基于LLM的核燃料后处理多模态知识库构建OA

With the rapid growth of scientific literature and engineering data,knowledge extraction and struc-tured modeling from multimodal data have become key challenges for intelligent engineering transformation.In recent years,large language model(LLM)have demonstrated strong capabilities in cross-modal under-standing and knowledge modeling,providing new technical pathways for multimodal knowledge base con-struction.However,in domains characterized by high specialization,dense terminology,and strict safety re-quirements,general-purpose LLM still face multiple challenges such as hallucinated outputs,knowledge dis-tortion and deployment constraints.To enable effective application of LLM in spent nuclear fuel reprocess-ing,we first review the development of LLM and their core techniques,and analyzes their roles in profes-sional knowledge base construction.We examine the representative studies of LLM in three domains,say,symbolic regression,chemical materials and healthcare,to summarize a technical framework and key techni-cal components of LLM-driven multimodal knowledge base construction,including structured knowledge ex-traction,knowledge representation and indexing,retrieval-augmented generation and reliability control.Then we focus on the knowledge base construction of spent nuclear fuel reprocessing,which is subject to strict safety constraints.An engineering validation is conducted on the structured extraction and knowledge inges-tion stages.A case study on equation-reaction structured extraction is presented to demonstrate the transfer-ability and practical feasibility of the obtained technical framework.Finally,current challenges in this domain are discussed,and future directions are outlined in terms of safe and controllable deployment,few-shot adap-tation,multimodal fusion,and expert knowledge integration,providing technical references for enabling LLM-driven intelligence in the nuclear industry.