Data-driven frameworks to robustly predict solubility parameter of diverse polymers.

Area of Science:

• Polymer Science
• Computational Chemistry
• Materials Science

Background:

• Accurate prediction of polymer solubility parameters is crucial for material selection and processing.
• Understanding the complex relationships between polymer properties and solubility is essential for developing new materials.
• Existing methods for determining solubility parameters can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for accurate forecasting of polymer solubility parameters.
• To identify key molecular descriptors influencing polymer solubility.
• To enhance the interpretability and predictive reliability of ML models in polymer science.

Main Methods:

• Utilized a dataset of 1,799 polymer solubility data points, preprocessed with Monte Carlo outlier detection.
• Trained and compared multiple ML algorithms including Linear Regression, SVM, Random Forests, Gradient Boosting Machines, ANNs, and CNNs.
• Evaluated model performance using R-squared, RMSE, MRD%, cross-plots, deviation plots, and SHAP analysis.

Main Results:

• CatBoost, ANN, and CNN models achieved superior accuracy in predicting polymer solubility parameters.
• Sensitivity analysis confirmed that all input features influenced the solubility parameter.
• SHAP analysis identified dielectric constant as the most significant predictor of polymer solubility.

Conclusions:

• ML models, particularly CatBoost, ANN, and CNN, offer an efficient and accurate approach for forecasting polymer solubility parameters.
• Key molecular descriptors, especially dielectric constant, play a vital role in determining polymer solubility.
• The developed models provide valuable insights into structure-property relationships, enhancing scientific understanding and predictive capabilities in polymer science.