Learning collective variables that respect permutational symmetry
Jiaxin Yuan, Shashank Sule, Yeuk Yin Lam, Maria Cameron · 2025
Read Learning collective variables that respect permutational symmetry by Jiaxin Yuan, Shashank Sule, Yeuk Yin Lam, Maria Cameron (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-2507-00408.txt.
Introduction
Jiaxin Yuan and co-authors build simplified models for clusters of identical particles in a 2025 paper. Such clusters can be described by a few collective variables, but a good description must respect the fact that swapping identical particles changes nothing. It must also respect shifts and rotations.
Their framework starts with a sorting-based summary of the cluster that ignores particle order. A learning step then finds the space where the cluster lingers, and autoencoders extract the key variables. A special quantity from the reduced model guides sampling of rare transitions between states. Tests on small Lennard-Jones clusters in two and three dimensions gave transition rates matching brute-force checks. The method helps study how atomic clusters rearrange.
Catalog record
| CREATOR | Jiaxin Yuan, Shashank Sule, Yeuk Yin Lam, Maria Cameron |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 129 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-00408
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning collective variables that respect permutational symmetry Jiaxin Yuan, Shashank Sule, Yeuk Yin Lam, Maria Cameron Abstract In addition to translational and rotational symmetries, clusters of identical interacting particles possess permutational symmetry. Coarse-grained models for such systems are instrumental in identifying metastable states, providing an effective description of their dynamics, and estimating transition rates. We propose a numerical framework for learning collective variables that respect translational, rotational, and permutational symmetries, and for estimating transition rates and residence times. It combines a sort-based featurization, residence manifold learning in the feature space, and learning collective variables with autoencoders whose loss function utilizes the orthogonality relationship (Legoll and Lelievre, 2010). The committor of the resulting reduced model is used as the reaction coordinate in the forward flux sampling and to design a control for sampling the transition path process. We offer two case studies, the Lennard-Jones-7 in 2D and the Lennard-Jones-8 in 3D. The transition rates and residence times computed with the aid of the reduced models agree with those obtained via brute-force methods. Keywords: coarse-graining, collective variables, permutational symmetry, autoencoders, diffusion maps, feature map, the Legoll-Lelievre orthogonality condition, Lennard-Jones, committor, forward flux sampling 1 Introduction Molecular dynamics (MD) simulations offer a platform for investigating the properties and behaviors of molecular systems at an atomic level. Typically, MD simulations aim to explore the configurational space, identify metastable states, and estimate the transition rates between them. The two main challenges in achieving these goals are high dimensionality and a wide range of timescales. The dimension of the configurational space is 3N , where N , the number of atoms, is often large. The shortest timescale of interatomic bond vibrations that determines the timestep of MD simulations is of the order of a femtosecond, while conformational changes of interest happen on much larger timescales. Collective variables (CVs), differentiable functions of atomic coordinates, facilitate the study of complex molecular systems. They are typically used for • biasing MD simulations, • defining the metastable states, and • visualizing the results. In small organic molecules, such as butane and alanine dipeptide, the dihedral angles formed by triplets of atoms along the molecule’s backbone serve as natural collective variables. The MD force fields often admit significant changes in dihedral angles while the lengths of covalent bonds and interbond angles change relatively little. However, for large and complex molecules or atomic clusters, finding suitable CVs is much more challenging. In 2005, Ma and Dinner MaDinner2005 proposed using a neural network to learn the committor, facilitating the study of isomerization in the alanine dipeptide. Since then, machine learning methods have shown great promise for identifying CVs and exploring the conformational space of biomolecules – see Refs. Gkeka2020review ; Noe2021review ; Ferguson2020review for a comprehensive overview and references therein. Over the past decade, researchers have successfully employed autoencoders Kramer1992AutoassociativeNN to discover CVs in molecular dynamics (MD) simulations. Time-lagged autoencoders Noe2018autoencoders are trained to predict future configurations, \mathbf{x}_{\tau}\rightarrow\mathbf{x}_{\tau+1} , and identify low-dimensional embeddings that capture