Learning glass transition temperatures via dimensionality reduction with data from computer simulations: Polymers as the pilot case

• DOI: 10.1063/5.0229161
• Publication Date: 2024-11-08T10:41:58Z
• Authors: Artem Glova; Mikko Karttunen
• Keywords: Dihedral angle

Machine learning methods provide an advanced means for understanding inherent patterns within large and complex datasets. Here, we employ the principal component analysis (PCA) diffusion map (DM) techniques to evaluate glass transition temperature (Tg) from low-dimensional representations of all-atom molecular dynamic simulations polylactide (PLA) poly(3-hydroxybutyrate) (PHB). Four descriptors were considered: radial distribution functions (RDFs), mean square displacements (MSDs), relative (RSDs), dihedral angles (DAs). By applying Gaussian Mixture Models (GMMs) analyze PCA DM projections by quantifying their log-likelihoods as a density-based metric, distinct separation into two populations corresponding melt states was revealed. This enabled Tg evaluation cooling-induced sharp increase in overlap between log-likelihood distributions at different temperatures. values derived RDF MSD using closely matched standard computer simulation-based dilatometric both PLA PHB models. not case PCA. The DM-transformed DA RSD data resulted agreement with experimental ones. Overall, fusion atomistic DMs complemented GMMs presents promising framework computing studying unified way across various glass-forming materials.