LLM-Assisted Design and Analytics of Next-Generation Optical Transport Networks

Optical networks are driving an increasing demand for network automation and intent-driven design, which requires translating high-level operator requests into network actions. In this paper, we present an LLM-based system for design and analytics of optical networks, where LLMs play the key role as reasoning agents, managing network analytics. The proposed framework integrates network queries with domain specific Python modules for topology generation, graph analytics, and equipment-aware referencing. Four open-source LLMs have been evaluated, namely Qwen-2.5-14B-Instruct, LLaMA-3.1-8B, Mistral-NeMo-12B, and Gemma-2-9B, using a benchmark of representative network engineering queries ranging in varying network tasks. Model outputs are assessed using a graded scoring criterion of full correctness, partial correctness, or hallucinations. The results show that the Qwen-2.5-14B-Instruct model achieved the highest overall accuracy of 96.97 % with the lowest hallucination, while smaller models such as LLaMA and Gemma delivered competitive performance. These findings indicate that LLMs have strong potential to reliably support optical network design and analytics in real-world applications.