Neural network based quantitative structural property relations (QSPRs) for predicting boiling points of aliphatic...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Physical Organic Chemistry

Background:

• Accurate prediction of boiling points for aliphatic hydrocarbons is crucial for chemical process design and safety.
• Traditional quantitative structure-property relationship (QSPR) methods often struggle with complex molecular structures and isomer differentiation.
• Neural networks offer a powerful approach for modeling complex relationships between molecular structure and physical properties.

Purpose of the Study:

• To develop and compare quantitative structure-property relationships (QSPRs) for predicting the boiling points of aliphatic hydrocarbons.
• To evaluate the performance of back-propagation neural networks and a modified Fuzzy ARTMAP architecture for this task.
• To identify key molecular descriptors that influence hydrocarbon boiling points, including stereoisomers.

Main Methods:

• Employed molecular connectivity indices, Kappa shape indices, dipole moment, and molecular weight as descriptors.
• Utilized a back-propagation neural network (7-4-1 architecture for alkanes, 7-10-1 for alkenes, 7-9- for composite) and a modified Fuzzy ARTMAP network.
• Validated models using test, validation, and overall datasets, calculating average absolute errors in Kelvin and percentage.

Main Results:

• The modified Fuzzy ARTMAP network achieved significantly lower average absolute errors compared to the back-propagation models for alkanes, alkenes, and composite aliphatic hydrocarbons.
• For alkanes, Fuzzy ARTMAP reduced errors to 0.31% (overall), and for alkenes, to 0.25% (overall).
• The developed QSPRs demonstrated accuracy within experimental error ranges, outperforming previously reported methods.

Conclusions:

• Modified Fuzzy ARTMAP provides a highly accurate and robust method for QSPR modeling of aliphatic hydrocarbon boiling points.
• The selected molecular descriptors, particularly dipole moment, are effective in capturing structural nuances relevant to boiling point prediction, including isomer differentiation.
• These QSPRs offer a valuable tool for predicting boiling points with high precision, surpassing existing regression and neural network approaches.