Revolutionizing optical networks: The integration and impact of large language models

Abstract

The increasing complexity and scale of optical networks demand advanced automation frameworks capable of adapting to dynamic service requirements, physical-layer impairments, and multi-vendor environments. Traditional solutions—based on static rule sets or narrowly scoped machine learning models—struggle to manage real-time performance, heterogeneous data, and domain-specific variability. Large Language Models (LLMs), built on transformer architectures, offer a paradigm shift by enabling context-aware reasoning, multi-task generalization, and natural language interpretation. These models can automate configuration generation, fault diagnosis, alarm correlation, and routing and spectrum assignment (RSA), while enhancing Quality of Transmission (QoT) estimation and scenario modeling.

This article provides a comprehensive survey of current automation approaches in optical networks, including software-defined networking (SDN), intent-based networking (IBN), machine learning (ML)-based orchestration, and cognitive control architectures. Special attention is given to emerging paradigms that integrate LLMs for intent interpretation, fault analysis, configuration generation, and reasoning.

Building on these foundations, we propose a hybrid framework that integrates LLMs with Digital Twin (DT) technologies to enable closed-loop control, predictive optimization, and explainable, intent-driven decision-making. Telemetry streams feed both DT simulations and LLM-based reasoning agents, supporting proactive reconfiguration and fault mitigation. To address LLM limitations—such as hallucinations and inference latency —the framework incorporates prompt engineering, retrieval-augmented generation (RAG), domain-specific fine-tuning, and simulation-based validation.

The proposed architecture paves the way for resilient, autonomous, and sustainable optical networks that can self-optimize and adapt in real time.