Machine Learning Symmetry Discovery for Integrable Hamiltonian Dynamics
Wanda Hou, Molan Li, Yi-Zhuang You · 2024
Read Machine Learning Symmetry Discovery for Integrable Hamiltonian Dynamics by Wanda Hou, Molan Li, Yi-Zhuang You (2024) 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-2412-14632.txt.
Introduction
Wanda Hou, Molan Li, and Yi-Zhuang You teach machines to discover hidden symmetries of tidy mechanical systems. Conserved quantities reveal a system's deep order but are hard to spot by eye. Their framework fits candidate conserved amounts with neural networks. A closure rule on Poisson brackets forces the candidates into a true Lie algebra.
A soft penalty keeps the found quantities from collapsing into copies. Two classic systems test the method: the Kepler problem and the round harmonic oscillator. The known non-Abelian algebras, SO(4) and SU(3), came back up to basis change. Chaotic and mixed systems are named as the next frontier.
Catalog record
| CREATOR | Wanda Hou, Molan Li, Yi-Zhuang You |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-14632
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine Learning Symmetry Discovery for Integrable Hamiltonian Dynamics Wanda Hou, Molan Li, Yi-Zhuang You Abstract We propose a data-driven Machine-Learning Symmetry Discovery (MLSD) framework for identifying continuous symmetry generators and their Lie-algebraic structure directly from phase-space trajectory data expressed in canonical coordinates. MLSD parameterizes candidate conserved quantities with neural networks and learns antisymmetric structure coefficients by enforcing Poisson-bracket closure, supplemented by a weak independence regularizer. We validate MLSD on two integrable benchmark systems—the three-dimensional Kepler problem and the three-dimensional isotropic harmonic oscillator—recovering the expected non-Abelian algebras (respectively 𝔰 𝔬 ( 4 ) \mathfrak{so}(4) and 𝔰 𝔲 ( 3 ) \mathfrak{su}(3) ) up to basis transformations. This work focuses on integrable benchmark dynamics, where global conserved quantities are well-defined and admit compact representations learnable from canonical-coordinate trajectories. Extending symmetry discovery to mixed or chaotic phase-space regimes is an important direction for future work. I Introduction The term symmetry originates from the Greek words syn (meaning “same”) and metron (meaning “measure”). At its core, symmetry reflects a fundamental concept in our understanding of the natural world: the ability to preserve identical measurable properties under transformations or dynamics. Much of the progress in physics as a science has come from uncovering symmetries within physical systems, deepening our insight into the laws governing the universe [ 21 , 15 , 18 , 2 ] . While the discovery of symmetries is invaluable, it is often a challenging task that typically requires the sophisticated expertise of physicists. To address this challenge, advancing field of machine learning [ 11 , 1 , 3 ] offers a promising set of tools for assisting symmetry discovery from data in controlled physical settings. One significant advantage of machine learning in this context is its ability to learn and compress from large volumes of data. This data-driven approach provides a systematic way to extract patterns and correlations from data that may not be evident from standard theoretical analysis. By harnessing the vast amounts of data generated in scientific research, machine learning holds the potential to accelerate discovery and provide new insights into the fundamental laws and symmetries governing physical systems [ 20 , 16 , 9 ] . Previous work has demonstrated learning-based approaches to uncovering continuous symmetries in classical mechanical systems, primarily in settings with regular dynamics where conserved quantities admit stable functional representations [ 22 , 14 , 13 , 23 , 17 , 4 , 19 , 12 ] . In this paper we similarly focus on benchmark Hamiltonian systems with integrable dynamics, where the relevant conserved quantities are well-defined on the sampled phase-space region and admit compact representations. A natural next step within this line of work is to develop a systematic framework for discovering continuous Lie group symmetries, particularly non-Abelian ones, directly from trajectory data, without assuming prior analytic knowledge of the conserved quantities. In this study, we propose a Machine-Learning Symmetry Discovery (MLSD) algorithm designed to identify continuous Lie group symmetries in classical mechanical systems with integrable Hamiltonian dynamics, using simulated time-evolution trajectory data. The MLSD algorithm identifies conservation laws that correspond to continuous symmetry transformations. After training, MLSD outputs a three-way tensor f , representing the Lie algebra structure coefficients of the discovered set of symmetry transformations. The structure coefficients provide a basis-independent