Optical networks form the backbone of today’s digital infrastructure, transporting the majority of global data traffic across long-haul, metro, and data center domains. As they evolve to accommodate exponential traffic growth and support latency-critical services such as 5G, 6G, and distributed AI, future optical networks face unprecedented demands for capacity, resilience, and automation. At the same time, physical impairments—including amplifier noise, Kerr nonlinearities, stimulated Raman scattering (SRS), filtering penalties, and transponder back-to-back limitations—accumulate along transmission paths and constrain the signal reach through its Quality of Transmission (QoT). While analytical models can capture these effects, their predictive accuracy in live networks is hindered by uncertain parameters and limited observability, forcing operators to adopt conservative margins that reduce the utilization efficiency of network resources such as capacity and power. This thesis aims to advance the design of Robust and Automatically-Driven Optical Networks, and the main contributions can be summarized in three parts as follows. 1. Experimental Quantification and Mitigation of QoT Degradation through Power Re-optimization: A key challenge in optical networks is that the practical "set-and-forget" power setting strategy, widely adopted in today’s operations, ignores cumulative load-dependent effects arising from incremental traffic loading and the absence of power re-optimization for already established channels. As a consequence, the QoT of existing services progressively degrades as additional services are loaded into the network. To address this, I experimentally quantify such QoT degradation in multi-node mesh testbeds and show its impact on network performance. I further propose both static and dynamic power re-optimization strategies integrated into a software-defined networking (SDN) control plane called "AI-Light". Experimental demonstrations confirm that these strategies can substantially recover SNR degradation and pave the way for automated and resilient network operation. 2. Development of Input Refinement (IR) Techniques for Digital Twins: The accuracy of analytical QoT models depends critically on physical parameters such as insertion losses, amplifier gains, and fiber characteristics. However, these parameters are often uncertain or hidden from operators in live networks. To overcome this limitation, I design and validate four complementary Input Refinement (IR) techniques that leverage in-service monitoring data to calibrate digital twins. Specifically: (i) Passive IR (PIR) achieves OMS-level accuracy from a single snapshot and extends to C+L systems; (ii) Active IR (AIR) perturbs amplifier gains to exploit SRS signatures, enabling span-level anomaly detection and localization; (iii) Incremental IR (IIR) leverages multiple snapshots for progressive refinement and supports closed-loop autonomous optimization; and (iv) a Hybrid IR+PPE method integrates power profile estimation with IR, achieving absolute calibration of insertion losses and longitudinal power evolution. Together, these methods establish accurate, robust, and operationally viable digital twins for QoT prediction and optimization. 3. Benchmarking of Transponders for Data Center Interconnect (DCI): Beyond backbone and metro networks, the rapid expansion of cloud and AI workloads has made data center interconnects (DCIs) a critical domain for optical networking. Hyperscale operators must balance cost, reach, and capacity when choosing among different coherent transponder families. To provide design guidelines, I conduct a systematic benchmarking of 800G ZR, ZR+, and high-performance transponders using simulations that incorporate experimentally measured amplifier dynamics in short-reach DCI networks. The study analyzes capacity scaling under varying span losses, multiplexing schemes, and spectrum expansion scenarios, identifying the trade-offs among different solutions. The results highlight effective design and loading strategies for cost-efficient and high-capacity DCI deployments. In summary, this thesis demonstrates that integrating analytical QoT models with refined monitoring, SDN control, and digital twins enables scalable automation, margin-efficient operation, and actionable design guidelines for next-generation optical networks.
