Geometric Deep Learning: SE(3)-Equivariant Networks — respecting symmetries in molecular, crystallographic, and point-cloud models

Geometric deep learning incorporates domain symmetries: rotations, translations, reflections. SE(3)-equivariant networks (SE(3) = 3D rotations + translations) preserve physical invariances, improving generalization and data efficiency.

Equivariance Principles

Invariance: f(g·x) = f(x) (output unchanged by transformation). Equivariance: f(g·x) = g·f(x) (output transforms same way as input). SE(3)-equivariance crucial for molecules: rotating/translating molecule shouldn't change predicted properties (invariance) but should transform atomic forces/velocities correspondingly (equivariance). Gauge-equivariance (additional generalization): permits learning different gauges (coordinate systems) for different atoms.

SE(3)-Transformer and Tensor Field Networks

SE(3)-Transformer: attention mechanism respecting SE(3) symmetry. Type-0 (scalar) features: invariant (attention scores computed from scalars). Type-1 (vector) features: equivariant (directional attention output transforms as vectors). Multi-head attention aggregates information across types. Transformer layers stack, building expressive SE(3)-equivariant networks.

e3nn Library and Point Cloud Processing

e3nn (Equivariant 3D Neural Networks): PyTorch library implementing SE(3)-equivariant layers. Tensor products combine representations respecting equivariance. Applications: point cloud classification (ModelNet, ScanNet), semantic segmentation (3D shape part labeling). PointNet++ with equivariance constraints improves robustness to rotations.

Molecular Applications

SchNet and DimeNet leverage SE(3) symmetry: interatomic distances (invariant), directional angles (equivariant). Message passing: h_i ← UPDATE(h_i, [h_j for neighbors j], relative geometry). Applications: predict molecular properties (atomization energy, dipole moment), forces (for MD simulation), and electron density. Equivariance enables: fewer training samples (symmetry is inductive bias), better generalization to new molecules, transferability across datasets.

Materials Science and Crystallography

Crystal structures have space group symmetries (1-230 space groups defining crystallographic constraints). E(3)-equivariant networks respect these symmetries, crucial for crystal property prediction (band gap, magnetic moments). NequIP (Neural Equivariant Interatomic Potential): SE(3)-equivariant GNN for molecular dynamics, achieving quantum mechanical (DFT) accuracy 100x faster. Applications: materials screening, alloy design, defect prediction.