e3nn-jax

What it is and what it does

e3nn-jax is a JAX implementation of Euclidean neural networks that enforce equivariance to 3D rotations, translations, and mirror symmetries. It wraps tensor data in an IrrepsArray structure that tracks irreducible representations (irreps) of the E(3) group, allowing neural networks to respect these geometric symmetries by design. This is particularly useful for molecular systems, point clouds, and other 3D data where physical symmetries should be preserved or exploited by the model.

The package provides operations like tensor products, norms, and convolutions that work directly on irreps-annotated arrays. It depends on jax, jaxlib, numpy, sympy, and attrs. The library is actively maintained, supports modern Python versions (3.9+), and is distributed under Apache 2.0, making it suitable for both research and commercial use.

Use it for:

• Train equivariant neural networks on molecular structures or point clouds where 3D symmetries matter for prediction accuracy.
• Build MACE-style interatomic potential models for molecular dynamics simulations with guaranteed rotation/translation equivariance.
• Prototype geometric deep learning models in JAX that respect Euclidean symmetries without manual constraint engineering.
• Compute tensor products and irrep decompositions for group-theoretic operations in physics-informed machine learning.
• Develop models on GPU/TPU hardware using JAX's compilation and automatic differentiation with built-in symmetry guarantees.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Implements equivariant convolutional neural networks for 3D rotations, translations, and mirrors using JAX, with an IrrepsArray data structure that tracks irreducible representations alongside tensor data.

Yes. The package is actively maintained, has low install friction, carries a permissive license, and fills a specific niche—equivariant neural networks in JAX with explicit irreps tracking. Install it if you need to build or experiment with E(3)-equivariant models; skip it if you don't work with 3D geometric data or prefer PyTorch.

Install

e3nn-jax on PyPI

pip install e3nn-jax

uv add e3nn-jax

poetry add e3nn-jax

Installing e3nn-jax

Before you install

Low friction installation with a pure Python wheel. Active maintenance as of April 2026 with recent commits. Supports Python 3.9, 3.10, and 3.11. Runtime dependencies are well-established scientific libraries (jax, jaxlib, numpy, sympy, attrs).

License in practice

Apache License 2.0 permits commercial and derivative use with attribution. No restrictions on modification or distribution, making it suitable for both research and production applications.

Quickstart

pip install e3nn-jax

import e3nn_jax as e3nn
import jax

array = e3nn.normal("0e + 1o", jax.random.PRNGKey(0))
norms = e3nn.norm(array)
tensor_product = e3nn.tensor_square(array)

Requires JAX and jaxlib, which may need GPU/TPU setup depending on your hardware; jaxlib installation can be non-trivial on some systems.

Verify before relying

• Whether the 44% speed advantage over PyTorch holds across different model architectures and hardware beyond the MACE/revMD-17 benchmark cited.
• Maturity and stability guarantees for production use in molecular dynamics or other scientific domains.
• Whether sympy is a runtime dependency or only needed for specific symbolic operations.

Package facts

Evidence: e3nn_jax-0.21.0-py3-none-any.whl

Tags

More Artificial Intelligence packages