{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,3,14]],"date-time":"2026-03-14T00:00:41Z","timestamp":1773446441048,"version":"3.50.1"},"reference-count":61,"publisher":"IOP Publishing","issue":"4","license":[{"start":{"date-parts":[[2024,10,29]],"date-time":"2024-10-29T00:00:00Z","timestamp":1730160000000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"},{"start":{"date-parts":[[2024,10,29]],"date-time":"2024-10-29T00:00:00Z","timestamp":1730160000000},"content-version":"tdm","delay-in-days":0,"URL":"https:\/\/iopscience.iop.org\/info\/page\/text-and-data-mining"}],"funder":[{"name":"Platform for Advanced Scientific Computing"},{"name":"European Research Council","award":["101001890-FIAMMA"],"award-info":[{"award-number":["101001890-FIAMMA"]}]},{"DOI":"10.13039\/501100009150","name":"National Center of Competence in Research Materials\u2019 Revolution: Computational Design and Discovery of Novel Materials","doi-asserted-by":"crossref","award":["182892"],"award-info":[{"award-number":["182892"]}],"id":[{"id":"10.13039\/501100009150","id-type":"DOI","asserted-by":"crossref"}]}],"content-domain":{"domain":["iopscience.iop.org"],"crossmark-restriction":false},"short-container-title":["Mach. Learn.: Sci. Technol."],"published-print":{"date-parts":[[2024,12,1]]},"abstract":"<jats:title>Abstract<\/jats:title>\n               <jats:p>Symmetry is one of the most central concepts in physics, and it is no surprise that it has also been widely adopted as an inductive bias for machine-learning models applied to the physical sciences. This is especially true for models targeting the properties of matter at the atomic scale. Both established and state-of-the-art approaches, with almost no exceptions, are built to be exactly equivariant to translations, permutations, and rotations of the atoms. Incorporating symmetries\u2014rotations in particular\u2014constrains the model design space and implies more complicated architectures that are often also computationally demanding. There are indications that unconstrained models can easily learn symmetries from data, and that doing so can even be beneficial for the accuracy of the model. We demonstrate that an unconstrained architecture can be trained to achieve a high degree of rotational invariance, testing the impacts of the small symmetry breaking in realistic scenarios involving simulations of gas-phase, liquid, and solid water. We focus specifically on physical observables that are likely to be affected\u2014directly or indirectly\u2014by non-invariant behavior under rotations, finding negligible consequences when the model is used in an interpolative, bulk, regime. Even for extrapolative gas-phase predictions, the model remains very stable, even though symmetry artifacts are noticeable. We also discuss strategies that can be used to systematically reduce the magnitude of symmetry breaking when it occurs, and assess their impact on the convergence of observables.<\/jats:p>","DOI":"10.1088\/2632-2153\/ad86a0","type":"journal-article","created":{"date-parts":[[2024,10,14]],"date-time":"2024-10-14T22:56:51Z","timestamp":1728946611000},"page":"04LT01","update-policy":"https:\/\/doi.org\/10.1088\/crossmark-policy","source":"Crossref","is-referenced-by-count":9,"title":["Probing the effects of broken symmetries in machine learning"],"prefix":"10.1088","volume":"5","author":[{"ORCID":"https:\/\/orcid.org\/0000-0002-1270-3016","authenticated-orcid":true,"given":"Marcel F","family":"Langer","sequence":"first","affiliation":[]},{"ORCID":"https:\/\/orcid.org\/0000-0001-5980-5813","authenticated-orcid":false,"given":"Sergey N","family":"Pozdnyakov","sequence":"additional","affiliation":[]},{"ORCID":"https:\/\/orcid.org\/0000-0003-2571-2832","authenticated-orcid":true,"given":"Michele","family":"Ceriotti","sequence":"additional","affiliation":[]}],"member":"266","published-online":{"date-parts":[[2024,10,29]]},"reference":[{"key":"mlstad86a0bib1","doi-asserted-by":"publisher","DOI":"10.1103\/RevModPhys.91.045002","article-title":"Machine learning and the physical sciences","volume":"91","author":"Carleo","year":"2019","journal-title":"Rev. Mod. Phys."