Machine-learning hidden symmetries

Ziming Liu, Max Tegmark

Introduction

In physics, symmetry is a cornerstone for understanding the fundamental laws of nature. Symmetries, such as rotational or translational invariance, allow physicists to simplify complex systems and derive conservation laws. However, many physical systems possess "hidden" symmetries that are only apparent in specific coordinate systems. Historically, discovering these coordinates required profound human intuition—for instance, it took nearly two decades to find the Gullstrand-Painlevé coordinates that reveal the translational symmetry of a Schwarzschild black hole.

In "Machine-learning hidden symmetries," Ziming Liu and Max Tegmark propose an automated framework to discover these hidden symmetries using neural networks and symbolic regression.

Methodology: The Neural Discovery Pipeline

The authors introduce a two-stage process to "un-obfuscate" physical laws. The first stage utilizes an Invertible Neural Network (INN) to learn a coordinate transformation $z \to z'$. Unlike standard networks, an INN ensures the transformation is smooth and reversible, preventing the model from collapsing data into trivial solutions.

The core of the training is the Symmetry Loss. The network is trained to minimize the violation of Partial Differential Equations (PDEs) that define specific symmetries (e.g., $\partial T / \partial z_j = 0$ for translational invariance). By treating symmetry as a constraint on a tensor field, the search becomes a differentiable optimization problem.

The automated pipeline uses an Invertible Neural Network to find coordinate transformations that minimize symmetry-related loss functions.

Following the neural discovery phase, the system employs Symbolic Refinement. Using tools like AI Feynman, the numerical transformation learned by the "black box" neural network is converted into a clean, human-readable mathematical formula. This ensures the results are not just accurate but also interpretable by physicists.

Experimental Evaluation

The framework was tested on six diverse physical systems, ranging from classical mechanics to general relativity:

  1. 1D Uniform Motion & Harmonic Oscillator: Rediscovering translational and rotational symmetries.
  2. 2D Kepler Problem: Identifying hidden canonical equivariance.
  3. Double Pendulum: Automatically discovering "normal modes" by minimizing modularity loss.
  4. Expanding Universe & Schwarzschild Black Hole: Finding symmetries in complex spacetime metrics.

Results

The model successfully identified hidden symmetries across all test cases. Notably, it rediscovered the Gullstrand-Painlevé metric for black holes, a task that historically challenged the physics community. The training process typically shows a sharp drop in loss as the network converges on the optimal coordinate transformation.

Loss curves across six physical systems (A-F) demonstrate the model's ability to minimize symmetry violations and discover hidden structures.

Beyond geometric symmetries, the model proved capable of identifying Hamiltonicity (energy conservation) and Modularity (system decomposability). This suggests that the framework can discover not just spatial symmetries, but the underlying dynamical rules of a system.

Discussion and Future Directions

While powerful, the current method requires users to pre-specify the types of symmetries (PDEs) to search for. Future work aims to make the "Lie generators" of these symmetries trainable, potentially allowing the AI to discover entirely new classes of mathematical symmetries. Additionally, while the method is robust to noise, scaling to high-dimensional systems like fluid dynamics remains a computational challenge.

By replacing "inspired guesses" with algorithmic search, this work provides a new lens for physics, potentially uncovering hidden order in complex experimental datasets where human intuition may fall short.

Summary updated

Abstract

We present an automated method for finding hidden symmetries, defined as symmetries that become manifest only in a new coordinate system that must be discovered. Its core idea is to quantify asymmetry as violation of certain partial differential equations, and to numerically minimize such violation over the space of all invertible transformations, parametrized as invertible neural networks. For example, our method rediscovers the famous Gullstrand-Painleve metric that manifests hidden translational symmetry in the Schwarzschild metric of non-rotating black holes, as well as Hamiltonicity, modularity and other simplifying traits not traditionally viewed as symmetries.
BibTeX
@misc{liu2021machine,
  author = {Ziming Liu and Max Tegmark},
  title = {{Machine-learning hidden symmetries}},
  year = {2021},
  eprint = {2109.09721},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.09721},
}
Figure