Atomistic Descriptors for Machine Learning Models of Solubility Parameters for Small Molecules and Polymers

Area of Science:

• Polymer Science
• Computational Chemistry
• Materials Science

Background:

• Accurate prediction of polymer solubility parameters is challenging due to data scarcity.
• Cohesive energy density and solubility parameters for polymers are difficult to obtain experimentally.
• Small molecules can serve as proxies for evaluating predictive models for polymers.

Purpose of the Study:

• To evaluate atomic and quantum chemical descriptors for machine learning models predicting polymer solubility parameters.
• To assess the accuracy of multilinear regression and kernel ridge regression models.
• To correlate predicted polymer solubility parameters with existing databases.

Main Methods:

• Utilized descriptors derived from atomic structure and quantum chemical calculations.
• Employed machine learning models, including multilinear regression and kernel ridge regression.
• Used experimental heat of vaporization (ΔHvap) of small molecules as a proxy property for model evaluation.

Main Results:

• Multilinear regression demonstrated good accuracy in predicting ΔHvap for small molecules (MAE: 2.63 kJ/mol training, 3.61 kJ/mol cross-validation).
• Kernel ridge regression showed comparable performance on the training set but slightly lower accuracy for polymer repeat units.
• Predicted Hildebrand solubility parameters for polymers correlated well with the CROW polymer database.

Conclusions:

• Atomistic descriptors are effective for machine learning models predicting polymer solubility parameters.
• The use of small molecule proxies provides a viable method for descriptor evaluation.
• This computational approach offers a reliable alternative for estimating polymer solubility parameters.