},{"key":"mlstad86a0bib2","doi-asserted-by":"publisher","first-page":"414","DOI":"10.1038\/s41586-021-04301-9","article-title":"Magnetic control of tokamak plasmas through deep reinforcement learning","volume":"602","author":"Degrave","year":"2022","journal-title":"Nature"},{"key":"mlstad86a0bib3","doi-asserted-by":"publisher","first-page":"9885","DOI":"10.1073\/pnas.0603553103","article-title":"Low-dimensional, free-energy landscapes of protein-folding reactions by nonlinear dimensionality reduction","volume":"103","author":"Das","year":"2006","journal-title":"Proc. Natl Acad. Sci. USA"},{"key":"mlstad86a0bib4","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.98.146401","article-title":"Generalized neural-network representation of high-dimensional potential-energy surfaces","volume":"98","author":"Behler","year":"2007","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib5","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.104.136403","article-title":"Gaussian approximation potentials: the accuracy of quantum mechanics, without the electrons","volume":"104","author":"Bart\u00f3k","year":"2010","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib6","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.108.058301","article-title":"Fast and accurate modeling of molecular atomization energies with machine learning","volume":"108","author":"Rupp","year":"2012","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib7","doi-asserted-by":"publisher","first-page":"59","DOI":"10.1038\/s41586-020-03072-z","article-title":"Origins of structural and electronic transitions in disordered silicon","volume":"589","author":"Deringer","year":"2021","journal-title":"Nature"},{"key":"mlstad86a0bib8","doi-asserted-by":"publisher","first-page":"512","DOI":"10.1038\/s41586-022-05036-x","article-title":"The first-principles phase diagram of monolayer nanoconfined water","volume":"609","author":"Kapil","year":"2022","journal-title":"Nature"},{"key":"mlstad86a0bib9","doi-asserted-by":"publisher","first-page":"746","DOI":"10.1038\/s41928-023-01030-x","article-title":"Device-scale atomistic modelling of phase-change memory materials","volume":"6","author":"Zhou","year":"2023","journal-title":"Nat. Electron."},{"key":"mlstad86a0bib10","doi-asserted-by":"publisher","first-page":"9759","DOI":"10.1021\/acs.chemrev.1c00021","article-title":"Physics-inspired structural representations for molecules and materials","volume":"121","author":"Musil","year":"2021","journal-title":"Chem. Rev."},{"key":"mlstad86a0bib11","doi-asserted-by":"publisher","first-page":"41","DOI":"10.1038\/s41524-022-00721-x","article-title":"Representations of molecules and materials for interpolation of quantum-mechanical simulations via machine learning","volume":"8","author":"Langer","year":"2022","journal-title":"Nature"},{"key":"mlstad86a0bib12","article-title":"Geometric and physical quantities improve E(3) equivariant message passing","author":"Brandstetter","year":"2022"},{"key":"mlstad86a0bib13","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.125.166001","article-title":"Incompleteness of atomic structure representations","volume":"125","author":"Pozdnyakov","year":"2020","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib14","doi-asserted-by":"publisher","first-page":"82","DOI":"10.1016\/j.trechm.2020.10.006","article-title":"Euclidean symmetry and equivariance in machine learning","volume":"3","author":"Smidt","year":"2021","journal-title":"Trends Chem."},{"key":"mlstad86a0bib15","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1038\/s41467-022-29939-5","article-title":"E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials","volume":"13","author":"Batzner","year":"2022","journal-title":"Nat. Commun."