This paper explores operationally reliable combinations of AI and physics through digital twins to optimize optical networks. Key methodologies and applications are discussed, addressing security, scalability, automation and sustainability.

We experimentally demonstrate sequential parameter refinement leveraging lightpath diversity in presence of various levels of uncertainty on parameter knowledge to iteratively improve QoT prediction accuracy by up to 1.8 dB.

We characterize SOP event localization with a commercial transponder on a 150 km link. Fast events achieved below 20 m of standard deviation. We propose a localization improvement of low-speed events, achieving constant performance up to 10 krad/s.

We study the use of longitudinal power monitoring (LPM) to improve SNR prediction for optical links where lumped power losses can be present. We define three scenarios: 1) no knowledge of the presence of power losses, 2) the amplifier gives information on the value of power loss and 3) LPM allows the estimation of both the loss location and the loss value. We determine the prediction gain between scenarios in a case with no uncertainty. We later introduce an uncertainty on the fiber attenuation. We show that LPM can significantly reduce the SNR prediction uncertainty interval down to 0.2 dB while for the amplifier scenario it is 1.7 dB.

We present a memory-efficient, computationally optimized longitudinal power monitoring method scalable to numerous spatial samples for embedded systems. Results show < 0.6 dB accuracy penalty across > 75% of the link versus linear least squares.

We propose a novel method for continual PDL monitoring that first detects PDL drift by tracking SNR distributional shifts at the receiver, and then provides a prediction of the updated PDL value. This approach achieves a root mean square error (RMSE) of approximately 0.15 dB over time.

We leverage binary classification models to provide digital twins with forecasting ability, to proactively mitigate the impact of failures on availability. Our preferred model achieves 98.5% accuracy on data collected from a live production network.

We experimentally measure the specific spectral signature of strong fiber bending and propose a metric to detect strong bending events. Leveraging telemetry from a production network we report short-lived anomalies which exhibit the aforementioned signature.

We propose a multi-stage computation for longitudinal power monitoring reducing the memory footprint compared to linear least square approach. We demonstrate a power accuracy penalty lower than 0.1dB over more than 85% of link.

As deploying large amounts of monitoring equipment results in elevated cost and power consumption, novel low-cost monitoring methods are being continuously investigated. A new technique called Power Profile Monitoring (PPM) has recently gained traction thanks to its ability to monitor an entire lightpath using a single post-processing unit at the lightpath receiver. PPM does not require to deploy an individual monitor for each span, as in the traditional monitoring technique using Optical Time-Domain Reflectometer (OTDR). In this work, we aim to quantify the cost and power consumption of PPM (using OTDR as a baseline reference), as this analysis can provide guidelines for the implementation and deployment of PPM. First, we discuss how PPM and OTDR monitors are deployed, and we formally state a new Optimized Monitoring Placement (OMP) problem for PPM. Solving the OMP problem allows to identify the minimum number of PPM monitors that guarantees that all links in the networks are monitored by at least n PPM monitors (note that using n>1 allows for increased monitoring accuracy). We prove the NP-hardness of the OMP problem and formulate it using an Integer Linear Programming (ILP) model. Finally, we also devise a heuristic algorithm for the OMP problem to scale to larger topologies. Our numerical results, obtained on realistic topologies, suggest that the cost (and power) of one PPM module should be lower than 2.6 times that of one OTDR for nation-wide and 10.2 times for continental-wide topology.

We predict anomalous fiber macro-bending conditions by using an artificial neural network, trained with data from a laboratory experiment. We discuss the performance and robustness of two feature extraction methods for different monitoring equipment capabilities.

We explore state-of-polarization sensing as a promising candidate for large-scale deployments thanks to longdistance monitoring capabilities. We further discuss scalability aspects related to monitoring placement to reduce telemetry streams, hence lowering alarms redundancy and data storage requirements. We also benchmark AI-based event classification techniques.

We propose a physics-based digital twin to predict the statistical QoT distribution of a realistic optical lightpath. We demonstrate up to 0.73 dB accuracy improvement in worst-case SNR prediction for short distance transmissions in linear regime.

We experimentally analyzed nonlinear effects from adjacent QPSK channels and ASE noise loading over a long-haul link, including the impact of CPE. ASE-induced penalties were smaller than predicted by format-dependent models, and CPE was found to significantly mitigate both intra- and interchannel nonlinearities.

We propose a method to localize and quantify anomalous polarization dependent loss through signal to noise ratio distributions monitored at the receiver. We demonstrate that anomalies can be localized with 100% accuracy and estimated with 0.01dB uncertainty in simulations and with 0.2dB uncertainty in experiments. (c) 2025 The Author(s)

We present an accuracy comparison between two longitudinal power profile estimation methods, the minimum mean square error (MMSE) method and the deconvolution of the correlation-based method. We first numerically evaluate the accuracy of each method and benchmark the deconvolution method versus the MMSE with respect to span input power, number of spans and modulation format. For both methods, we obtain a similar accuracy - root-mean square error (RMSE) of 0.14 dB at 5 dBm - at a given input power. We also show that using 16-QAM or 64-QAM for the modulation format instead of QPSK allows the deconvolution method to be more accurate and robust with the number of spans. Secondly, we perform a first experimental demonstration of the deconvolution method and assess the accuracy of both methods with commercial transponders in an optical networking testbed. We measure the RMSE of both - MMSE and deconvoluted - profiles with respect to the optical time-domain reflectometer (OTDR). We obtain close RMSE values as the average RMSE difference between both profiles is below 0.28 dB for the lightpath excluding the first span.

We propose a digital twin of coherent receivers based on an extended physical model for predicting quality of transmission under receiver power variations. We experimentally validate it, demonstrating an accuracy improved by up to 1.5dB.

Forecasting state of polarization (SOP) in aerial fibers poses several challenges due to exposure to dynamic und unpredictable environmental factors, such as wind-induced vibrations, leading to multi-timescale SOP variations. Traditional statistical and deterministic approaches may struggle to accurately model the complex, nonlinear, and multi-scale dynamics of these SOP fluctuations, making them less effective for real-world scenarios. To address these challenges, this study proposes a weather-adaptive forecasting framework combining discrete wavelet transform and artificial neural networks to predict SOP changes across varying timescales. By integrating environmental data, the model effectively captures rapid fluctuations and gradual drifts, significantly improving forecasting accuracy. Validation using field data across two locations demonstrates the model's reliability for both short-term and long-term predictions. Including environmental factors namely wind speed, temperature, and humidity improves long-term forecasting accuracy by over 64.5% compared to models taking as input SOP change data alone. Wind gusts enhance short-term forecasting under extremely windy conditions. Our proposed framework outperforms conventional techniques, offering a reliable framework for SOP change prediction and advancing proactive optical system monitoring.

We review methods for optical link tomography, particularly receiver-based power profile estimation. We highlight key works including a demonstration over 10,000 km and an accuracy comparison between linear least squares method and deconvoluted correlation-based method.

We demonstrate linear and nonlinear SNR estimation over a commercial 130 Gbaud coherent transponder by using received signal features. The obtained linear and nonlinear SNR estimation error were within 0.7 dB and 1.3 dB, respectively.