Machine learning surrogate models of many-body dispersion interactions in polymer melts Journal: Machine Learning: Science and Technology

Zhaoxiang Shen, Raúl I. Sosa, Jakub Lengiewicz, Alexandre Tkatchenko, Stéphane P.A. Bordas · 2025

Read Machine learning surrogate models of many-body dispersion interactions in polymer melts Journal: Machine Learning: Science and Technology by Zhaoxiang Shen, Raúl I. Sosa, Jakub Lengiewicz, Alexandre Tkatchenko, Stéphane P.A. Bordas (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15149.txt.

Introduction

Zhaoxiang Shen and four co-authors train a machine learning stand-in for a costly physics calculation, in a 2025 paper. Many-body dispersion forces are the subtle attractions that hold liquids and soft matter together, and computing them exactly is slow. Polymer melts need those forces described well, which blocks large simulations. The team builds a surrogate model on a trimmed neural network that keeps only the most relevant atomic links.

The model was checked on melts of polyethylene, polypropylene, and polyvinyl chloride. It predicted the forces with high accuracy and held up on new structures. A fast surrogate lets researchers run melt simulations at sizes and lengths that the exact method cannot reach, which serves plastics design and recycling research.

Catalog record

CREATORZhaoxiang Shen, Raúl I. Sosa, Jakub Lengiewicz, Alexandre Tkatchenko, Stéphane P.A. Bordas
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-15149

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Machine learning surrogate models of many-body dispersion interactions in polymer melts Journal: Machine Learning: Science and Technology Zhaoxiang Shen, Raúl I. Sosa, Jakub Lengiewicz, Alexandre Tkatchenko, Stéphane P.A. Bordas 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. 1 Introduction Van der Waals (vdW) dispersion interactions play a critical role in diverse physical and chemical phenomena, from layered material cohesion [ 1 ] and protein folding [ 2 ] to gecko adhesion [ 3 ] . Accurately capturing vdW dispersion forces is essential for understanding and predicting the properties of numerous engineering and biological systems. Well established pairwise (PW) models, such as Lennard-Jones (LJ) potential, Grimme’s dispersion corrections [ 4 ] , or the Tkatchenko–Scheffler (TS) method [ 5 ] , are widely used due to their simplicity and computational efficiency. While effective for certain systems, they inherently neglect the quantum many-body nature of vdW dispersion interactions, often leading to discrepancies with experimental observations [ 6 , 7 , 8 ] . For this reason, the many-body dispersion (MBD) method [ 9 , 10 ] becomes more preferable because it offers a more sophisticated and accurate approach by explicitly capturing collective long-range electron correlations. Numerous studies have shown that MBD surpasses PW models in reproducing experimental or nearly exact quantum-mechanical results, achieving superior accuracy in diverse scenarios, including thin-layer delamination [ 11 ] , modeling of supramolecular systems [ 6 ] , and polymorphs prediction for molecular crystals [ 12 ] . The major obstacle of using MBD is its computational complexity, which scales as O(N^{3}) with the number of atoms, N , thus severely limits its application to large-scale simulations. Recent efforts have attempted to accelerate MBD through physically motivated approximations and algorithmic optimizations of the analytical formulation [ 13 , 14 ] , but these approaches still require further refinement to achieve practical applicability in large-scale simulations. An alternative and more common approach is to develop a surrogate model that offers a fast yet accurate approximation, addressing the problem from the efficiency side while gradually improving accuracy. In molecular simulations, this is typically referred to as a force field (FF) model, which aims to approximate the high-fidelity energy and forces of molecular systems using functional forms based on simple geometric descriptors, such as interatomic distances, bond angles, and torsions [ 15 , 16 , 17 , 18 ] . While classical FF models are computationally efficient, they suffer from limited accuracy due to their inadequate description of polarization and many-body interactions. To overcome these drawbacks, machine learning force fields (MLFFs) [ 19 , 20 , 21 , 22 , 23 ] have emerged as promising solutions, offering quantum-level accuracy with competitive computational efficiency. MLFFs, which include a wide range of models from kernel-based methods to artificial neural networks (e.g., sGDML [ 24 ] , SchNet [ 25 , 26 ] , PhysNet [ 27 ] , GEMS [ 28 ] , among others [ 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 ] ), aim to learn the statistical relationship between chemical structures and the corresponding potential energy as well as atomic forces. Despite their advantages, many MLFFs are primarily trained for short-distance interactions and continue to rely on classical PW models for vdW interactions [ 42 ] . Some approaches, such as SpookyNet [ 43 ] and SO3LR [ 44 ] , incorporate MBD-corrected reference data into their training while explicitly embedding parametric PW vdW functions within their architectures. These approaches, however, remain essentially parametric fittings to MBD and are limited in capturing many-body vdW effects beyond intramolecular interactions. To the best of our knowledge, there is currently no MLFF model explicitly designed as an MBD surrogate capable of capturing many-body vdW effects in large-scale systems. The objective in this work is to develop an ML surrogate model for MBD interactions that is specifically designed for large-scale systems . Here, we focus on a particular class of molecular systems, the polymer melts, which can benefit from an accurate MBD description of vdW interactions, and which also requires a fast model for large-scale molecular simulations. This choice of systems follows the conclusions made in our previous work [ 45 ] , where we have identified classes of systems where MBD effects are most pronounced, namely, extended systems with flexible degrees of freedom whose physical properties are predominantly governed by vdW interactions, such as polymer crystals, polymer melts, and proteins. Among these, polymer melts emerge as the ideal initial target for our surrogate modeling. Polymer melts [ 46 , 47 ] are characterized by their disordered amorphous arrangement of polymer chains, where the chains exhibit significant mobility and fluid-like behavior while maintaining their high-molecular-weight structure. Accurately capturing vdW interactions is essential for predicting the mechanical and thermodynamic properties of polymer melts, as vdW forces directly influence key behaviors such as chain mobility, density, and viscosity [ 48 , 49 ] , while the desired molecular simulations of polymer melts often involve more than 100k atoms [ 50 , 51 , 52 , 53 ] . Despite extensive efforts that have been made to enable such large-scale molecular simulations [ 54 , 55 , 56 , 57 , 58 ] , incorporating MBD into these simulations still requires a dedicated ML architecture to provide an efficient surrogate model. In this paper, we introduce a trimmed SchNet architecture that is specifically tailored for MBD force predictions in polymer melts. It is based on the SchNet architecture [ 25 , 26 ] , which we found an optimal choice for our problem due to its simplicity and superior adaptability (see also a more detailed discussion in Section 2.2 ). SchNet has demonstrated strong performance in molecular energy and force predictions, which is attributed to its radial basis function (rbf) encoding of structural information and continuous-filter convolutions. It remains scalable with respect to system size and ensures rotational invariance for energy prediction and rotational equivariance for force prediction. In our adaptation, we reduce the dimension and depth of the model while trimming the connection graph to retain only the most relevant interactions, which are then encoded using a reduced number of trainable rbfs. In addition, inspired by

Read the complete text