},{"key":"mlstad86a0bib16","first-page":"pp 11423","article-title":"MACE: higher order equivariant message passing neural networks for fast and accurate force fields","volume":"vol 35","author":"Batatia","year":"2022"},{"key":"mlstad86a0bib17","first-page":"pp 37334","article-title":"TensorNet: Cartesian tensor representations for efficient learning of molecular potentials","volume":"vol 36","author":"Simeon","year":"2023"},{"key":"mlstad86a0bib18","first-page":"pp 29400","article-title":"So3krates: equivariant attention for interactions on arbitrary length-scales in molecular systems","volume":"vol 35","author":"Frank","year":"2022"},{"key":"mlstad86a0bib19","doi-asserted-by":"publisher","DOI":"10.1063\/5.0087042","article-title":"Unified theory of atom-centered representations and message-passing machine-learning schemes","volume":"156","author":"Nigam","year":"2022","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib20","article-title":"The design space of E(3)-equivariant atom-centered interatomic potentials","author":"Batatia","year":"2022"},{"key":"mlstad86a0bib21","article-title":"Symmetry breaking and equivariant neural networks","author":"Kaba","year":"2023"},{"key":"mlstad86a0bib22","first-page":"pp 5105","article-title":"PointNet++: deep hierarchical feature learning on point sets in a metric space","author":"Qi","year":"2017"},{"key":"mlstad86a0bib23","doi-asserted-by":"publisher","first-page":"pp 3172","DOI":"10.1109\/CVPR46437.2021.00319","article-title":"PAConv: position adaptive convolution with dynamic kernel assembling on point clouds","author":"Xu","year":"2021"},{"key":"mlstad86a0bib24","doi-asserted-by":"publisher","first-page":"pp 16239","DOI":"10.1109\/ICCV48922.2021.01595","article-title":"Point transformer","author":"Zhao","year":"2021"},{"key":"mlstad86a0bib25","first-page":"pp 79469","article-title":"Smooth, exact rotational symmetrization for deep learning on point clouds","volume":"vol 36","author":"Pozdnyakov","year":"2023"},{"key":"mlstad86a0bib26","doi-asserted-by":"publisher","first-page":"493","DOI":"10.1038\/s41586-024-07487-w","article-title":"Accurate structure prediction of biomolecular interactions with AlphaFold 3","volume":"630","author":"Abramson","year":"2024","journal-title":"Nature"},{"key":"mlstad86a0bib27","article-title":"Forces are not enough: benchmark and critical evaluation for machine learning force fields with molecular simulations","volume":"2023","author":"Fu","year":"2023","journal-title":"Trans. Mach. Learn. Res."},{"key":"mlstad86a0bib28","doi-asserted-by":"publisher","DOI":"10.1088\/1361-648X\/abcbdb","article-title":"Black-box inhomogeneous preconditioning for self-consistent field iterations in density functional theory","volume":"33","author":"Herbst","year":"2021","journal-title":"J. Phys.: Condens. Matter"},{"key":"mlstad86a0bib29","doi-asserted-by":"publisher","first-page":"709","DOI":"10.1038\/nnano.2007.320","article-title":"A charge-driven molecular water pump","volume":"2","author":"Gong","year":"2007","journal-title":"Nat. Nanotechnol."},{"key":"mlstad86a0bib30","doi-asserted-by":"publisher","first-page":"555","DOI":"10.1038\/nnano.2010.152","article-title":"Static charges cannot drive a continuous flow of water molecules through a carbon nanotube","volume":"5","author":"Wong-ekkabut","year":"2010","journal-title":"Nat. Nanotechnol."},{"key":"mlstad86a0bib31","doi-asserted-by":"publisher","first-page":"1110","DOI":"10.1073\/pnas.1815117116","article-title":"Ab initio thermodynamics of liquid and solid water","volume":"116","author":"Cheng","year":"2019","journal-title":"Proc. Natl Acad. Sci. USA"},{"key":"mlstad86a0bib32","doi-asserted-by":"publisher","first-page":"890","DOI":"10.1103\/PhysRevLett.80.890","article-title":"Comment on \u201cGeneralized gradient approximation made simple\u201d","volume":"80","author":"Zhang","year":"1998","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib33","doi-asserted-by":"publisher","first-page":"6158","DOI":"10.1063\/1.478522","article-title":"Toward reliable density functional methods without adjustable parameters: the PBE0 model","volume":"110","author":"Adamo","year":"1999","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib34","doi-asserted-by":"publisher","DOI":"10.1063\/1.3382344","article-title":"A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu","volume":"132","author":"Grimme","year":"2010","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib35","doi-asserted-by":"publisher","first-page":"2453","DOI":"10.1038\/s41467-022-29939-5","article-title":"E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials","volume":"13","author":"Batzner","year":"2022","journal-title":"Nat. Commun."},{"key":"mlstad86a0bib36","doi-asserted-by":"publisher","first-page":"207","DOI":"10.1109\/LSP.2015.2503295","article-title":"Gauss-Legendre sampling on the rotation group","volume":"23","author":"Khalid","year":"2016","journal-title":"IEEE Signal Process. Lett."