Machine-learning accelerated density-explicit polymer field theory simulations

Recently, the density-explicit framework, which describes the thermodynamic properties of multicomponent polymer systems as functionals of both auxiliary and density fields, has attracted growing interest in polymer field theory simulations. This is due to the flexibility of the framework in accommodating various forms of intermolecular potentials, and in particular, the easy generalization to many-body interactions beyond pair potentials. However, numerical simulations based on the formalism are generally more expensive owing to the increased number of fields, compared to the conventional auxiliary field framework. In this work, we develop deep neural networks with efficient low-dimensional and largely local feature representations that are applicable across different spatial resolutions and dimensions to accelerate polymer field theory simulations based on this framework. Our results may serve as a stepping stone toward accurate and efficient prediction of the phase behavior of complex block copolymer mesophases, as well as a blueprint for developing machine learning-assisted field theoretic simulation tools for the computational study of polymers and soft matter systems more broadly.