Deep learning model for predicting phase diagrams of block copolymers
Takeshi Aoyagi
Abstract
Block copolymers show various microphase-separated structures depending on their chain architecture and the interaction parameters between the different chemical structures of monomeric unit χ. Self-consistent field theory is a powerful tool to predict such phase-separated structures, and phase diagrams of block copolymers have been reported using self-consistent field theory. However, obtaining stable morphology of each polymer structure and the χ parameter requires intensive computational study. We applied deep learning to predict phase diagrams from metastable structures obtained by crude, cost-effective self-consistent field calculation. We used a 3D convolutional neural network for classification of the metastable structures, and a limited number of sets of block copolymer structures and χ parameters with stable phase labels were used for training. After the model was trained using the training set, it successfully assigned the metastable structures of a wide variety of diblock and triblock copolymers to the correct stable phases. This approach is capable of predicting the phase diagrams of block copolymers effectively without intensive self-consistent field calculation.