Many-Body Coarse-Grained Molecular Dynamics with the Atomic Cluster Expansion
Yangshuai Wang, Gabor Csanyi, Christoph Ortner · 2025
Read Many-Body Coarse-Grained Molecular Dynamics with the Atomic Cluster Expansion by Yangshuai Wang, Gabor Csanyi, Christoph Ortner (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-2502-04661.txt.
Introduction
Yangshuai Wang, Gabor Csanyi, and Christoph Ortner present a 2025 method for coarse-grained molecular dynamics built on the Atomic Cluster Expansion. Full molecular dynamics gives fine views of atoms, but it can reach just small sizes and short times. Coarse tools lump atoms into beads, so they can run far larger systems.
Such tools are in theory exact, yet in practice most aim for a rough match. The team use a math frame that names how atom clusters add up, to build coarse models that are fast, clear to read, and close to true. Many-body effects get a close look.
Catalog record
| CREATOR | Yangshuai Wang, Gabor Csanyi, Christoph Ortner |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-04661
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Many-Body Coarse-Grained Molecular Dynamics with the Atomic Cluster Expansion Yangshuai Wang, Gabor Csanyi, Christoph Ortner Abstract . Molecular dynamics (MD) simulations provide detailed insight into atomic-scale mechanisms but are inherently restricted to small spatio-temporal scales. Coarse-grained molecular dynamics (CGMD) techniques allow simulations of much larger systems over extended timescales. In theory, these techniques can be quantitatively accurate, but common practice is to only target qualitatively correct behaviour of coarse-grained models. Recent advances in applying machine learning methodology in this setting are now being applied to create also quantitatively accurate CGMD models. We demonstrate how the Atomic Cluster Expansion parameterization (Drautz, 2019) can be used in this task to construct highly efficient, interpretable and accurate CGMD models. We focus in particular on exploring the role of many-body effects. 1. Introduction Molecular dynamics (MD) simulations offer insights into atomic-level processes but are inherently limited in their ability to efficiently capture phenomena across larger spatio-temporal scales. Bridging the gap between micro-scale atomic resolution and meso-scale phenomena remains a significant challenge in understanding complex systems [ 17 , 23 ] . Coarse-graining (CG) techniques are a promising tool to extend simulation capabilities to larger systems and longer timescales. These techniques work by grouping atoms into larger particles, allowing the focus to shift from detailed atomic interactions to broader, system-level behaviors. The resulting coarse-grained molecular dynamics (CGMD) methods have been widely applied and extensively studied in a wide variety of application scenarios [ 9 , 20 , 24 , 26 , 29 , 42 , 52 , 69 ] . In CGMD, accurately representing the potential of mean force (PMF) is crucial for capturing system behavior [ 29 , 48 , 51 ] . Therefore, the precise and efficient description of the PMF is a key challenge in this field. Due to many parallels between modeling interatomic potentials (or, force fields) and the PMF, advances in machine learning interatomic potentials (MLIPs) [ 2 , 5 , 7 , 13 , 55 ] can be leveraged for this task. Works applying MLIP techniques to model the PMF include [ 14 , 15 , 21 , 25 , 37 , 41 , 49 , 66 , 75 ] . We refer to the recent review [ 54 ] for a more comprehensive discussion. In the same way as interatomic interactions, the interaction between coarse-grained particles also involve many-body effects that cannot always be ignored [ 18 , 65 , 73 ] . While the deep learning techniques cited above do incorporate some many-body effects, a systematic, efficient and interpretable approach for non-bonded many-body terms remains an ongoing challenge in developing accurate CG models. In the field of MLIPs, extensive research has focused on this aspect. Of particular note are the Moment Tensor Potential [ 55 ] and the Atomic Cluster Expansion (ACE) [ 13 ] methods which provide systematic and flexible but computationally efficient solutions. Although the ACE method and its extensions [ 2 , 6 , 13 , 39 , 70 ] has been widely used to model interatomic interactions, to the best of our knowledge, it has not been applied to approximate the PMF in the CGMD context. In this work, we investigate the use of the ACE method to construct