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

Abstract

Descriptors derived from atomic structure and quantum chemical calculations for small molecules representing polymer repeat elements were evaluated for machine learning models to predict the Hildebrand solubility parameters of the corresponding polymers. Since reliable cohesive energy density data and solubility parameters for polymers are difficult to obtain, the experimental heat of vaporization ΔHvap of a set of small molecules was used as a proxy property to evaluate the descriptors. Using the atomistic descriptors, the multilinear regression model showed good accuracy in predicting ΔHvap of the small-molecule set, with a mean absolute error of 2.63 kJ/mol for training and 3.61 kJ/mol for cross-validation. Kernel ridge regression showed similar performance for the small-molecule training set but slightly worse accuracy for the prediction of ΔHvap of molecules representing repeating polymer elements. The Hildebrand solubility parameters of the polymers derived from the atomistic descriptors of the repeating polymer elements showed good correlation with values from the CROW polymer database.

Solvent selection for polymers enabled by generalized chemical fingerprinting and machine learning

“…16 Others have used descriptors derived from atomic structure and quantum chemical calculation for small molecules representing polymer repeat units to predict the Hildebrand solubility parameters using kernel ridge regression and multi-linear regression models. 17 Most recently, Liu et al collected data on 81 polymers and 1221 solvents and created more easily interpretable regression models to predict Chi, Hildebrand, and Hansen parameters. 18 They featurized their polymers and solvents using RDKit generated chemical fingerprints such as the count, density and weighted sum for atoms, and 2nd order features generated from the 3d structures of trimers and solvents such as LUMO, HOMO and heat of formation.…”

Artificial Intelligence for Conjugated Polymers

“…(2) Classification algorithms, including the k-nearest neighbors (KNN) algorithm, decision trees, random forests, logistic regression, and support vector machines, are routinely used to categorize data into classes. They contribute to the prediction of properties like glass transition temperature, 31 solubility parameters, 32 and mechanical properties of conjugated polymers. 33 (3) Neural networks, particularly deep learning architectures like convolutional neural networks (CNNs) and recurrent neural networks (RNNs), possess the ability to capture intricate hierarchical patterns in data.…”

Efficiency Roll-off Suppression in Solution-Processed Organic Light-Emitting Diodes through a Bipolar Host Design

“…The larger voids inside SiDBFDBA would give simultaneously higher possibility for the solvent penetrating to dissolve materials while also providing more space for molecular motion, which lowers the T g , as well (Figure D) . To have good solubility in a solvent, the Hildebrand solubility parameter difference (Δδ) between solvent and material should be approximately less than 2; a higher value of Δδ would lead to poor solubility because of the immiscibility between material and solvent. , All three hosts have good solubility in chlorobenzene with Δδ < 2 (Figure F and Figure S35). Among them, SiDBFDBA has relatively better solubility with the lowest Δδ because of its larger voids and free volume.…”