughout the network lifecycle—this paper proposes a dynamically updated DT framework for fiber channel performance prediction. Methodologically, it introduces the first DT dynamic update mechanism for optical networks, integrating physics-informed neural networks (PINNs), partial differential equation (PDE)-constrained hybrid modeling, real-time parameter identification, and a closed-loop feedback architecture. This enables synchronous, adaptive updates of multi-physical parameters—including Raman gain, amplifier frequency response, and connection loss—across C- and L-bands. Experimental results demonstrate a 100× speedup in prediction over conventional numerical methods; a 1.4 dB reduction in performance estimation error following device replacement; and validation of high accuracy (sub-dB), low latency (millisecond-level), and physical consistency in both large-scale simulations and live C+L-band field trials.

This work addresses key challenges in autonomous control of heterogeneous optical systems—namely, weak task comprehension, difficulty in multi-device coordination, and poor fault tolerance—by introducing AgentOptics, the first AI control framework that integrates an agent-based architecture with the Model Context Protocol (MCP). Through a structured tool abstraction layer, the framework maps natural language instructions to 64 standardized optical operations, enabling end-to-end task orchestration from single devices to full system-level workflows. The contributions include a unified tool interface, a multi-task evaluation benchmark, and integrated mechanisms for natural language understanding, multi-step coordination, error recovery, and closed-loop optimization. Evaluated on 410 tasks, AgentOptics achieves success rates of 87.7%–99.0%, substantially outperforming code-generation baselines (≤50%) and demonstrating robust system-level performance across five real-world scenarios, including DWDM configuration and 5G fronthaul optimization.

This work proposes the first intent-driven, closed-loop intelligent agent management framework for optical networks that complies with the T-API standard. Built upon the ReAct reasoning-and-acting paradigm, the framework innovatively integrates domain-specific composite tools with large language model (LLM) invocation interfaces. Experimental results, validated by domain experts, demonstrate that the proposed approach achieves a task accuracy of 90% while reducing token consumption by a factor of three compared to generic tool abstractions. This substantial improvement in both efficiency and precision underscores the framework’s effectiveness in enhancing autonomous network management capabilities.

This work addresses the lack of end-to-end automation and autonomous control in current multi-vendor, multi-layer IP-over-DWDM (IPoDWDM) networks, which hinders efficient service lifecycle management. To overcome this limitation, the authors propose a distributed, vendor-agnostic multi-MCP architecture that uniquely integrates MCP with Agentic AI. By synergistically combining SDN-based control, GNPy optical-layer modeling, real-time optical telemetry, and closed-loop feedback mechanisms, the proposed framework enables cross-vendor, cross-layer autonomous intelligent control and end-to-end service automation. Experimental validation on a real-world IPoDWDM testbed demonstrates that the approach significantly enhances network operational efficiency and intelligence.

This study addresses the trade-off between spectral efficiency and service disruption in reconfigurable optical networks by proposing a multi-period network planning framework that, for the first time, integrates multi-step deep learning–based traffic forecasting with multi-period resource allocation. An encoder–decoder model predicts traffic demands over multiple future time slots, which then drives both an integer linear programming (ILP) formulation and an efficient heuristic algorithm to optimize network configuration. While the ILP yields optimal solutions, the heuristic achieves comparable performance with significantly lower computational latency. Experimental results demonstrate that the length of the prediction horizon critically influences levels of resource over-provisioning, under-provisioning, and service disruption. The proposed framework flexibly accommodates diverse operator preferences—balancing spectrum conservation against disruption tolerance—while maintaining stringent quality-of-service guarantees.

Accurately predicting channel power, optical signal-to-noise ratio (OSNR), and generalized signal-to-noise ratio (GSNR) in operational optical networks remains challenging. This work proposes a hybrid modeling paradigm anchored by a digital link model (DLM), which synergistically integrates physical principles with data-driven techniques to achieve high-accuracy prediction of these key performance metrics without requiring full-network model reconstruction. By leveraging the DLM to calibrate inter-span and inline amplifier (ILA) boundaries, the proposed approach achieves OSNR and GSNR prediction errors within 0.39 dB and 0.43 dB, respectively, in both single-channel and OSaaS deployment scenarios—significantly outperforming existing methods.

This work addresses the lack of open-source tools capable of evaluating architectures and optimizing energy efficiency in 6G optical access–metro–core converged networks. To bridge this gap, we present SixGman, an open-source platform featuring a modular design and standardized interfaces, which enables the first end-to-end assessment of performance, cost, and energy consumption for novel architectures that bypass electrical-layer aggregation at HL3. The platform integrates key functionalities including dual-homing routing, quality-of-transmission estimation, joint spectrum–fiber allocation, total cost of ownership modeling, and energy consumption quantification, complemented by visualization capabilities. Experimental results on the Telefónica MAN157 topology demonstrate that the HL3-bypass architecture reduces total cost of ownership by up to 17.5% and cumulative energy consumption by 29.1%, while simultaneously improving traffic distribution and decreasing end-to-end latency.

This work addresses the critical challenge of unreliable decision-making in machine learning–driven optical networks, which can severely degrade service quality and system stability. To mitigate this risk, the paper proposes a runtime decision validation method grounded in explainable artificial intelligence (XAI). Prior to executing control actions, the approach generates explanations by analyzing feature importance and interaction patterns, then assesses decision credibility based on the coherence of these explanations and their consistency with underlying physical principles. As the first application of XAI to runtime verification in optical networks, the method effectively intercepts a substantial number of erroneous decisions in transmission quality classification tasks while preserving a high degree of automation, thereby significantly enhancing overall system reliability.

This study addresses the lack of unified autonomous capabilities in IP over DWDM (IPoDWDM) networks operating in multi-vendor environments. To this end, it proposes an agent-based AI architecture driven by a Multi-layer Control Platform (MCP), marking the first application of agent technology to the full lifecycle management of IPoDWDM networks. The approach integrates GNPy-based optical-layer simulation, real-time telemetry, and closed-loop control mechanisms to enable automated coordination and self-optimization across vendors in end-to-end multilayer networks. Experimental validation on a real-world testbed demonstrates the feasibility and effectiveness of the proposed solution in dynamic service provisioning, self-healing upon failures, and cross-layer resource coordination, significantly enhancing the network’s level of autonomy.

This work addresses the significant degradation in generalization performance of machine learning models when deployed across heterogeneous domains in optical networks, where varying topologies and operational configurations hinder adaptability. To tackle this challenge, the authors propose a representation learning framework that jointly integrates contrastive and classification learning, enabling the co-optimization of task-specific objectives and domain-invariant features within a unified latent space. This approach effectively captures relationships that are both task-relevant and invariant across domains. Notably, it is the first to combine contrastive and classification learning explicitly for cross-domain generalization in optical networks and supports few-shot fine-tuning. Evaluated on optical channel quality estimation, the method achieves superior cross-domain performance, substantially outperforming existing baselines with only a small number of labeled samples from the target domain.