From SMILES to scattering: Automated high-throughput atomistic polyurethane simulations compared with WAXS data

D Robe and A Menzel and AW Phillips and E Hajizadeh, COMPUTATIONAL MATERIALS SCIENCE, 256, 113931 (2025).

DOI: 10.1016/j.commatsci.2025.113931

A critical bottleneck in high throughput molecular modeling is the manual declaration of force field parameters. An expert operator must consider the particular environment of each atom to specify its interactions. We address this challenge by developing an end-to-end fully automated workflow, which integrates and extends several software tools (LAMMPS, RDKit, RadonPy, Signac, Psi4, and Freud) to construct, execute, and analyze molecular dynamics simulations of polymers en masse without any operator. We study polyurethanes as a class of materials with a non-trivial multi-block structure and a wide range of achievable properties. Our workflow receives SMILES strings representing hard, soft, and chain extender monomers, and procedurally constructs fully specified models with varied chemistry, molecular weight, and hard component volume fraction. This automatic modeling of polyurethanes required novel implementation of explicit representations of full chemical structures, as well as neighborhood-dependent atomic charges. With these considerations, automatically constructed models reproduced the experimental structure data from WAXS experiments, in spite of model assumptions and computational limitations. Simulations with varying hard segment content indicate that the structure factor interpolates linearly between the extremes of nearly pure hard or soft systems. The effects of temperature, block length, and block connectivity are also investigated systematically. This capability enables fully autonomous high-throughput expansion of computational data sets necessary for machine learning, material screening, and inverse design.

Return to Publications page