Development and application of machine learning models for the prediction of critical properties of dynamically crosslinked polymers

The ongoing environmental crisis is driving scientists to develop "greener" polymers, either by adopting more sustainable materials or alternative chemistries.This effort led to the discovery of a fascinating group of dynamically crosslinked polymers, called vitrimers.Vitrimers bridge the gap between traditional thermosets and thermoplastics, combining excellent mechanical abilities with recyclability.To accelerate vitrimer research while minimizing experimental cost, this thesis explores the use of machine learning to predict four key properties: crosslink density, glass transition and topology freezing temperature and gel fraction.This was achieved through processing manually collected data from literature along their respective extracted chemical descriptors.Despite several difficulties during database curation, an experimental data set comprising of roughly 200 observations was created.Using SMILES notation for small molecules and PSMILES, a variation created for polymer repeating units, a variety of descriptors was extracted leveraging a popular cheminformatics package.After rigorous preprocessing-including feature filtering and elimination, handling of missing values, and normalization to achieve a standard value range-two dimensionality reduction methods were applied.First up was Principal Component Analysis (PCA) which involves combining extracted molecular descriptors and features into principal components, dramatically reducing model's inputs.The second hybrid methodology involved combines Pearson correlation analysis with feature selection algorithms like Mutual Information and Random Forest (CA/FS).Model training was done after applying a 5-fold cross validation split into training and test sets, and validation was carried out through metrics like R 2 , root mean squared error and mean absolute error highlighted the impact of both methodologies on results.To further investigate exhibited behavior, plots including parity plots, box plots and error distribution plots were created.More specifically, in Tg prediction Gradient Boosting and Random Forest models achieved cross-validated R² values of 0.85, confirming the effectiveness of ensemble-based methods.This also applied in the case of Tv where Random Forest reached an R² of 0.70.PCA performed poorly because it only captures linear relationships well, which was not the case for Tg and Tv.Contrastingly, PCA yielded better results when used to predict crosslink density, with the Random Forest model achieving an R² of 0.77, suggesting it captures broader structural trends.Prediction of gel fraction was unsuccessful, (R² < 0.40), indicating descriptors may not capture relevant information or the target is too variable.Finally, hyperparameter tuning, while an important step in machine learning pipelines, did not improve on results massively, except for the case of the Gaussian Process model whose R 2 values improved from -54.81 to 0.83 when predicting Tg.These findings demonstrate the promise and limitations of ML approaches in polymer informatics for emerging materials like vitrimers and highlight a way forward.