Machine learning surrogate models of many-body dispersion interactions in polymer melts

TL;DR

The study addresses the prohibitive cost of many-body dispersion (MBD) calculations in large-scale polymer melts by developing a trimmed SchNet-based surrogate that predicts center-atom MBD forces. The model uses trimmed interaction graphs, trainable radial-basis encoding with $N_{rbf}=100$, and unit-specific batching, trained on PE, PP, and PVC datasets with $N_{cut}=1000$ and a cutoff near $14\,\AA$, achieving fast inference and strong generalization. It analyzes key design choices, demonstrates robustness across mixtures and a temperature range, and provides Hessian-level insight into the learned force landscape, including implications for cutoff optimization. The authors also show practical MD integration via open-source data and code, achieving substantial speedups over direct MBD and outlining future work to couple with existing force fields and extend to other condensed-phase systems.

Abstract

Accurate prediction of many-body dispersion (MBD) interactions is essential for understanding the van der Waals forces that govern the behavior of many complex molecular systems. However, the high computational cost of MBD calculations limits their direct application in large-scale simulations. In this work, we introduce a machine learning surrogate model specifically designed to predict MBD forces in polymer melts, a system that demands accurate MBD description and offers structural advantages for machine learning approaches. Our model is based on a trimmed SchNet architecture that selectively retains the most relevant atomic connections and incorporates trainable radial basis functions for geometric encoding. We validate our surrogate model on datasets from polyethylene, polypropylene, and polyvinyl chloride melts, demonstrating high predictive accuracy and robust generalization across diverse polymer systems. In addition, the model captures key physical features, such as the characteristic decay behavior of MBD interactions, providing valuable insights for optimizing cutoff strategies. Characterized by high computational efficiency, our surrogate model enables practical incorporation of MBD effects into large-scale molecular simulations.

Figures (9)

• Figure 1: A PE melt structure. (a) is the original unfolded configuration of the unit cell. (b) is the folded unit cell according to the periodic boundary condition. (c) shows an atomic cluster with 1000 atoms cut off from (b), in which the center atom is marked as red.
• Figure 2: Illustration of the trimmed SchNet. The trimmed connections graph is depicted for a toy system that consists of $N=9$ atoms, where only connections to the central target atom (red solid lines) and a few extra connections to its two nearest neighbors (yellow solid lines) are retained, while all other peripheral connections (gray dashed lines) are trimmed. The detailed architectures of rbf encoding, interaction, and projection are shown in the bottom.
• Figure 3: Segments of the three types of polymer used in this work, with carbon atoms in dark gray, hydrogen atoms in white, and chlorine atoms in green. The monomer of each polymer (ethylene for PE, propylene for PP, and vinyl chloride for PVC) is highlighted in non-transparent shading.
• Figure 4: Training and validation convergence of the model on three polymer datasets, following the proposed training strategy.
• Figure 5: Histogram of angular error ($\mathcal{E}_\text{angle}\,(^\circ)$) for the model trained on each single-polymer dataset and tested on the corresponding test dataset.