A LoRA fine-tuning and retrieval-augmented generation-based method for assessing submarine survivability under damage-induced flooding

Objective To address the lack of domain-specific knowledge and the "hallucination" problem encountered when general-purpose large language models (LLM) are applied to submarine damage control under damage-induced flooding conditions, this paper proposes a large language model adaptation method for submarine survivability assessment based on low-rank adaptation (LoRA) and retrieval-augmented generation (RAG).

Methods Based on authoritative sources, including domain-specific knowledge, academic monographs, technical reports, and national standards related to submarine damage control under hull-breach flooding conditions, a domain-specific fine-tuning dataset comprising 10 089 samples was developed, together with a modular knowledge base designed to support individual survivability-index calculation tasks. LoRA was employed to align the base model with domain-specific knowledge, thereby enhancing its understanding of damage-control terminology, flooding scenarios, and the task logic underlying survivability assessment. Meanwhile, RAG was integrated to retrieve information relevant to the current calculation task, including applicable operating conditions, parameter definitions, and calculation formulas, thereby providing the model with the necessary knowledge support for survivability-index calculations.

Results In the knowledge question-answering task, the LoRA-fine-tuned large language model for submarine survivability assessment under damage-induced flooding achieved higher scores than the baseline model on both the bilingual evaluation understudy (BLEU) and recall-oriented understudy for gisting evaluation (ROUGE) metrics. In the submarine survivability-index calculation task under damage-induced flooding, the accuracy of the model’s calculation results increased from 33.33% for the baseline model to 93.33%, while the compliance rate of the calculation process increased from 13.33% to 96.67%.

Conclusion The proposed large language model for submarine survivability assessment under damage-induced flooding can effectively enhance the knowledge understanding and computational reasoning capabilities of general-purpose large language models in submarine damage control under flooding conditions. It provides standardized and traceable decision-making data to support emergency response in submarine damage-control scenarios.