Abstract
Curie's principle states that "when effects show certain asymmetry, this asymmetry must be found in the causes that gave rise to them". We demonstrate that symmetry equivariant neural networks uphold Curie's principle and can be used to articulate many symmetry-relevant scientific questions into simple optimization problems. We prove these properties mathematically and demonstrate them numerically by training a Euclidean symmetry equivariant neural network to learn symmetry-breaking input to deform a square into a rectangle and to generate octahedra tilting patterns in perovskites.
Introduction: The Challenge of Symmetry Breaking
In physics and materials science, identifying the sources of "symmetry breaking" is fundamental to understanding phase transitions, such as superconductivity or ferroelectricity. When a system transitions from a high-symmetry state to a lower-symmetry state, the missing symmetry is described by an "order parameter." Traditionally, finding these parameters requires deep domain expertise and manual group-theory analysis using crystallographic character tables.
While standard machine learning models are often "black boxes" that ignore physical symmetries, and "invariant" models only capture scalar properties, this paper utilizes Euclidean symmetry-equivariant neural networks. These networks ensure that inputs, outputs, and internal features rotate and translate predictably with the physical system, acting as strict enforcers of physical laws.
Methodology: Networks as Symmetry Compilers
The authors utilize the e3nn PyTorch framework to build 3D Euclidean equivariant networks. The core of their approach rests on two mathematical pillars:
- Curie's Principle: An equivariant network's output must have equal or higher symmetry than its input. It cannot "break" symmetry on its own.
- Gradient Symmetry: The gradients of an invariant loss function with respect to the input can have lower symmetry than the input itself.
To discover a missing order parameter, the authors train a network to map a high-symmetry input to a lower-symmetry target. Because of Curie's Principle, the network initially produces a symmetrically degenerate "blur." By freezing the network and backpropagating loss gradients to update the inputs, the model is forced to learn the exact mathematical form—the irreducible representations (irreps)—of the missing symmetry-breaking information.
Results - Toy Model: Square to Rectangle
The first demonstration involves a 2D geometry task: deforming a square (high symmetry, $D_{4h}$) into a rectangle (lower symmetry, $D_{2h}$).
As shown in Figure 1, the network successfully identified the specific irreps required for the deformation. This result perfectly matches the manual derivations a physicist would perform using a $D_{4h}$ character table, demonstrating that the network can autonomously discover the geometric components of symmetry breaking.
Results - Material Science: Perovskite Octahedral Tilting
The method was further applied to complex 3D materials: perovskite crystal structures ($ABX_3$). These crystals often transition from a highly symmetric cubic parent structure ($Pm\bar{3}m$) to lower-symmetry structures through the tilting of their $BX_6$ octahedra.
In this experiment (Figure 3), the network recovered the exact pseudovector order parameters corresponding to the $M^{3+} \oplus R^{4+}$ irreps for the $Pnma$ structure. Furthermore, by constraining the learned parameters, the network generated intermediate structures like $Imma$, showcasing its ability to map the entire symmetry-breaking pathway.
Conclusion and Future Outlook
By leveraging equivariance, the authors have turned neural networks into "symmetry compilers" that bypass tedious manual lookups in crystallographic tables. This approach provides geometrically interpretable order parameters that are immediately useful to researchers. Future work may extend this technique to discover unknown environmental parameters in experimental physics, such as hidden magnetic field anisotropies, further democratizing advanced symmetry analysis.
Summary updated