Le reti ottiche costituiscono la spina dorsale dell’infrastruttura digitale odierna, trasportando la maggior parte del traffico dati globale attraverso domini long-haul, metropolitani e dei data center. Con la loro evoluzione per accogliere una crescita esponenziale del traffico e supportare servizi a bassa latenza come il 5G, il 6G e l’intelligenza artificiale distribuita, le future reti ottiche affrontano richieste senza precedenti in termini di capacità, resilienza e automazione. Allo stesso tempo, gli effetti fisici degradanti — tra cui il rumore degli amplificatori, le non linearità di Kerr, la diffusione Raman stimolata (SRS), le penalità di filtraggio e le limitazioni back-to-back dei transponder — si accumulano lungo i percorsi di trasmissione e limitano la portata del segnale attraverso la sua Qualità di Trasmissione (QoT). Sebbene i modelli analitici possano descrivere tali effetti, la loro accuratezza predittiva nelle reti operative è ostacolata da parametri incerti e limitata osservabilità, costringendo gli operatori ad adottare margini conservativi che riducono l’efficienza di utilizzo delle risorse di rete, come capacità e potenza. Questa tesi mira ad avanzare la progettazione di reti ottiche robuste e a guida automatica, e i principali contributi possono essere riassunti in tre parti come segue. 1. Quantificazione sperimentale e mitigazione del degrado di QoT tramite riottimizzazione della potenza: Una sfida fondamentale nelle reti ottiche è che la pratica strategia di impostazione della potenza “set-and-forget”, ampiamente adottata nelle operazioni odierne, ignora gli effetti cumulativi del caricamento incrementale del traffico. Di conseguenza, i nuovi canali provisionati possono degradare significativamente la QoT dei servizi già esistenti. Per affrontare questo problema, quantifico sperimentalmente tale degrado di QoT in banchi di prova (testbed) a maglia multi-nodo e mostro il suo impatto sulle prestazioni di rete. Propongo inoltre strategie di riottimizzazione della potenza, sia statiche sia dinamiche, integrate in un piano di controllo SDN (Software-Defined Networking) denominato “AI-Light”. Le dimostrazioni sperimentali confermano che tali strategie possono recuperare sostanzialmente la degradazione di SNR e aprire la strada a un funzionamento di rete automatizzato e resiliente. 2. Sviluppo di tecniche di Input Refinement (IR) per gemelli digitali: L’accuratezza dei modelli analitici di QoT dipende in modo critico da parametri fisici come le perdite di inserzione, i guadagni degli amplificatori e le caratteristiche della fibra. Tuttavia, tali parametri sono spesso incerti o non accessibili agli operatori nelle reti operative. Per superare questa limitazione, progetto e valido quattro tecniche complementari di Input Refinement (IR) che sfruttano dati di monitoraggio in servizio per calibrare i gemelli digitali. In particolare: (i) il Passive IR (PIR) raggiunge un’accuratezza a livello di OMS a partire da un singolo snapshot e si estende ai sistemi C+L; (ii) l’Active IR (AIR) perturba i guadagni degli amplificatori per sfruttare le firme SRS, consentendo il rilevamento e la localizzazione di anomalie a livello di tratta; (iii) l’Incremental IR (IIR) utilizza snapshot multipli per un raffinamento progressivo e supporta l’ottimizzazione autonoma in ciclo chiuso; e (iv) un metodo ibrido IR+PPE integra la stima del profilo di potenza con IR, ottenendo una calibrazione assoluta delle perdite di inserzione e dell’evoluzione longitudinale della potenza. Insieme, questi metodi stabiliscono gemelli digitali accurati, robusti e operativamente validi per la previsione e l’ottimizzazione della QoT. 3. Benchmarking dei transponder per l’interconnessione dei data center (DCI): Oltre alle reti backbone e metropolitane, la rapida espansione dei carichi di lavoro legati al cloud e all’intelligenza artificiale ha reso le interconnessioni tra data center (DCI) un dominio cruciale per le reti ottiche. Gli operatori hyperscale devono bilanciare costo, portata e capacità nella scelta tra diverse famiglie di transponder coerenti. Per fornire linee guida di progettazione, conduco un benchmarking sistematico di transponder a 800G ZR, ZR+ e ad alte prestazioni utilizzando simulazioni che incorporano dinamiche di amplificatori misurate sperimentalmente in reti DCI a corta portata. Lo studio analizza la scalabilità della capacità in condizioni variabili di perdita di tratta, schemi di multiplazione e scenari di espansione dello spettro, identificando i compromessi tra le diverse soluzioni. I risultati evidenziano strategie efficaci di progettazione e caricamento per implementazioni DCI ad alta capacità ed economicamente efficienti. In sintesi, questa tesi dimostra che l’integrazione di modelli analitici di QoT con monitoraggio raffinato, controllo SDN e gemelli digitali consente automazione su larga scala, funzionamento con margini efficienti e linee guida progettuali concrete per le reti ottiche di nuova generazione.
