SAMA-IR: Sensitivity-Analysis Guided Memory-Aware Input Refinement Methodology for Large-Scale Optical Networks With Field-Trial Validation

The unprecedented demands on optical network capacity and stability necessitate the development of autonomous driving optical networks (ADONs), where accurate physical-layer modeling plays a fundamental role. Digital twins (DTs), while promising for precise quality of transmission (QoT) estimation, face significant challenges in maintaining modeling accuracy due to uncertain input parameters. Although input refinement (IR) approaches have been investigated to address this issue, they often struggle with massive uncertain inputs from various network components, especially in large-scale networks. The high dimensionality of input parameter spaces and complex physical interactions among different inputs make IR intractable. To tackle this challenge, this paper presents SAMA-IR, a comprehensive methodology for refining uncertain inputs of physical-layer models in large-scale optical networks. SAMA-IR introduces two key innovations: a sensitivity analysis (SA) framework that prioritizes parameter refinement based on their impact on system performance, and a memory aware (MA) weighting mechanism that maintains physical consistency throughout the refinement process. SAMA-IR is validated through field trials on a 440-km link with commercial equipment, where it demonstrated superior performance in both QoT metric and physical consistency metric. Using only 20 training samples, SAMA-IR reduces the Q-factor estimation error from −2.52±0.96 dB to −0.03±0.55 dB. Compared to conventional IR approaches, SAMA-IR can improve the optical power estimation accuracy by up to 2.32 dB at the 90th percentile. The results suggest that SAMA-IR offers a practical solution for maintaining accurate modeling with strong reliability and interpretability, particularly in scenarios involving complex parameter interactions.