Machine Learning-Assisted Performance Prediction for Hollow-Core Fiber Interconnects Under Dynamic Traffic Loads

Abstract

Hollow-core fiber interconnects are emerging as a transformative substrate for low-latency, high-capacity data center and high-performance computing fabrics, yet their performance under dynamic traffic loads remains difficult to predict because optical, thermal, mechanical, and scheduling dynamics interact across multiple timescales. This paper examines machine learning-assisted performance prediction for hollow-core fiber interconnects from a system-level perspective. It argues that accurate prediction cannot be reduced to link-level regression or isolated optical performance monitoring; rather, it requires an integrated architecture that couples physical-layer telemetry, switch and routing state, workload traces, energy telemetry, and governance metadata. The paper analyzes structural trade-offs among latency, capacity, energy efficiency, reliability, and fairness when predictive models are embedded in network control loops. It discusses data acquisition and feature engineering challenges, including sparse failure events, distribution drift, adversarial manipulation, and the need for explainable decisions in critical infrastructure. It also considers deployment constraints, sustainability implications, supply-chain dependencies, and policy questions regarding transparency, accountability, and equitable access. Through cross-domain comparisons with cloud networking, transport software-defined networking, cyber-physical systems, and high-performance computing, the paper shows that hollow-core fiber interconnects are not merely faster pipes but socio-technical infrastructures whose benefits depend on institutional coordination and robust model governance. The conclusion outlines a research agenda for validated digital twins, federated and energy-aware learning, uncertainty quantification, and standards that make machine learning-assisted prediction auditable and operationally trustworthy.