Nucleation patterns of polymer crystals analyzed by machine learning models

The understanding of polymer crystallization is a longstanding challenge within the field of polymer science. We employ machine learning (ML) algorithms to analyze conformation patterns during nucleation of early crystallites in undercooled melts based on molecular dynamics simulation data. Our focus is on data-driven techniques that establish decision boundaries to detect the crystalline state of individual monomers without prior knowledge instead of 'classical' methods that manually position these boundaries [C. Luo and J.-U. Sommer, Macromolecules 44 (2011), 1523]. In particular, we utilize self-supervised auto-encoders to compress the local fingerprint information and apply a Gaussian mixture model to distinguish between ordered and disordered states. The high specificity of the method allows us to uncover the intricate temporal patterns related to crystalline order, even before any clear indications of the transition became evident in thermodynamic properties, such as specific volume. We identify a pre-transition point characterized by the highest crystallization efficiency, determined by the fraction of monomers preserved in the crystalline phase as compared to those entering that phase.