a computationally efficient many-body coarse-grained “ACE-CG” model. Traditionally, constructing data-driven CG models involves matching mean forces, a process typically requiring computationally intensive constrained or restrained simulations [ 25 , 66 , 75 ] . To overcome this difficulty we utilize the force-matching scheme to optimize the ACE-CG model parameters [ 12 , 22 ] . In this scheme one directly derives free energy gradients from standard atomic forces (often called “instantaneous forces” in the CGMD literature) sampled from reference molecular dynamics trajectories. We test the ACE-CG method on two prototype systems: star polymers and methanol fluids. We demonstrate the ACE-CG models’ capability to accurately represent equilibrium properties such as radial distribution functions (RDFs) and angular distribution functions (ADFs). Furthermore we demonstrate the importance of including many-body effects to improve qualitative and/or quantiative accuracy of the predicted equilibrium properties, compared to pairwise potentials. Our results demonstrate the capabilities of the ACE formalism in the CGMD context, achieving systematically improvable accuracy at high computational efficiency. The scope of this work is limited to the prediction of equilibrium properties determined by the conservative potential function of a set of extensive CG variables. For the conformational free energy of non-extensive CG variables, several machine learning-based approaches have been developed [ 34 , 43 , 58 , 72 , 74 ] ; see also a recent review [ 45 ] and the references therein. Moreover, to accurately predict dynamic properties, it is necessary to properly introduce memory and coherent noise terms arising from unresolved variables into the CG model [ 19 , 28 , 33 , 40 ] . These are limitations still to be addressed to establish ACE-CG as a general-purpose framework for CGMD simulation. 2. Models and Methods Previous works [ 25 , 66 , 75 ] explored the use of machine learning methods to construct coarse-grained (CG) potentials. We provide a review of the general approach, the force matching principle, and then show how the framework can be adapted to develop a systematic and efficient many-body coarse-grained model using the Atomic Cluster Expansion (ACE). 2.1. Consistent coarse-grained models A key objective in coarse-graining is to ensure that the equilibrium distribution of a system under a CG model matches that of the reference atomistic model. This principle aligns with the concept of consistent CG models [ 46 , 47 ] . Preserving the equilibrium distribution enables consistent CG models to accurately describe equilibrium properties. In principle one can extend the approach to non-equilibrium properties such as time correlation functions [ 19 , 38 , 61 ] , but this is beyond the scope of the present work. We consider an atomistic system with n atoms with positions {\bf r}_{i} , momenta {\bf p}_{i} , {\bf r}=\{{\bf r}_{1},\ldots,{\bf r}_{n}\},\quad{\bf p}=\{{\bf p}_{1},\ldots,{\bf p}_{n}\}, as well as species z_{i} (e.g. atomic number) and masses m_{i} . We assume that atoms are in the canonical (NVT) ensemble in equilibrium, under a potential energy u(\{{\bf r}_{i},z_{i}\}_{i=1}^{n}) . Implicit in the definition is a computational cell, and we allow arbitrary length n input configurations. Analogously, the state of the system in the CG model is defined by positions {\bf R}_{I} , momenta {\bf P}_{I} of N CG variables, {\bf R}=\{{\bf R}_{1},\ldots,{\bf R}_{N}\},\qquad{\bf P}=\{{\bf P}_{1},\ldots,{\bf P}_{N}\}, as well as effective species \zeta_{I} and effective masses M_{I} . As is common, we use upper-case variables to refer to the CG model and lower-case variables to refer to the atomistic model (except for \zeta_{I} to avoid confusion with common usage of Z_{I} as atomic number). We assume that the two descriptions are explicitly connected via projections \{\mathcal{M}_{1},\ldots,\mathcal{M}_{N}\} such that {\bf R}_{I}=\mathcal{M}_{I}({\bf r}) . The projections are many-to-one mappings: many atomistic configurations are mapped to the same CG configuration. For the sake of simplicity of presentation we assume that the projections are linear in the position, {\bf R}_{I}=\mathcal{M}_{I}({\bf r})=\sum_{i\in\mathcal{J}_{I}}w_{Ii}{\bf r}_{i}, with \sum_{i}w_{Ii}=1 , where \mathcal{J}_{I} denotes the set of atoms that form the CG particle I . The center of mass projection is particularly prevalent throughout the