Context Aromatic hydrocarbons such as benzene, toluene, and ethylbenzene are extensively used as solvents in coatings, resin, and artificial leather industries. Azeotropic mixtures involving these compounds are commonly encountered in chemical manufacturing, where accurate azeotropic temperature and composition are essential for designing and optimizing separation processes such as extractive and pressure-swing distillation. In this study, two quantitative structure–property relationship (QSPR) models were developed to predict the azeotropic temperature and composition of binary mixtures containing aromatic hydrocarbons using only molecular structural information. The models show excellent agreement with experimental data (R2 = 0.9454 and 0.9448, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">${R}_{adj}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>R</mi> <mrow> <mi mathvariant="italic">adj</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation> = 0.9400 and 0.9413). Internal validation via leave-one-out cross-validation yields <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${R}_{cv}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>R</mi> <mrow> <mi mathvariant="italic">cv</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation> = 0.9308 and 0.9364, while external validation using an independent test set yields <InlineEquation ID="IEq3"> <EquationSource Format="TEX">${Q}_{ext}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>Q</mi> <mrow> <mi mathvariant="italic">ext</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation> = 0.8939 and 0.9364, indicating strong robustness and superior predictive performance compared to previously reported models. Methods Molecular geometries were optimized using HyperChem 8.0, employing MM + and PM3 methods. Molecular descriptors were calculated using the Online Chemical Modeling Environment (OCHEM). Binary mixture descriptors were derived from pure-component descriptors via Kay’s mixing rule. The genetic function approximation (GFA) algorithm was used to select the most relevant descriptors, and predictive models were constructed using multiple linear regression (MLR). Model robustness and predictive capacity were evaluated using leave-one-out cross-validation and an external test set, with applicability domains assessed via Williams plots. All computational procedures and modeling analyses were performed using OCHEM, SPSS, and HyperChem 8.0.

High-accuracy QSPR models for azeotropic property prediction of binary aromatic hydrocarbon mixtures: a genetic function approximation approach

期刊论文
发表日期：

2026年5月13日

查看全文

Liping Lv,
Xingyan Zeng,
Lin Han,
Xiaogang Guo,
Shirui Sun,
Huisheng Huang,
Xiaohong Wu,
Deng Tang

摘要

Context

Aromatic hydrocarbons such as benzene, toluene, and ethylbenzene are extensively used as solvents in coatings, resin, and artificial leather industries. Azeotropic mixtures involving these compounds are commonly encountered in chemical manufacturing, where accurate azeotropic temperature and composition are essential for designing and optimizing separation processes such as extractive and pressure-swing distillation. In this study, two quantitative structure–property relationship (QSPR) models were developed to predict the azeotropic temperature and composition of binary mixtures containing aromatic hydrocarbons using only molecular structural information. The models show excellent agreement with experimental data (R² = 0.9454 and 0.9448, ${R}_{adj}^{2}$ $R_{adj}^{2}$ = 0.9400 and 0.9413). Internal validation via leave-one-out cross-validation yields ${R}_{cv}^{2}$ $R_{cv}^{2}$ = 0.9308 and 0.9364, while external validation using an independent test set yields ${Q}_{ext}^{2}$ $Q_{ext}^{2}$ = 0.8939 and 0.9364, indicating strong robustness and superior predictive performance compared to previously reported models.

Methods

Molecular geometries were optimized using HyperChem 8.0, employing MM + and PM3 methods. Molecular descriptors were calculated using the Online Chemical Modeling Environment (OCHEM). Binary mixture descriptors were derived from pure-component descriptors via Kay’s mixing rule. The genetic function approximation (GFA) algorithm was used to select the most relevant descriptors, and predictive models were constructed using multiple linear regression (MLR). Model robustness and predictive capacity were evaluated using leave-one-out cross-validation and an external test set, with applicability domains assessed via Williams plots. All computational procedures and modeling analyses were performed using OCHEM, SPSS, and HyperChem 8.0.