Accurate prediction of polymer properties is essential for their effective application and devel- opment. However, traditional experimental methods are often prohibitively time-consuming and expensive. This thesis employs an integrated approach to predict various polymer prop- erties, including mechanical, rheological, and thermal characteristics. By combining group interaction modelling, molecular dynamics methods, and machine learning, we predicted 14 distinct physical properties of both homo-polymers and binary copolymers with varying monomer compositions.

Group interaction modelling was utilized to calculate properties such as glass transition temperature, heat capacity, elastic modulus, linear thermal expansion, and Poissons ratio for homo-polymers. The accuracy of this method was found to depend heavily on the precision of input parameters, such as the Debye temperature. A substantial portion of this study employed machine learning techniques to predict polymer properties. For homo-polymers with sufficiently large datasets, we applied random forest algorithms. Molecular descriptors derived from chemical structures were used to train the model. This approach yielded robust predictive performance, with squared correlation coefficient values ranging from 0.83 to 0.955 across six different properties of homo-polymers. Our results demonstrated that machine learning models outperformed group interaction modelling, highlighting the superior reliability of machine learning approaches.

One challenge in extending the machine learning approach was the lack of sufficient datasets. To address this, we employed transfer learning to predict properties of homo-polymers with smaller datasets. A neural network initially trained on heat capacity at constant pressure was adapted to predict four additional properties: specific heat capacity at constant volume, shear modulus, flexural strength, and dynamic viscosity. Despite the small dataset sizes (ranging from 13 to 18 samples), the transfer learning models achieved high accuracy, illustrating the effectiveness of transfer learning in leveraging limited data for diverse property predictions.

Further expanding our research, we explored the domain of copolymers, which present a broad range of chemical compositions and potential applications. By integrating molecular dynamics simulations with machine learning, we predicted seven physical properties of 140 binary copolymers with varying monomer compositions. Using random forest models with molecular descriptors and graph neural networks with graph representations, we found that random forest models excelled in predicting properties such as density and heat capacity at constant pressure and volume, while graph neural networks outperformed in predicting properties like volume thermal expansion, density, and bulk modulus. This dual approach underscores the importance of selecting appropriate molecular representations for accurate property predictions.

Collectively, these studies demonstrate the efficacy of combining machine learning, transfer learning, group interaction modelling, and molecular dynamics to predict polymer properties both accurately and efficiently. This thesis offers a comprehensive framework for accelerating polymer design and application through advanced computational methods, paving the way for future innovations in material science.