Extensive Databases and Group Contribution QSPRs of Ionic Liquids Properties. 1. Density

A new group contribution (GC) quantitative structure-property relationship (QSPR) for estimating density (ρ) of pure ionic liquids (ILs) as a function of temperature (T) and pressure (p) is developed on the basis of the most comprehensive collection of volumetric data reported so far (in total 41 250 data points, deposited for 2267 ILs from diverse chemical families). The model was established based on a carefully revised, evaluated, and reduced data set, whereas the adopted GC methodology follows the approach proposed previously [Ind. Eng. Chem. Res. 2012, 51, 591−604]. However, a novel approach is proposed to model both temperature and pressure dependence. The idea consist of an independent representation of reference density ρ0 at T0 = 298.15 K and ρ0 = 0.1 MPa and dimensionless correction f(T, P)  ρ­(T, p)/ρ0 for other conditions of temperature and pressure. Three common machine learning algorithms are employed to represent the quantitative structure–property relationship between the studied property end points, GCs, T, and p, namely, multiple linear regression, feed-forward artificial neural network, and least-squares support vector machine. On the basis of detailed statistical analysis of the resulting models, including both internal and external stability checks by means of common statistical procedures such as cross-validation, y-scrambling, and “hold-out” testing, the final model is selected and recommended. An impact of type of cation and anion of the accuracy of calculations is highlighted and discussed. Performance of the new model is finally demonstrated by comparing it with similar methods published recently in the literature.本研究构建了一种全新的分组贡献（Group Contribution, GC）定量结构-性质关系（Quantitative Structure-Property Relationship, QSPR）模型，用于以温度（T）和压力（p）为变量估算纯离子液体（Ionic Liquids, ILs）的密度（ρ），建模基于目前已报道的最全面的体积数据集：共包含41250个数据点，涵盖2267种来自不同化学家族的离子液体。本模型基于经过精心修正、评估与精简的数据集构建，所采用的GC方法沿用了此前提出的方案[Ind. Eng. Chem. Res. 2012, 51, 591−604]。不过本研究提出了一种全新的方法来同时建模温度与压力的依赖性，其核心思路是对参考温度T0=298.15K与参考压力p0=0.1MPa下的参考密度ρ0进行独立表征，并针对其他温压条件定义无量纲校正项$f(T,p) equiv ho(T,p)/ ho_0$。本研究采用三种常见的机器学习算法，构建目标物性端点、GC基团、温度与压力之间的定量结构-性质关系，分别为多元线性回归、前馈人工神经网络以及最小二乘支持向量机。通过对所得模型开展详细统计分析，包括借助交叉验证、y随机打乱（y-scrambling）以及"hold-out"测试等常用统计手段进行内部与外部稳定性验证，最终筛选并推荐了最优模型。研究重点探讨了阳离子与阴离子类型对计算精度的影响。最后通过将本模型与近期文献中发表的同类方法进行对比，验证了其优异性能。