Reproduce “E(n) Equivariant Normalizing Flows” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Victor Garcia Satorras, Emiel Hoogeboom, Fabian B. Fuchs, Ingmar Posner, Max Welling

E-NF uses an E(n)-equivariant graph network to define an ODE whose integration maps molecular positions and features invertibly to a simple latent distribution. It treats positions equivariantly and atom features invariantly; discrete-feature likelihood is trained as a lower bound. On QM9, its joint negative log-likelihood is −59.7 versus −29.3 for a rotation-augmented baseline, while 4.9% of generated molecules pass a distance-based valence test.

Key insights

  • E(n) equivariance, when integrated as an inductive bias within generative models, consistently improves performance for 3D data by ensuring generated samples inherently respect geometric symmetries and more accurately capture data distributions.
  • Continuous-time normalizing flows provide a flexible and expressive framework for integrating equivariant dynamics, allowing for both exact likelihood computation and controlled, high-quality generation of complex 3D structures.
  • Lifting discrete node features into a continuous space and modeling them within the flow enables a single generative model to comprehensively generate both 3D coordinates and associated invariant properties of molecules simultaneously.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.