In the world of 3D deep learning—whether modeling the folding of a protein or the forces in a new battery material—symmetry is a physical requirement, not just a preference. If you rotate a molecule, its energy should remain the same (invariance), and the forces on its atoms should rotate accordingly (equivariance).
To achieve this, researchers have traditionally relied on SO(3)-equivariant neural networks. These models use spherical harmonics and Clebsch-Gordan tensor products to ensure that the network’s internal representations transform perfectly with 3D rotations. However, these models face a "complexity wall": the math required to mix these geometric features scales at $O(L^6)$, where $L$ is the resolution (degree) of the representation. This makes it nearly impossible to model high-resolution features like $d$ and $f$ atomic orbitals or to scale to large systems.
A new wave of research is breaking this bottleneck by simplifying 3D rotational math into more efficient forms, most notably by reducing 3D operations to 2D local frames.
The SO(3) Bottleneck and the $O(L^6)$ Curse
To represent directional information, equivariant networks use irreducible representations (irreps). Think of these as a hierarchy of shapes: $L=0$ is a scalar (a sphere), $L=1$ is a vector (a dipole), and higher $L$ values represent increasingly complex angular patterns.
The standard way to combine two such features is the Clebsch-Gordan tensor product. As YuQing Xie, Tess Smidt, and colleagues explain in Asymptotically Fast Clebsch-Gordan Tensor Products with Vector Spherical Harmonics, the naive implementation of this product is computationally brutal. Because every component of one feature must interact with every component of another across multiple degrees, the cost explodes as $O(L^6)$. This forces researchers to stick to low-resolution representations (usually $L \leq 2$), leaving finer geometric details on the table.
The SO(2) Shortcut: Axis Alignment
The most successful strategy for breaking this complexity wall is a geometric "trick": reducing SO(3) to SO(2).
In a 3D graph, an interaction usually happens between two points (atoms) along an edge. Saro Passaro and C. Lawrence Zitnick demonstrated in Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs that if you rotate your local coordinate system so that the z-axis aligns perfectly with the edge between two atoms, the symmetry of the interaction simplifies.
Instead of needing to be robust to any rotation in 3D space (SO(3)), the interaction now only needs to be robust to rotations around that specific edge (SO(2)). This transformation makes the underlying Clebsch-Gordan matrices extremely sparse. The result is a reduction in complexity from $O(L^6)$ to $O(L^3)$.
This "axis alignment" approach has become the backbone of state-of-the-art models:
- EquiformerV2: Yi-Lun Liao and Brandon M Wood used this to scale equivariant transformers to $L=8$, achieving new records on the Open Catalyst dataset.
- QHNetV2: Haiyang Yu, Shuiwang Ji, and Yuchao Lin applied this to predict quantum Hamiltonian matrices, achieving a 4x speedup while handling complex orbital interactions that were previously too expensive (Efficient Prediction of SO(3)-Equivariant Hamiltonian Matrices via SO(2) Local Frames).
Cartesian Bases and Tensor Decomposition
While axis alignment simplifies the math of spherical harmonics, another group of researchers argues for moving away from spherical harmonics entirely.
In Equivariant message passing for the prediction of tensorial properties and molecular spectra, Kristof Schütt and his team introduced PAINN, which performs equivariant updates directly using Cartesian vectors. By avoiding the expansion into spherical harmonics, they achieved massive speedups in simulating molecular spectra. Similarly, Dian Jin and colleagues developed CEITNet, which uses Cartesian geometric bases to predict high-order crystal tensors (like elasticity) up to 10x faster than traditional methods (Efficient Equivariant High-Order Crystal Tensor Prediction via Cartesian Local-Environment Many-Body Coupling).
For those who stick with tensor products, Tensor Decomposition Networks (TDNs) offer a different path. Yuchao Lin and Jianwen Xie showed that the Clebsch-Gordan product can be approximated using low-rank CP decomposition, slashing complexity to $O(L^4)$ and enabling up to 80x speedups for high-order features (Tensor Decomposition Networks for Fast Machine Learning Interatomic Potential Computations).
Scaling to the Limit
The ultimate goal of these efficiency gains is to move from small molecular "toy" systems to real-world materials.
Recent breakthroughs like E2Former-V2 by Lin Huang and others have shifted computation from edges to nodes and utilized custom GPU kernels to scale equivariant modeling to 100,000 atoms (E2Former-V2: On-the-Fly Equivariant Attention with Linear Activation Memory). Meanwhile, Alexander Maeder and Chen Hao Xia have demonstrated that these efficient equivariant frameworks can be distributed across hundreds of GPUs to simulate semiconductor devices with nearly 200,000 atoms (Distributed Equivariant Graph Neural Networks for Large-Scale Electronic Structure Prediction).
By rethinking the fundamental math of 3D rotations—whether through SO(2) reduction, Cartesian bases, or algorithmic decompositions—the field has transformed equivariance from a computational luxury into a scalable tool for the next generation of scientific discovery.