},{"key":"mlstad86a0bib37","doi-asserted-by":"publisher","DOI":"10.1063\/5.0215869","article-title":"i-PI 3.0: a flexible and efficient framework for advanced atomistic simulations","volume":"161","author":"Litman","year":"2024","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib38","doi-asserted-by":"publisher","first-page":"3225","DOI":"10.1021\/acs.jctc.5b00301","article-title":"Transferable atomic multipole machine learning models for small organic molecules","volume":"11","author":"Bereau","year":"2015","journal-title":"J. Chem. Theory Comput."},{"key":"mlstad86a0bib39","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.120.143001","article-title":"Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics","volume":"120","author":"Zhang","year":"2018","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib40","article-title":"Frame averaging for invariant and equivariant network design","author":"Puny","year":"2021"},{"key":"mlstad86a0bib41","first-page":"pp 9013","article-title":"FAENet: frame averaging equivariant GNN for materials modeling","author":"Duval","year":"2023"},{"key":"mlstad86a0bib42","first-page":"pp 15546","article-title":"Equivariance with learned canonicalization functions","author":"Kaba","year":"2023"},{"key":"mlstad86a0bib43","article-title":"Equivariance via minimal frame averaging for more symmetries and efficiency","author":"Lin","year":"2024"},{"key":"mlstad86a0bib44","article-title":"Equivariant frames and the impossibility of continuous canonicalization","author":"Dym","year":"2024"},{"key":"mlstad86a0bib45","first-page":"pp 18582","article-title":"Learning probabilistic symmetrization for architecture agnostic equivariance","volume":"vol 36","author":"Kim","year":"2023"},{"key":"mlstad86a0bib46","doi-asserted-by":"publisher","first-page":"1990","DOI":"10.1063\/1.463137","article-title":"Reversible multiple time scale molecular dynamics","volume":"97","author":"Tuckerman","year":"1992","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib47","doi-asserted-by":"publisher","first-page":"214","DOI":"10.1016\/j.cpc.2018.09.020","article-title":"I-PI 2.0: a universal force engine for advanced molecular simulations","volume":"236","author":"Kapil","year":"2019","journal-title":"Comput. Phys. Commun."},{"key":"mlstad86a0bib48","doi-asserted-by":"publisher","first-page":"357","DOI":"10.1146\/annurev.pc.25.100174.002041","article-title":"Chemical dynamics","volume":"25","author":"Farrar","year":"1974","journal-title":"Annu. Rev. Phys. Chem."},{"key":"mlstad86a0bib49","doi-asserted-by":"publisher","first-page":"552","DOI":"10.1002\/1439-7641(20010917)2:8\/9<552::AID-CPHC552>3.0.CO;2-5","article-title":"Strong acceleration of chemical reactions occurring through the effects of rotational excitation on collision geometry","volume":"2","author":"Miklavc","year":"2001","journal-title":"ChemPhysChem"},{"key":"mlstad86a0bib50","doi-asserted-by":"publisher","first-page":"1170","DOI":"10.1021\/ct900563s","article-title":"Colored-noise thermostats \u00e0 la Carte","volume":"6","author":"Ceriotti","year":"2010","journal-title":"J. Chem. Theory Comput."},{"key":"mlstad86a0bib51","doi-asserted-by":"publisher","DOI":"10.1063\/1.2408420","article-title":"Canonical sampling through velocity rescaling","volume":"126","author":"Bussi","year":"2007","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib52","doi-asserted-by":"publisher","first-page":"26","DOI":"10.1016\/j.cpc.2008.01.006","article-title":"Stochastic thermostats: comparison of local and global schemes","volume":"179","author":"Bussi","year":"2008","journal-title":"Comput. Phys. Commun."},{"key":"mlstad86a0bib53","doi-asserted-by":"publisher","first-page":"5757","DOI":"10.1038\/s41467-020-19606-y","article-title":"Liquid water contains the building blocks of diverse ice phases","volume":"11","author":"Monserrat","year":"2020","journal-title":"Nat. Commun."},{"key":"mlstad86a0bib54","doi-asserted-by":"publisher","first-page":"1518","DOI":"10.1063\/1.473300","article-title":"Unit cells for the simulation of hexagonal ice","volume":"106","author":"Hayward","year":"1997","journal-title":"J. Chem. Phys."