Validation of Universal Neural Network Potentials for Bulk and Thermodynamic Properties of Polymers

Atomic-scale simulations are increasingly important in polymer materials research because many properties are governed by molecular structure, chain conformations, local packing, interfacial interactions, and chemical reactions. Predictive modeling therefore requires an accurate description of intrachain and intermolecular interactions. Density functional theory (DFT) provides high accuracy, but its cost limits accessible system sizes and time scales. Classical molecular dynamics (MD) can access larger systems and longer time scales, but its reliability depends on force-field parameterization and its transferability is limited for complex chemical environments and reactive processes. 

Machine-learning interatomic potentials (MLIPs) offer a promising route to bridge the gap between DFT-level accuracy and practical computational feasibility. Matlantis PFP is a pre-trained universal neural network potential trained on large DFT datasets [1]. PFP enables atomistic MD simulations beyond the typical length and time scales accessible by DFT calculations.

In this work, we evaluate the applicability of PFP to simulations of polymer materials. The model was trained on DFT reference data generated using the r2SCAN and PBE functionals. All MD simulations were carried out using LAMMPS. Polymer structures were equilibrated by high-temperature and high-pressure relaxation using the PCFF force field, followed by NPT simulations at 300 K and 1 atm using PCFF and PFP.

Benchmark polymer systems were used to validate the accuracy of PFP. Densities calculated using PFP showed good agreement with experimental data and higher accuracy than those from PCFF. These results suggest that PFP reasonably describes interactions relevant to polymer chain packing. We will further discuss the results for thermodynamic properties such as glass transition temperatures and demonstrate PFP for computational polymer materials developments. 

[1] S. Takamoto et. al., Nat Commun 2022, 13, 2991.