Exploring the Strengths and Limitations of Polymer Chemistry Informed Neural Networks

ABSTRACT Kinetic models are essential tools for providing a fundamental understanding of polymerization processes. However, they are constrained by the current level of understanding of the system under study and often rely on numerous unknown parameters that are difficult to measure experimentally or estimate reliably. To address these challenges, polymer chemistry‐informed neural networks (PCINNs) have been developed in which the domain knowledge embedded in these kinetic models is combined with data‐driven machine learning tools. Herein, we investigate the robustness and limitations of this methodology through an in‐silico investigation in which an imperfect kinetic model is combined with small datasets. As a case study, high‐temperature solution polymerization of methyl methacrylate was considered. It is demonstrated that PCINNs are able to utilize imperfect kinetic models in order to reach high levels of predictive performance, with reliable extrapolation at reaction temperatures significantly beyond the range of the original dataset. Integrating domain knowledge was found to substantially reduce variance in the outcomes of the final model when using a relatively small dataset. Finally, it is shown that PCINNs can be used to provide insights into the causes of poor predictive performance of kinetic models, offering the ability to use data‐driven learning to inform first‐principles approaches.