虚假关系
傅里叶变换红外光谱
渗透
聚合物
谱线
膜
人工智能
管道(软件)
计算机科学
磁导率
生物系统
巴勒
傅里叶变换
化学
模式识别(心理学)
分析化学(期刊)
机器学习
特征(语言学)
配对
材料科学
合成膜
比例(比率)
算法
吸收(声学)
特征选择
吸收光谱法
作者
Mehryar Jafari,Khashayar Salehi,Bernardo Castro-Dominguez
标识
DOI:10.1016/j.memsci.2026.126184
摘要
Gas-separation polymer membranes are typically screened through expensive permeation experiments or molecular-descriptor-based machine learning that relies on idealized repeat-unit representations unable to capture batch variability, additives, or processing history. We demonstrate that FTIR-derived spectral descriptors, combined with lightweight polymer-family metadata, enable accurate prediction of gas permeability, outperforming all structure-based baselines tested. A curated dataset of 843 FTIR spectral records from 128 polymers was assembled, pairing mid-infrared spectra with experimentally reported permeabilities for six industrially relevant gases (He, H 2 , CO 2 , O 2 , N 2 and CH 4 ). A CatBoost gradient-boosted regressor is trained on principal-component-analysis-reduced spectral features (95% retained variance), one-hot polymer-family descriptors and log 10 -transformed targets under 5-fold group cross-validation, in which all spectra of a given polymer are constrained to the same fold to prevent near-duplicate spectral leakage. A leverage-based Williams-plot diagnostic defines the model applicability domain; within this domain, the model achieves R 2 = 0.94 in log 10 -space and R 2 = 0.90 on the original Barrer scale for CO 2 , with a median error factor of 1.2. The same pipeline applied to the remaining five gases gives consistently high log-space accuracy, while original-scale performance varies with the width of each gas's permeability distribution, a pattern explained by the underlying target statistics. Back-projected CatBoost feature importance identifies three chemistry-grounded predictive bands: C–O/C–O–C stretching, a CO 2 -related combination/atmospheric absorption region, and aliphatic C–H stretching, confirming that the model has learned genuine structure-property relationships rather than spurious correlations. Because FTIR is acquired in minutes on routinely available instruments, this framework enables rapid, instrument-light screening of polymer membranes before any permeation experiment is run .
科研通智能强力驱动
Strongly Powered by AbleSci AI