Toward Deep-SCFT: A Preliminary Study Using Neural Networks

Abstract

The phase states of block copolymer (BCP) melt systems, their dependence on system parameters, and the underlying physics have been central topics in polymer science for decades. Among theoretical and simulation approaches, the field theoreitcal approach based on mean field framework is widely adopted. Within this framework, self-consistent field theory (SCFT)—which seeks fields for composition exchange and pressure that satisfy mutual self-consistency—has become a widely used method for its simplicity and computational efficiency. To obtain self consistent solutions from given initial conditions and parameters, numerical schemes such as simple mixing and Anderson mixing are commonly used. Recently, there have been attempts to leverage neural networks and deep learning, but there has been only limited success. In this talk, we present our recent efforts to incorporate neural networks into SCFT calculation. Training data were generated from conventional SCFT runs, and a network was trained to predict converged solutions. The resulting model reached SCFT solutions with fewer iterations than Anderson mixing and was generalized beyond the training conditions to nearby regions of parameter space.