Abstract:

To evaluate the potential of large language models (LLM) in assisting the modeling and prediction of the pulsed column extraction process for spent fuel reprocessing, this study designed five case studies of progressively increasing complexity to assess four mainstream LLM. Model outputs were obtained through structured prompt engineering and quantitatively scored across multiple dimensions via a double-blind evaluation by three domain experts. The results indicate that while all LLM demonstrated excellent performance in instruction adherence, their capabilities declined significantly when addressing diagnostic problems involving complex physicochemical coupling or ambiguous information, lacking the analytical depth to diagnose critical engineering issues precisely. The study concludes that the most suitable role for current LLM is as “intelligent research assistants” to domain experts, jointly forming a highly efficient “human-computer collaborative” research paradigm, rather than acting as independent decision-makers. This paradigm can reduce days of modeling preparation to half an hour. However, all model outputs must undergo rigorous expert review and revision to avoid potential risks such as factual illusions.

Key words: large language model, spent fuel reprocessing, solvent extraction, mathematical modeling, computer simulation

摘要：

为评估大语言模型（LLM）在辅助乏燃料后处理过程脉冲萃取柱萃取过程中建模与预测的潜力，选取4种主流LLM，设计了5个复杂度递进的萃取过程案例，通过结构化提示工程获取模型输出，并采用3位领域专家双盲评估的方式多维度对结果进行量化评分。结果表明，所有LLM在遵循指令方面表现优异，但在处理涉及复杂物理化学耦合或信息模糊的诊断性问题时所有模型的表现均显著下降，分析深度不足，未能精准剖析工程症结。结论认为，LLM当前最合适的定位是作为领域专家的“智能研究助理”，共同构成一种高效的“人机协同”科研新范式，而非独立的决策者。该范式可将数日的建模准备工作缩短至半小时内，但所有模型输出均需经过专家严格审查与修正，以规避“事实性幻觉”等潜在风险。

关键词: 大语言模型, 乏燃料后处理, 溶剂萃取, 数学模拟, 计算机模拟