Welcome to e3nn!

This is the website for the e3nn repository
    https://github.com/e3nn/e3nn/

E(3) is the Euclidean group in dimension 3. That is the group of rotations, translations and mirror. e3nn is a pytorch library that aims to create E(3) equivariant neural networks.

Getting Started

How to use the Resources

If you’d like to generally learn what e3nn is and what it’s used for, check out some of previously recorded talks and skimming some of the papers. If you’d like to try out e3nn, the first step is to install it.

Once you have e3nn up and running, try out the Jupyter notebooks in e3nn_tutorial. This will give you a feel for the primary data types and classes in e3nn. You can also read e3nn_book in parallel and dive into more rigorous math resources. This reading will be helpful for parsing even the most techincal parts of the relevant papers.

If you want to talk to other folks using e3nn, you are most welcome to join our recurring monthly meetings which are shown in the e3nn calendar and our Slack.

We want to help eliminate any bottlenecks in making Euclidean equivariant neural networks and e3nn accessible to a broad audience, especially scientists whose primary expertise is not machine learning. If you have an application in mind but not sure how to structure a network around it or other questions not satisfied by this page’s resources, feel free to reach out to Tess (tsmidt at lbl gov).

Installation

Installing e3nn and dependencies if you are only going to use the cpu is straightforward. Things get a bit more complicated if you want to GPU. If you want to use GPUs with e3nn you need to:

• be careful to install a version of torch_geometric that matches the same CUDA version that your torch installation uses (this is general advice and not e3nn specific)
• AND make sure you have the appropriate NVIDIA compilers (nvcc) for your CUDA installation so you can compile e3nn CUDA kernels.

The full instructions for installing e3nn can be found here and here.

Resources

Math that’s good to know

e3nn makes use of group theory and representation theory. You don’t need to be knowledgeable on these topics to start using e3nn but it might be useful to have some relevant resources handy. Some of our favorite resources are:

• Group Theory: Application the Physics of Condensed Matter by Dresselhaus, Dresselhaus, and Jorio
• Lie Algebras in Particle Physics by H. Georgi
• Lecture Note on Group Theory in Physics by D. Arovas
• Group Theory in a Nutshell for Physicists by A. Zee
• Linear Representations of Finite Group by J-P. Serre

e3nn_tutorial

This link has our quick tutorial and this website / respository has many illustrative notebooks on how to use e3nn. If you run into an error while running these notebooks – please make a pull request!

e3nn_book [in progress]

We are currently in the progress of compiling a more user-friendly introduction to the core concepts used in e3nn. You can see the (very incomplete) draft of the book here.

Previous Talks

Some previous recorded talks on e3nn.

2020/09 Unintended Features of E(3)NNs, Workshop on Equivariance and Data Augmentation (video // slides), University of Pennsylvania, September 4, 2020

2020/09 Lecture on Symmetry and Equivariance in ML, Berkeley Lab Deep Learning School (video // slides), Berkeley Lab, September 3, 2020.

2020/07 Neural Networks with Euclidean Symmetry for Physical Sciences, 1st Workshop on Scientific-Driven Deep Learning (SciDL) (video // slides), July 1, 2020

2020/01 An autoencoder for discrete geometry, Applied Machine Learning Days – AI and the Molecular World (video), EPFL, January 27-28, 2020, Lausanne, Switzerland

2019/12 Euclidean Neural Networks for Emulating Ab Initio Calculations and Generating Atomic Geometries, eScience Institute Seminar (video), University of Washington, Seattle, WA

2019/09 Euclidean Neural Networks for Emulating Ab Initio Calculations and Generating Atomic Geometries, Workshop I: From Passive to Active: Generative and Reinforcement Learning with Physics, (video // slides), IPAM at UCLA, Los Angeles, CA

Poster

A poster overview of the e3nn framework.