representation that allows MLSD to identify both Abelian and non-Abelian symmetry algebras within the considered setting. Within this deliberately scoped benchmark setting, MLSD provides an end-to-end pipeline that (i) learns candidate conserved quantities from trajectories, (ii) recovers a basis-independent Lie-algebra signature via structure coefficients and Killing-form analysis, and (iii) reports stability and reproducibility diagnostics suitable for systematic comparison across symmetry dimensions. The paper is organized as follows: we first review the definition of continuous symmetries in classical systems, then explain each loss term designed in MLSD to discover symmetries using neural networks, and finally examine our proposed algorithm on two tasks: discovering \mathrm{SO}(4) symmetry in the three-dimensional Kepler problem; and discovering \mathrm{SU}(3) symmetry in the three-dimensional harmonic oscillator. Figure 1: The Machine Learning Symmetry Discovery (MLSD) framework takes time evolution data from a classical system and feeds the canonical coordinates into neural networks to predict a physical quantity. By optimizing the loss function, the predicted observables converge to conserved quantities, and the corresponding structure coefficients reveal the underlying symmetry group. II Methodology Our goal of symmetry discovery differs from those explored in previous studies. For example, [ 22 , 23 ] focuses on identifying transformations that preserve time evolution trajectories, while [ 14 , 13 ] examines whether specific proposed transformations can be learned by neural networks. In contrast, our approach seeks to uncover symmetry transformations that extend beyond preserving the shape of trajectories and instead focus solely on preserving energy, as similarly discussed in [ 17 , 4 , 19 ] , without assuming analytic forms for the conserved quantities. In this section, we begin by reviewing the concept of symmetry in classical systems as it pertains to our study. We then introduce the architecture of the Machine Learning Symmetry Discovery (MLSD) framework, which is designed to identify these symmetries. II.1 Continuous symmetry and symmetry group of a physical system The time evolution dynamics of a physical system is encoded in the Hamiltonian H({\bm{x}}) , which depends on the canonical coordinates {\bm{x}}:=({\bm{q}},{\bm{p}}) , with {\bm{q}} and {\bm{p}} being the conjugate position and momentum variables. A continuous symmetry of the system is defined by the invariance of the Hamiltonian under a continuous family of canonical transformations generated by a conserved quantity. This can be expressed through the vanishing Poisson bracket between H({\bm{x}}) and the conserved quantity G_{i}({\bm{x}}) , where i\in\{1,2,...,n\} and n denote the dimension of the symmetry group. The identification of the continuous symmetry group (Lie group) follows directly from the Lie algebra structure coefficients f^{ijk} , which are obtained from the commutators between the quantities. \{H,G_{i}\}=0,\forall\ i\in\{1,2,...,n\}. (1) \{G_{i},G_{j}\}=\sum_{k}f^{ijk}G_{k},\forall\ i,j\in\{1,2,...,n\}. (2) Here the Poisson bracket \{A,B\} between two canonical functions A({\bm{x}}) and B({\bm{x}}) is defined by the following vector-matrix-vector multiplication \{A,B\}:=(\nabla_{\bm{x}}A)^{\intercal}\mathbf{J}(\nabla_{\bm{x}}B), (3) where \nabla_{\bm{x}} is the gradient operator, and the matrix \mathbf{J}=\begin{pmatrix}&\mathds{1}_{d\times d}\\ -\mathds{1}_{d\times d}&\end{pmatrix} (4) characterizes the symplectic metric of the canonical coordinate {\bm{x}} in the phase space. Since the three-way tensor f^{ijk} is anti-symmetric by definition, we can assign G_{0}({\bm{x}}):=H({\bm{x}}) together with f^{0jk}=f^{i0k}=f^{ij0}=0 . Therefore Eq. ( 1 ) and Eq. ( 2 ) can be combined for compactness as \{G_{i},G_{j}\}=\sum_{k}f^{ijk}G_{k},\forall\ i,j\in\{0,1,2,...,n\}. (5) The goal in this work is to discover a set of independent conserved quantities G_{i}({\bm{x}}) and associated structure coefficients f^{ijk} that describe the symmetry algebra, within the expressivity of the chosen parametrization and the phase-space coverage of the dataset. The special G_{0} component corresponds to the Hamiltonian H , relating to energy conservation. Its dependence H({\bm{x}}) on {\bm{x}} can be learned from a series of time evolution data {\bm{x}}(t) , assuming the underlying principle of Hamiltonian dynamics of classical mechanics. II.2 Learning continuous symmetry transformations Considering a classical system in spatial dimension d , let {\bm{X}}=\{{\bm{x}}(t)\in\mathbb{R}^{2d}:\dot{{\bm{x}}}=\{{\bm{x}},H\}\} denote a dataset of time-series samples, where each sample is a trajectory of the canonical coordinates {\bm{x}}(t) evolving under Hamiltonian dynamics generated by an unknown Hamiltonian H . Throughout this work, we assume that the dataset {\bm{X}} consists of simulated trajectories expressed in