polyDYN: A Machine Learning Force Field for Predictive Molecular Dynamics Simulations of Polymers

The development of force fields that combine accuracy, computational efficiency, and broad chemical applicability remains a central challenge in polymer simulation. Quantum mechanical methods provide highly reliable descriptions of atomic interactions but remain computationally prohibitive for the length and timescales needed to study polymer properties, while classical force fields enable large-scale simulations at the expense of accuracy and transferability. Machine learning force fields (MLFFs) address these limitations, combining near quantum-mechanical performance with computational costs approaching those of classical. Among MLFFs, message-passing neural networks (MPNNs) are particularly promising because of their ability to learn complex atomic interactions directly from local chemical environments while preserving fundamental physical symmetries. In this work, we develop a polymer-specific force field based on the Polarizable Atom Interaction Neural Network (PaiNN) architecture and train it on twelve chemically distinct polymer systems spanning aromatic, aliphatic, heteroatomic, and branched chemistries. Using the resulting model, we perform MD simulations to predict key thermophysical and mechanical polymer properties, achieving excellent agreement with experiment. This work demonstrates the potential of polymer-specific MLFFs to enable high-fidelity simulations spanning diverse chemistries and properties.