Papers

(in reserve chronological order)

• 2020/10 On the Universality of Rotation Equivariant Point Cloud Networks
  • Nadav Dym, Haggai Maron
  • Proves the expressivity of rotation equivariant neural networks.
• 2020/09 Direct prediction of phonon density of states with Euclidean neural network
  • Zhantao Chen, Nina Andrejevic, Tess Smidt, Zhiwei Ding, Yen-Ting Chi, Quynh T. Nguyen, Ahmet Alatas, Jing Kong, Mingda Li
  • e3nn is used to predict phonon density of states (DOS) from crystal structure. Trained network is used to identify materials with high specific heat.
• 2020/09 Euclidean Symmetry and Equivariance in Machine Learning
  • Tess E. Smidt
  • A mini review on invariant vs. equivariant ML models.
• 2020/08 Relevance of Rotationally Equivariant Convolutions for Predicting Molecular Properties
  • Benjamin Kurt Miller, Mario Geiger, Tess E. Smidt, Frank Noé
  • Includes benchmark of e3nn on QM9.
• 2020/07 Finding Symmetry Breaking Order Parameters with Euclidean Neural Networks (code)
  • Tess E. Smidt, Mario Geiger, Benjamin Kurt Miller
  • Demonstrates how to use e3nn to perform a symmetry analysis to resolve order parameters for 2nd order phase transitions.
• 2020/06 Hierarchical, rotation-equivariant neural networks to predict the structure of protein complexes
  • Stephan Eismann, Raphael J.L. Townshend, Nathaniel Thomas, Milind Jagota, Bowen Jing, Ron Dror
• 2020/06 SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks
  • Fabian B. Fuchs, Daniel E. Worrall, Volker Fischer, Max Welling
• 2019/06 Cormorant
  • Brandon Anderson, Truong-Son Hy, Risi Kondor
  • Introduces n-body convolutions.
• 2018/07 3D Steerable CNNs*
  • Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, Taco Cohen
• 2018/06 Clebsch-Gordan Networks*
  • Risi Kondor, Zhen Lin, Shubhendu Trivedi
• 2018/02 Tensor Field Networks* (code)
  • Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, Patrick Riley

*indicates foundational paper in the development of Euclidean neural networks.

Related work

• 2019/11 General E(2)-Equivariant Steerable CNNs
  • Maurice Weiler, Gabriele Cesa
• 2016/12 Harmonic Networks: Deep Translation and Rotation Equivariance
  • Daniel E. Worrall, Stephan J. Garbin, Daniyar Turmukhambetov, Gabriel J. Brostow
  • SE(2) Equivariant networks

If there are any papers that you think should be on this list but are missing, please email Tess tsmidt at lbl gov with a citation and description that matches the examples above.

Slack

The e3nn developers and several collaborators discuss ideas and help each other out with projects via Slack. If you’d like to join the Slack, please send an email to Tess tsmidt at lbl gov.

Recurring Meetings / Events

e3nn has two recurring monthly meetings in addition to other events on our calendar.

• A monthly developers / collaboration 1.5 hour meeting on the first Wednesday of every month (typically) at 10 am Pacific Time. We typically go around and each give a brief update on what we’ve been working on and if there are any difficulties we are running into.
• A monthly tutorials / documentation hackathon on the second Wednesday of every month starting at 10 am Pacific Time and ending around 6 pm Pacific Time. This is a great opportunity to ask questions.

Everyone is welcome to join these meetings – yes, that means you! The meeting links are in the calendar events show below. Feel free to reach out to Tess tsmidt at lbl gov to introduce yourself if you are new to joining the meetings.

Calendar

with Upcoming Tutorials / Meetings / Hackathons

e3nn Team

Core-development team

(aka the people answering pull-requests)

• Mario Geiger (e3nn’s BDFL)
• Tess Smidt (tsmidt at lbl gov)
• Ben Miller
• Kostiantyn Lapchevskyi

Collaborators and Contributors

Josh Rackers Thomas Hardin Simon Batzner Boris Kozinsky Eugene Kwan

Frank Noé Claire West Zhantao Chem Nina Andrejevic Mingda Li