Generative block polymer phase discovery

Block polymers spontaneously self-assemble into mesoscale phases, driven by a thermodynamic balance between chain stretching and interfacial tension. Self-consistent field theory (SCFT) can accurately predict known phases but suffers from a fundamental problem for phase discovery: how can one perform the calculations without knowing the structure in advance? We overcame this limitation by training a deep convolutional generative adversarial network that learns from existing phases to suggest new initial guesses for SCFT calculations. Applied to linear diblock copolymer melts, this approach not only reproduced known network phases, but also uncovered hundreds of possible new phases. Overall, this integrated approach opens up new possibilities for discovering new phases in block polymers.