[ICLR 2024] EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations

[ICLR 2023 Spotlight] Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs

Tools for exploiting Morphological Symmetries in robotics

EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers

Library to make any existing neural network architecture equivariant

Annotated implementations of equivariant (graph) neural networks in Jax: EGNN, SEGNN, NequIP.

High-performance CUDA kernels for equivariant graph neural networks (MACE, NequIP, Allegro). 10-20x faster than e3nn.

Interactive exploration of equivariant neural networks on homogeneous spaces, with a focus on the sphere S² as SO(3)/SO(2). From Lecture 8 of the Lie groups course with Quantum Formalism

[ICML'25] "Rethinking Addressing in Language Models via Contextualized Equivariant Positional Encoding" by Jiajun Zhu, Peihao Wang, Ruisi Cai, Jason D. Lee, Pan Li, Zhangyang Wang

Torch-based library for ML problems with symmetry priors. It provides equivariant neural network modules, models, and utilities for leveraging group symmetries in data.

Variational algorithms for finding ground states of lattice gauge theories using gauge equivariant neural network ansätze. Phys. Rev. B 110, 165133 (2024).

Geometric deep learning (EGNN) + diffusion modeling for RNA 3D structure prediction.

A convolution-equivariant deep-learning approach to segmentation of histopathology data

D4-equivariant CNN for pixel-level avalanche debris segmentation from bi-temporal Sentinel-1 SAR. Matches SOTA Swin-UNet on pixel F1 with 3.8x fewer params; beats it on instance-level detection F1.

A neural completion system for Islamic geometric line patterns that guarantees exact N-fold rotational symmetry by construction, enforced through orbit-tied prediction or inference-time projection and verified by a label-free rotation audit.

Implement SE(3)-equivariant graph attention transformers for efficient and expressive molecular modeling in PyTorch.

GeomAlgLab — Interactive Geometric Algebra & Equivariant Neural Networks Laboratory. 6 modules: multivector playground, rotor explorer, projective GA, conformal GA, Clifford neural nets, GATr architecture arena. Zero dependencies.

Compiling equivariant quantum neural networks can break their symmetry at the structural level, with up to 22.5% CZ-count asymmetry between symmetry-related sparse-graph inputs on IBM Heron's Fez. We classify ansätze by where the input enters and fix the affected case via compile-once-and-relabel.

A training-free, channel-level audit that tests whether a noisy quantum neural network still respects an intended symmetry once the ideal circuit is composed with hardware noise.

Generate Biomedicines — generative biology platform for de novo protein design (Chroma model)