Abstract

The estimation of thermophysical properties of chemical compounds holds considerable importance across a multitude of fields, ranging from material science to process design. Among these properties, the normal boiling point (NBP) stands out as a pivotal parameter in scientific and engineering contexts, as it elucidates the state of a substance under standard conditions commonly encountered in nature. Additionally, NBP is extensively documented in various reference materials and databases. Given its paramount importance and widespread availability, prediction methodologies for other properties such as critical temperature, liquid density, vapor pressure, surface tension, liquid viscosity, liquid thermal conductivity, and flash point often rely on NBP as a foundational metric. Accurate prediction of NBP becomes imperative when experimental data are lacking. Consequently, several prediction techniques have been proposed. These approaches encompass group contribution methodologies, which analyze chemical moieties within molecules, and quantitative structure-property relationships (QSPR), which aim to correlate molecular descriptors with various features of the substance. Although in recent years, Machine Learning (ML) and Deep Learning (DL) techniques have gained traction for predicting molecular properties, determining the optimal strategies for feature selection, feature vectorization, and algorithm choice remains an ongoing challenge. In this study, multiple feature selection methods were explored and evaluated alongside various ML/DL algorithms. Specifically, four techniques were assessed namely functional group moieties, molecular descriptors, SMILES enumeration, and molecular graphs (GNNs). The findings suggest that the utilization of molecular graphs, particularly employing GNNs, yields superior prediction capabilities compared to alternative methods and traditional non-ML approaches, average an average error rate of 2.5% on the tested compounds. Also, the findings of this work will be useful in improving the prediction of solid as well as temperature dependent thermophysical properties. To enhance accessibility, a user-friendly online tool based on GNNs, making this technique readily accessible to the broader scientific community.

Degree

College and Department

Ira A. Fulton College of Engineering; Chemical Engineering

Rights

https://lib.byu.edu/about/copyright/