Accelerating self-consistent field theoretic simulations for disordered systems with deep learning

Abstract

Polymer science holds a pivotal role in areas such as advanced materials design, drug delivery systems, and biological systems, where being able to efficiently predict polymer thermodynamics and self-assembly is crucial. Self-consistent field theory (SCFT) offers a theoretical framework with many successful predictions that have guided experiments. However, there are classes of systems that are challenging to describe with SCFT due to their computational expense, such as anisotropic systems and worm-like chain models. In this study, we take a first step toward alleviating these challenges by developing a machine-learning approach that can predict density fields directly from the potential fields without the need for computing chain propagators, which is typically the computationally demanding step of a field theory. By integrating different types of neural network models into SCFT, we compared the performance of the models and developed a robust and computationally efficient model for Gaussian chain models that form disordered, inhomogeneous (microphase separated) structures. Our model is able to achieve a speedup of more than three times for the same size systems and up to 100 times for larger systems in our tested systems. The results of this work demonstrate one strategy for how deep learning can be leveraged to improve the efficiency of large-scale SCFT simulations, and the methods herein could be readily extended to other, more computationally demanding models.