Applications of variational autoencoders for polymer nanocomposite structure generation

Polymer nanocomposites (PNCs) are used in automobile parts, aircraft design, materials for energy storage, soft electronics, and photonics applications. In these applications the desired PNC properties are a function of the PNC morphology (i.e., dispersed/aggregated nanorods with/without orientational alignment or percolation). Molecular simulations can be useful tools that can predict PNC morphology for variety of PNC design. However, generating numerous uncorrelated structures from simulations of PNCs at high (melt-like) packing fraction can be tedious and computationally intensive for some design parameters (e.g., long polymer chains, complex nanofiller design). In this talk, we present a novel deep learning workflow using variational autoencoders that takes as input a handful of uncorrelated structures to generate a larger ensemble of uncorrelated PNC structures with similar morphological features as the input. Our workflow will aid in the fast and automatic generation of configurations of 3D structures of amorphous soft material with prescribed structural characteristics / properties.