Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 46: Poster II

CPP 46.53: Poster

Donnerstag, 12. März 2026, 09:30–11:30, P5

Benchmarking the Transferability of Machine Learning Interatomic Potentials for Polymers — •Mirko Fischer and Andreas Heuer — Institute for Physical Chemistry, University of Münster, Corrensstraße 28/30, 48149 Münster

Machine Learning Interatomic Potentials (MLIPs) enable molecular dynamics (MD) simulations with nearly quantum-chemical (QM) accuracy and have been successfully applied to various molecular systems. However, their systematic application to polymer systems has not yet been explored. A key challenge arises from the large molecular size, which makes QM reference calculations for the training data computationally demanding. Moreover, relaxation processes and diffusive behavior in polymers require long simulation times, making a direct comparison between MLIP-based and ab initio MD simulations infeasible.

In this study, we first train Atomic Cluster Expansion (ACE) potentials for small oligomers on MD reference data, thereby circumventing the need for expensive QM reference simulations and enabling long, cost-efficient MD trajectories. We then benchmark the transferability of the trained ACE potentials to longer polymer chains with respect to density, structural, and dynamic properties. The insights gained allow us to identify an optimal oligomer chain length that balances training cost and transferability. Based on this, we can efficiently train a QM-accurate potential for polymers on QM reference data in a second step.

Keywords: Polymers; Machine Learning; Molecular Dynamics; Atomic Cluster Expansion