LLM-Orchestrated Digital Twin for Smart Optical Network Operations

The growing complexity of optical network planning demands automated and accessible solutions beyond expert-driven workflows. Although digital twins enable accurate physical-layer modeling and optimization, their use is limited by technical interfaces. This paper proposes a large language model (LLM)-orchestrated framework that enables intent-driven interaction with optical network digital twins (DTs) using natural language. A locally deployed, non-fine-tuned Mistral-7B model serves as a reasoning and orchestration layer, translating user queries into structured and validated digital twin operations through prompt-guided tool selection. A lightweight agent ensures parameter consistency before execution, enabling reliable automation of tasks such as topology generation, routing, and quality-of-transmission (QoT) analysis via deterministic backend computations. The framework is evaluated on a domain-specific benchmark covering multi-step network operations. It achieves a Domain-Grounded Accuracy of 89.22\% with average latency of 15.11 seconds, while maintaining low computational overhead and minimal hallucination. These results demonstrate that compact, locally deployed LLMs, combined with structured tool integration, provide a practical alternative to fine-tuned or cloud-based solutions for smart optical network operations.