Neural Network-Driven\nExploration in Polymerization-Induced\nSelf-Assembly: From 2D to 3D Pseudo-Phase Diagram

We explore the application of machine learning (ML) to\npredict\nthe morphology of poly(glycerol monomethacrylate)-poly(2-hydroxypropyl\nmethacrylate) (PGMA–PHPMA) diblock copolymer nano-objects prepared\nvia polymerization-induced self-assembly (PISA) in aqueous media.\nTraditional studies typically map copolymer morphology using two-dimensional\n(2D) pseudo-phase diagrams, plotting variables such as the mean degree\nof polymerization (<i>Xn</i>) of the solvophobic block against\nthe copolymer concentration (also known as the solids content). In\ncontrast, our approach utilizes deep neural networks (DNNs) trained\non literature data to generate detailed three-dimensional (3D) morphology\nmaps. These maps include the molecular weight of the solvophilic block,\nproviding a comprehensive volumetric view that reveals more complex\nrelationships and transitional morphologies that are difficult to\ncapture in 2D representations. This 3D modeling approach enriches\nour understanding by highlighting critical thresholds and nuanced\ntransitions within the morphology landscape. Such advanced modeling\nnot only deepens our understanding of how PGMA molecular weight influences\ncopolymer morphology but also significantly reduces the need for extensive\nexperimental trials. Consequently, it simplifies the creation of accurate\npseudo-phase diagrams across a broad range of aqueous PISA formulations.\nExperimental validation confirms the accuracy of our models, demonstrating\nthe potential of ML to make predictive modeling more accessible to\nchemists and paving the way for future research on other PISA formulations.\nThe data set, along with all codes for model training and evaluation,\nis publicly accessible via both Zenodo and GitHub platforms.