Quinones and hydroquinones are small organic molecules with numerous applications: battery electrolytes, pharmaceuticals, and sensors, to name a few. An understanding of their fundamental properties, such as melting points, is essential to incorporating these compounds into relevant technologies. In this study, two different approaches were investigated to predict the melting points of quinone- and hydroquinone-based molecules. In one approach, molecular features were calculated with the Mordred molecular descriptor calculator and used to train a ridge regression and a random forest machine learning (ML) model. In the other, a simpler featurization that captures key enthalpic and entropic descriptors was applied in three model variants: a thermodynamic model, a ridge regression model, and a random forest model. On the basis of the average absolute error (AAE), the Mordred-calculated features in the ridge regression model outperformed the thermodynamic features across all models for the quinone data set (AAE: 29.5 °C), but the thermodynamic features in the thermodynamic model resulted in lower prediction errors for the hydroquinone data set (AAE: 35.7 °C). These results emphasize the importance of including intermolecular interaction descriptors, especially for classes of molecules in which these interactions are expected to be strong (e.g., hydrogen bonding in hydroquinones). As a byproduct of this study, we also consolidate (from previously published data) four new data sets with quinone and hydroquinone melting points and other key features.

Quinones and hydroquinones are small organic molecules with numerous applications: battery electrolytes, pharmaceuticals, sensors, to name a few. An understanding of their fundamental properties, such as melting points, is essential to incorporate these compounds into relevant technologies. In this study, two different approaches were investigated to predict the melting points of quinone and hydroquinone-based molecules. In the first approach, molecular features were calculated with the Mordred molecular descriptor calculator and used to train a ridge regression and a random forest machine learning (ML) model. In the second, a simpler featurization that captures key enthalpic and entropic descriptors was applied in three model variants - a thermodynamic model, a ridge regression model, and a random forest model. The Mordred-calculated features in the ridge regression model outperformed the thermodynamic features across all models for the quinone dataset (average absolute errors (AAE): 29.5 °C), but the thermodynamic features in the thermodynamic model resulted in lower prediction errors for the hydroquinone dataset (AAE: 35.4 °C). These results emphasize the importance of including intermolecular interaction descriptors, especially for classes of molecules in which these interactions are expected to be strong (e.g. hydrogen bonding in hydroquinones). As a byproduct of this study, we also consolidate (from previously published data) four new datasets with quinone and hydroquinone melting points and other key features.