Robust and automatically-driven optical networks
YANG, XIN
2025/2026
Abstract
Optical networks form the backbone of today’s digital infrastructure, transporting the majority of global data traffic across long-haul, metro, and data center domains. As they evolve to accommodate exponential traffic growth and support latency-critical services such as 5G, 6G, and distributed AI, future optical networks face unprecedented demands for capacity, resilience, and automation. At the same time, physical impairments—including amplifier noise, Kerr nonlinearities, stimulated Raman scattering (SRS), filtering penalties, and transponder back-to-back limitations—accumulate along transmission paths and constrain the signal reach through its Quality of Transmission (QoT). While analytical models can capture these effects, their predictive accuracy in live networks is hindered by uncertain parameters and limited observability, forcing operators to adopt conservative margins that reduce the utilization efficiency of network resources such as capacity and power. This thesis aims to advance the design of Robust and Automatically-Driven Optical Networks, and the main contributions can be summarized in three parts as follows. 1. Experimental Quantification and Mitigation of QoT Degradation through Power Re-optimization: A key challenge in optical networks is that the practical "set-and-forget" power setting strategy, widely adopted in today’s operations, ignores cumulative load-dependent effects arising from incremental traffic loading and the absence of power re-optimization for already established channels. As a consequence, the QoT of existing services progressively degrades as additional services are loaded into the network. To address this, I experimentally quantify such QoT degradation in multi-node mesh testbeds and show its impact on network performance. I further propose both static and dynamic power re-optimization strategies integrated into a software-defined networking (SDN) control plane called "AI-Light". Experimental demonstrations confirm that these strategies can substantially recover SNR degradation and pave the way for automated and resilient network operation. 2. Development of Input Refinement (IR) Techniques for Digital Twins: The accuracy of analytical QoT models depends critically on physical parameters such as insertion losses, amplifier gains, and fiber characteristics. However, these parameters are often uncertain or hidden from operators in live networks. To overcome this limitation, I design and validate four complementary Input Refinement (IR) techniques that leverage in-service monitoring data to calibrate digital twins. Specifically: (i) Passive IR (PIR) achieves OMS-level accuracy from a single snapshot and extends to C+L systems; (ii) Active IR (AIR) perturbs amplifier gains to exploit SRS signatures, enabling span-level anomaly detection and localization; (iii) Incremental IR (IIR) leverages multiple snapshots for progressive refinement and supports closed-loop autonomous optimization; and (iv) a Hybrid IR+PPE method integrates power profile estimation with IR, achieving absolute calibration of insertion losses and longitudinal power evolution. Together, these methods establish accurate, robust, and operationally viable digital twins for QoT prediction and optimization. 3. Benchmarking of Transponders for Data Center Interconnect (DCI): Beyond backbone and metro networks, the rapid expansion of cloud and AI workloads has made data center interconnects (DCIs) a critical domain for optical networking. Hyperscale operators must balance cost, reach, and capacity when choosing among different coherent transponder families. To provide design guidelines, I conduct a systematic benchmarking of 800G ZR, ZR+, and high-performance transponders using simulations that incorporate experimentally measured amplifier dynamics in short-reach DCI networks. The study analyzes capacity scaling under varying span losses, multiplexing schemes, and spectrum expansion scenarios, identifying the trade-offs among different solutions. The results highlight effective design and loading strategies for cost-efficient and high-capacity DCI deployments. In summary, this thesis demonstrates that integrating analytical QoT models with refined monitoring, SDN control, and digital twins enables scalable automation, margin-efficient operation, and actionable design guidelines for next-generation optical networks.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Yang.pdf
accessibile in internet per tutti
Descrizione: pdf of thesis
Dimensione
22.07 MB
Formato
Adobe PDF
|
22.07 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/255197