发表机构
Oak Ridge National Laboratory; University of Michigan; University of Tennessee Knoxville（橡树岭国家实验室; 密歇根大学; 田纳西大学诺克斯维尔分校）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
PACE-SIMS是带AI智能体质量控制的自主SIMS表征工作流，经实验验证可高效完成WOx薄膜研究，可迁移至其他分析技术的破坏性测量场景。
AI 中文摘要
飞行时间二次离子质谱（ToF-SIMS）被广泛应用于各类材料与体系的局部化学研究中。然而，其操作需要耗费大量专家时间：训练有素的研究人员必须全程监督数据采集，在实验过程中动态调整参数，这类工作通常会持续数天的研究周期。在此，我们提出PACE-SIMS，一种将SIMS研究作为人机协作的智能体工作流。研究人员指定科学问题与质量要求后，AI智能体将构建计划，经批准后自主执行，在检查点暂停以评估每次测量结果，并进行修正、重试或升级操作。为验证该方法，我们将其应用于18O富集WOx薄膜的化学成分研究。在这项盲法随机双极性研究（时长8.1小时，共35次测量）中，智能体完成了固定脚本会遗漏的3项非预设修正，且全部4项预测均与密封真实值吻合。该运行还获得了可迁移的测量科学成果，包括成分校准、两种离子极性间5.3%的同位素读数偏移以及沉积的示踪剂输送机制，而研究人员投入的关注时长不足两小时。所开发的智能体架构并非仅适用于SIMS，可应用于其他分析技术，主要目标是优化方法适用性较差的破坏性测量。
英文摘要
Time-of-flight secondary ion mass spectrometry (ToF-SIMS) is widely used for local chemical investigations across a broad range of materials and systems. However, its operation is expensive in expert time: a trained researcher must supervise the acquisition continuously, tuning parameters as the experiment proceeds, often across a campaign spanning multiple days. Here, we present PACE-SIMS (Pause-Assess-Correct-Execute for SIMS), an agentic workflow that runs a SIMS study as a human-AI collaboration. In this workflow, the researcher specifies the scientific questions and quality requirements, and an AI agent builds the plan and, after approval, executes it autonomously, pausing at checkpoints to judge each measurement and to correct, retry, or escalate. To validate the approach, we applied it to a study of chemical composition in 18O-enriched WOx films. During this randomized, blind two-polarity study (8.1 hours, 35 measurements), the agent applied three rule-governed corrections and diagnosed an unanticipated source excursion without requiring event-specific control logic to be programmed in advance. The same run returned transferable measurement science, including a composition calibration and the deposition's tracer-delivery mechanism, from less than two hours of researcher attention. The developed agentic architecture is not specific to SIMS and can be applied to other analytical techniques, with the primary target being destructive measurements, for which optimization-based methods are poorly suited.