},{"key":"mlstad86a0bib55","article-title":"Note that even though this is a stringent test for rotational symmetry breaking, the energies are unlikely to fully capture the physics of proton ordering, given that PET, as most MLPs, is a local model and misses an explicit description of long-range electrostatics."},{"key":"mlstad86a0bib56","article-title":"Equivariant frames and the impossibility of continuous canonicalization","author":"Dym","year":"2024"},{"key":"mlstad86a0bib57","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.98.066401","article-title":"Efficient and accurate Car\u2013Parrinello-like approach to Born\u2013Oppenheimer molecular dynamics","volume":"98","author":"K\u00fchne","year":"2007","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib58","doi-asserted-by":"publisher","DOI":"10.1103\/PhysRevLett.118.015703","article-title":"Accelerating ab initio molecular dynamics and probing the weak dispersive forces in dense liquid hydrogen","volume":"118","author":"Mazzola","year":"2017","journal-title":"Phys. Rev. Lett."},{"key":"mlstad86a0bib59","doi-asserted-by":"publisher","DOI":"10.1088\/2632-2153\/abd51e","article-title":"Atomic permutationally invariant polynomials for fitting molecular force fields","volume":"2","author":"Allen","year":"2021","journal-title":"Mach. Learn.: Sci. Technol."},{"key":"mlstad86a0bib60","doi-asserted-by":"publisher","first-page":"187","DOI":"10.1007\/s41095-021-0229-5","article-title":"PCT: point cloud transformer","volume":"7","author":"Guo","year":"2021","journal-title":"Comput. Vis. Media"},{"key":"mlstad86a0bib61","doi-asserted-by":"publisher","DOI":"10.24435\/materialscloud:kz-3b","article-title":"Data and code for \u201cProbing the effects of broken symmetries in machine learning\u201d","author":"(EPFL)","year":"2024","journal-title":"Materials Cloud Archive"}],"container-title":["Machine Learning: Science and Technology"],"original-title":[],"link":[{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0","content-type":"text\/html","content-version":"am","intended-application":"text-mining"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0\/pdf","content-type":"application\/pdf","content-version":"am","intended-application":"text-mining"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0","content-type":"text\/html","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0\/pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0\/pdf","content-type":"application\/pdf","content-version":"am","intended-application":"syndication"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0\/pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0\/pdf","content-type":"application\/pdf","content-version":"am","intended-application":"similarity-checking"},{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0\/pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2024,11,13]],"date-time":"2024-11-13T14:59:48Z","timestamp":1731509988000},"score":1,"resource":{"primary":{"URL":"https:\/\/iopscience.iop.org\/article\/10.1088\/2632-2153\/ad86a0"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2024,10,29]]},"references-count":61,"journal-issue":{"issue":"4","published-online":{"date-parts":[[2024,10,29]]},"published-print":{"date-parts":[[2024,12,1]]}},"URL":"https:\/\/doi.org\/10.1088\/2632-2153\/ad86a0","relation":{},"ISSN":["2632-2153"],"issn-type":[{"value":"2632-2153","type":"electronic"}],"subject":[],"published":{"date-parts":[[2024,10,29]]},"assertion":[{"value":"Probing the effects of broken symmetries in machine learning","name":"article_title","label":"Article Title"},{"value":"Machine Learning: Science and Technology","name":"journal_title","label":"Journal Title"},{"value":"paper","name":"article_type","label":"Article Type"},{"value":"\u00a9 2024 The Author(s). Published by IOP Publishing Ltd","name":"copyright_information","label":"Copyright Information"},{"value":"2024-08-06","name":"date_received","label":"Date Received","group":{"name":"publication_dates","label":"Publication dates"}},{"value":"2024-10-14","name":"date_accepted","label":"Date Accepted","group":{"name":"publication_dates","label":"Publication dates"}},{"value":"2024-10-29","name":"date_epub","label":"Online publication date","group":{"name":"publication_dates","label":"Publication dates"}}]}}