equivariant-neural-networks

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

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

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

Tools for exploiting Morphological Symmetries in robotics

Library to make any existing neural network architecture equivariant

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

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

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.

Designing antibody CDRs with flexible CDR definition, using equivariant graph neural networks. Bioinformatics (2024).

[KDD26 Oral AI4S Track] The official implementation of the paper "EqCollide: Equivariant and Collision-Aware Deformable Objects Neural Simulator"

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

3D pharmacophore-conditioned molecular diffusion with an E(3)-equivariant EGNN backbone. Generates shape-complementary, drug-like molecules conditioned on pharmacophore point clouds and PMI/SSD shape descriptors. Inspired by ShEPhERD (Adams et al., ICLR Oral 2025). PyTorch · e3nn · RDKit.

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

Fused Metal kernels for e3nn tensor products on Apple GPUs — 3.1-6.7x faster than stock e3nn, no build step

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.

Cross-neighbour Hermitian density operators for equivariant machine-learned interatomic potentials