发表机构
National Institute for Materials Science; RIKEN KEIKI Co., Ltd（物质材料研究机构; 理化学仪器有限公司）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究针对光电子产额谱分析的异方差问题，提出基于1/n-Scan方法与集成残差度量的自主自诊断框架，可检测材料电子态分析中的模型失效，为闭环自主材料探索提供支撑。
AI 中文摘要
光电子产额谱（PYS）被广泛用于评估材料的电子态。随着自动化材料发现的推进，从环境大气下的PYS数据中无监督提取物理信息变得重要。传统的固定指数分析和对数变换存在异方差性，这会在低信号区域破坏估计的稳定性。为解决该问题，我们提出一种基于1/n-Scan方法的数据驱动分析框架，该框架直接在原始信号空间中运行，消除了对数变换方法固有的几何偏差。我们进一步集成了自诊断质量评估系统，该系统用赤池权重量化估计不确定性，同时结合独立的残差度量——归一化平均绝对误差（NMAE）、均方根误差与平均绝对误差之比（RMR）、杜宾-沃森（DW）统计量，以及宏观度量（ΔR²）——以区分硬件相关的数据退化与物理模型不匹配。将该框架应用于大气中不同掺杂的硅（Si）和多晶金（Au）参比样品，我们在不假设发射机制的情况下，自主检测到重掺杂p型Si中单组分近似的失效，该失效源于两个具有不同阈值的发射组分的重叠，表现为统计异常：DW降至临界值以下，同时ΔR²辅助升高，尽管NMAE和RMR表现良好，这一结果通过双组分拟合得到独立验证（赤池信息准则差ΔAIC≈57，DW从0.8恢复至2.0）。对于重掺杂n型Si，该框架将其逐渐的表面演化归类为单组分描述内的变化。该框架为闭环自主材料探索提供了一种鲁棒的自诊断分析引擎。
英文摘要
Photoemission yield spectroscopy (PYS) is widely used for evaluating the electronic states of materials. As automated materials discovery advances, unsupervised extraction of physical information from ambient-air PYS data becomes important. Conventional fixed-exponent analyses and logarithmic transformations suffer from heteroscedasticity, which destabilizes estimation in low-signal regions. To address this, we propose a data-driven analysis framework based on the 1/n-Scan method, which operates directly in the original signal space, removing the geometric bias inherent in log-transform approaches. We further integrate a self-diagnostic quality-evaluation system that quantifies estimation uncertainty with Akaike weights, together with independent residual metrics--the normalized mean absolute error (NMAE), the RMSE-to-MAE ratio (RMR), the Durbin-Watson (DW) statistic, and a macroscopic metric ($ΔR^2$)--that distinguish hardware-related data degradation from a physical-model mismatch. Applying the framework to differently doped Si and a polycrystalline Au reference in air, we demonstrate autonomous detection, without assumptions on the emission mechanism, of the breakdown of the single-component approximation in heavily doped p-type Si, arising from the overlap of two emission components with different thresholds, as a statistical anomaly--a decrease in DW below its critical value with an auxiliary increase in $ΔR^2$, despite sound NMAE and RMR--independently confirmed by a two-component fit ($Δ$AIC $\approx$ 57, DW recovering from 0.8 to 2.0). For heavily doped n-type Si, the gradual surface evolution was classified as a change within the single-component description. This framework provides a robust, self-diagnosing analysis engine for closed-loop autonomous materials exploration.
Comments30 pages, 4 figures; supplementary material (28 pages) is provided as an ancillary file. Code and sample data: https://github.com/s-yagyu/n-pys
