MACHINE LEARNING METHODS AND MODELS FOR POLYMERIC SYSTEMS

Artificial intelligence (AI) and Machine learning (ML), a subfield of AI, are important tools for the development of data-driven predictive models for a variety of applications. The use of ML models is extended to many different research areas, such as pharmaceutical research, material design and engineering, and process control. Along these lines, the inherent complexity of polymeric systems and the need for the development of new task-specific polymeric materials necessitate the use of modeling and computational tools to accelerate materials discovery. In this thesis, ML techniques are applied to different polymer systems to predict gas sorption capacity, glass transition temperature, and monomers reactivity ratios. Herein first ML models were developed to predict the CO2 uptake of Ionogels. Due to the complex nature and the intrinsic non-ideality of these materials, information about the interactions between Ionogels constituents, polymer, ionic liquid (IL), and CO2 are largely unknown. There are many well-known simulation platforms that have been developed for gas sorption prediction; however, extracting accurate polymer properties for the simulations still presents many obstacles. Therefore, finding the correct method to predict the gas sorption accurately is important. In this study, the Multiple-Layer Perception (MLP) is used to predict the CO2 uptake of Ionogels. Without requiring a large dataset of complex polymer properties, the MLP model showed a high prediction accuracy up to 96 % and 94 % for testing and external dataset. Moreover, compared to commonly used simulation models, the MLP needs a significantly low computing time of ~400 µs to achieve this result. Therefore, the MLP has a potential of substituting the traditional simulations to improve the gas sorption prediction for future research. In addition, ML models were developed to predict the glass transition temperature of polymers. Among the different properties, glass transition temperature is a key property in determining the specific application of a polymer and selection of the processing conditions. Image- and graph-based models are the two alternative state-of-the-art ML models to accurately predict chemical properties. In this study, the use of modifying an original limited dataset and investigating the learning performance of both Graph Attention Network (GAN) and Convolutional Neural Network (CNN) were examined. Because learning with graph structural data with different fingerprints requires effective representation of their graph structure, GAN can achieve a high prediction accuracy of 86% compared to CNN, where a prediction accuracy is 48%. It was concluded that GAN outperforms CNN in different categories, such as computing efficiency, size of dataset, and predictive capability. Hence, with a task of predicting chemical properties with structural coordinate systems, the graph-based models tend to learn the structures better, opening a new direction for future research in improving the property prediction and accelerating the chemistry discovery. Lastly, ML models for polymerization reaction engineering, specifically for the acrylate-based polymers, commonly employed in electronics applications, such as anti-reflective coatings and photoresists were developed. Kinetic modeling is central for developing optimal reaction conditions. For the copolymers, the reactivity ratio plays an essential role in determining several crucial properties of the material, such as molecular weight (MW). Therefore, determining the reactivity ratio accurately is critical in modeling polymerization processes, product design, and manufacturing scale-up. Considering the potential of the graph-based models, GAN is utilized in this study. Along with GAN, a supervised learning approach, K-fold cross-validation, and Bayesian optimization method were applied to optimize the predictive capability of the model and to predict a pair of reactivity ratios. As a result, we can achieve a high testing prediction performance of 85%. With the model’s superior performance, we can rely on its capability to predict the reactivity ratio pair of an unknown copolymer. Advisor: Mona Bavaria