Lightweight Adaptation of Foundation Models for Optical Network Failure Management

Machine Learning is increasingly used for optical network automation, yet current solutions require extensive model engineering and lack generalization across heterogeneous domains. This paper introduces a foundation-model pipeline for optical failure management that addresses this limitation by reusing a single frozen LLM backbone as a universal encoder for optical telemetry, while lightweight adapter modules enable efficient task specialization across domains. Evaluations on both an experimental optical dataset and a large-scale simulated soft-failure classification dataset show that adapters preserve most of the predictive performance of full fine-tuning, achieving an F1 of ≈ 0.885 on the experimental dataset and an accuracy of ≈ 0.91 on the simulated dataset, while reducing computational cost by up to one order of magnitude compared to full fine-tuning. Adapters also remain robust under aggressive compression, maintaining F1 and accuracy within (0.86–0.88) and (0.865–0.945) ranges, respectively, even at reduction factors up to 96×. Few-shot experiments further demonstrate strong data efficiency: with only 64–128 labeled samples, the proposed architecture reaches near-perfect performance in one transfer direction and reliably adapts in the more fine-grained case. These results highlight the viability of Parameter-Efficient Fine-Tuning-enabled foundation models for scalable, cross-domain optical failure management.