Diagnosing temperature-induced bias and dispersion in ATR-MIR protein measurement using spectral-window PLS models

This dataset supports the article “Temperature-response diagnostics for spectral-window ATR-MIR PLS protein prediction in low-solids skim milk”. The study tested the hypothesis that sample temperature acts as a structured non-analyte measurement-state variable that affects ATR-MIR PLS protein predictions differently across spectral windows, and that internal residual-temperature correction can improve benchmark agreement without necessarily reducing prediction dispersion. The active source system was one low-solids reconstituted bovine skim-milk formulation prepared from one skim-milk powder batch. Eight calibration formulations were prepared gravimetrically to build and internally validate seven one-latent-variable spectral-window PLS models: Full MIR, PhosCov, Amide I, Amide II, Amide III, Amide ALL and Amide I_II_III. The matched low-solids formulation was calibration sample S2W, with Rpw = 5.06% and Kjeldahl protein = 1.610%. Low-solids ATR-MIR spectra were collected during one passive warming sequence at 2, 5, 8, 11, 14, 17 and 20 °C, plus fixed refrigerated and ambient technical scans. The repository contains calibration formulation data, Kjeldahl/reference information, raw and processed ATR-MIR spectral files, PLS model metadata, leave-one-formulation-out cross-validation outputs, low-solids warming and fixed-temperature prediction tables, diagnostic summaries, leave-one-temperature-out residual-correction outputs, MATLAB analysis scripts, final article tables, figure-source files, supplementary tables and figures, and the laboratory report used to support the reference benchmark. The data show that temperature-response behaviour was spectral-window dependent. PhosCov had the lowest RMSECV and warming-sequence dispersion, Full MIR had the lowest RMSE relative to the 1.610% benchmark, and Amide I/II were the most temperature-sensitive windows. Leave-one-temperature-out linear residual correction reduced PhosCov RMSE and mean bias but did not reduce dispersion. The files should be interpreted as a controlled chemometric source-data envelope for temperature-aware ATR-MIR calibration diagnostics, not as external validation across batches, matrices, instruments or process environments.本数据集支撑论文《低固形物脱脂牛奶中基于光谱窗口衰减全反射-中红外光谱（ATR-MIR）偏最小二乘（PLS）的蛋白质预测温度响应诊断》。该研究验证了如下假设：样品温度作为结构化非分析物测量状态变量，在不同光谱窗口下对ATR-MIR PLS蛋白质预测的影响存在差异，且内部残余温度校正可提升基准一致性，同时未必降低预测离散度。 本次研究的活性源体系为一批以脱脂乳粉制备的低固形物复原牛脱脂乳配方。通过重量法制备8份校准配方，用于构建并内部验证7个单潜变量光谱窗口PLS模型：全中红外（Full MIR）、PhosCov、酰胺I区（Amide I）、酰胺II区（Amide II）、酰胺III区（Amide III）、全酰胺区（Amide ALL）以及酰胺I_II_III组合区。匹配的低固形物配方为校准样本S2W，其Rpw=5.06%，凯氏定氮蛋白质含量为1.610%。在2、5、8、11、14、17与20℃的单次被动升温过程中采集低固形物ATR-MIR光谱，同时设置固定冷藏与室温环境的技术扫描样本。 本数据集仓库包含校准配方数据、凯氏定氮/参考信息、原始与预处理ATR-MIR光谱文件、PLS模型元数据、留一配方交叉验证输出结果、低固形物升温与固定温度预测表格、诊断汇总结果、留一温度残余校正输出结果、MATLAB分析脚本、论文最终表格、图表源文件、补充表格与图表，以及用于支撑参考基准的实验室报告。 数据表明，温度响应行为具有光谱窗口依赖性：PhosCov的交叉验证均方根误差（RMSECV）与升温序列离散度最低，全中红外模型相对于1.610%基准的均方根误差（RMSE）最低，而酰胺I/II区是温度敏感性最高的光谱窗口。留一温度线性残余校正可降低PhosCov的RMSE与平均偏差，但未能降低离散度。本数据集文件应被视为面向温度感知ATR-MIR校准诊断的受控化学计量学源数据集合，而非适用于不同批次、基质、仪器或工艺环境的外部验证数据集。