the slow dynamics of diffusion processes. The RAVE algorithm Tiwary2018RAVE , a widely adopted method, utilizes a variational autoencoder to learn the latent distribution and interprets CVs in terms of pre-selected order parameters. Ferguson’s group Ferguson2018autoencoders_on-the-fly ; Ferguson2018autoencoders_innovations developed neural network architectures within the autoencoder framework to accommodate periodicity and eliminate translational and rotational invariance in CV discovery. In recent work by Belkacemi et al. belkacemi2023autoencoders , CVs learned by autoencoders Kramer1992AutoassociativeNN allowed the authors to identify six metastable states in HSP90 and map them to a 2D space. In this work, we address the problem of learning CVs for a class of systems that lack obvious physically motivated reaction coordinates: systems of interacting particles with permutational symmetry. Examples are the Lennard-Jones clusters modeling heavy rare gas (Ar, Xe, Cr) clusters at low temperatures Wales_2004_book , and sticky spheres such as micron-size colloidal particles interacting via the depletion force HolmesCerfon2017Sticky . The goal of this work is to • develop an algorithm for learning CVs respecting the permutational symmetry that allows us to (i) define the metastable states and (ii) accurately estimate the transition rate between them, and • apply it to Lennard-Jones clusters and compare the utility of the learned CVs with the standard ones, the second and third central moments of the coordination numbers Tribell02010LJ7 ; Tiwary2019CNum . Our original goal was to learn CVs such that the transition rates in the resulting coarse-grained model match those in the original model. However, we have realized that this requirement is incompatible with the invariance of the coarse-grained model under permutational symmetry. Therefore, we have split this project into two. In the first study, Sule et al. Sule2025butane explore the task of preserving rate under rotational and translational symmetry through a case study of a normal-butane molecule. Only a thoughtfully chosen feature map and a carefully designed CV result in an accurate rate, while other seemingly reasonable options yield exaggerated rate estimates. This work is devoted to learning CVs respecting rotational, translational, and permutational symmetry and using the committor of the resulting coarse-grained model in forward flux sampling Allen_2009 and stochastic control YUAN_2024_optimalcontrol to find the transition rates and residence times. We adjust a framework for identifying collective variables (CVs) proposed in the parallel work Sule2025butane for systems with permutational symmetry. The framework consists of five steps. 1. A feature map that achieves invariance under rotational, translational, and permutational symmetry. We map the atomic coordinates into a vector invariant under rotations and translations, e.g., a vector of pairwise distances squared or a vector of coordination numbers Tribell02010LJ7 ; Tiwary2019CNum , and then sort this vector. 2. Learning the residence manifold . This step is done by means of diffusion maps CoifmanLafon2006 ; Coifman_2008 ; banisch2017_target_measure_diffusionMap and diffusion nets 2015diffusionnets . 3. Learning the CVs whose level sets are orthogonal to the residence manifold. To do it, we first design a confining potential V_{1} whose zero-level set is the residence manifold. Then, we learn CVs whose gradients are orthogonal to the gradient of V_{1} at the residence manifold via the use of autoencoders following Refs. Noe2018autoencoders ; Ferguson2018autoencoders_on-the-fly ; Ferguson2018autoencoders_innovations ; Tiwary2018RAVE ; Pande2018autoencoders ; Bolhuis2021autoencodes ; Belkacemi_2021 . This orthogonality requirement is motivated by the orthogonality condition obtained by Legoll and Lelievre Legoll_2010 for a scalar CV and then extended to vector CVs by Duong et al. Sharma_2018 . This condition reduces the coarse-grained error measured by the relative entropy of the probability density of the CV following the original process and the probability density of the coarse-grained process in this CV. 4. Computing the free energy and the diffusion tensor for the reduced dynamics in the learned CVs. This step is standard. The calculation of the diffusion matrix follows the recipe proposed in Maragliano et al. Maragliano . The calculation of the free energy consists of flattening the energy landscape by means of well-tempered metadynamics Branduardi2012MetadynamicsWA ; Dama2014WelltemperedMC ; Valsson2016metad , binning the stochastic trajectory in the flattened landscape, and then restoring the free energy. 5. Computing the committor for the low-dimensional coarse-grained model . This committor is used as the reaction coordinate for the forward flux sampling Allen_2005 and to design a controller for sampling transition trajectories YUAN_2024_optimalcontrol . Using this framework, we have conducted two