[Paper Review] N-body Networks: a Covariant Hierarchical Neural Network Architecture for Learning Atomic Potentials

Risi Kondor|arXiv (Cornell University)|Mar 5, 2018

Machine Learning in Materials ScienceMaterials Science17 references68 citations

TL;DR

Introduces N-body networks, a covariant, hierarchical neural architecture for learning atomic potential energy surfaces, ensuring SO(3) rotation covariance via irreducible representations and Clebsch–Gordan decompositions, with Fourier-domain operations.

ABSTRACT

We describe N-body networks, a neural network architecture for learning the behavior and properties of complex many body physical systems. Our specific application is to learn atomic potential energy surfaces for use in molecular dynamics simulations. Our architecture is novel in that (a) it is based on a hierarchical decomposition of the many body system into subsytems, (b) the activations of the network correspond to the internal state of each subsystem, (c) the "neurons" in the network are constructed explicitly so as to guarantee that each of the activations is covariant to rotations, (d) the neurons operate entirely in Fourier space, and the nonlinearities are realized by tensor products followed by Clebsch-Gordan decompositions. As part of the description of our network, we give a characterization of what way the weights of the network may interact with the activations so as to ensure that the covariance property is maintained.

Motivation & Objective

• Motivate the need for accurate, scalable learning of atomic potential energy surfaces in molecular dynamics.
• Propose a hierarchical, compositional neural architecture that models subsystems with explicit rotational covariance.
• Develop a representation-theoretic framework to guarantee SO(3) covariance throughout the network.
• Provide practical aggregation gates that preserve symmetry while enabling learning of multiscale interactions.

Proposed method

• Introduce a composition scheme where nodes carry physical subsystems and internal states that transform covariantly under rotations.
• Define SO(3)-covariant vectors and isotypic components via group representations (irreps) and Clebsch–Gordan decompositions.
• Formulate a polynomial, covariance-preserving aggregation function Phi that expresses outputs as sums over irreducible fragments phi^ell_m with linear mixing via matrices W^ell.
• Show that the network’s nonlinearities are realized through tensor products and Clebsch–Gordan transforms computed in Fourier (frequency) space, avoiding time-domain nonlinearities.
• Propose zeroth- and first-order interaction gates to efficiently capture local and pairwise interactions among child subsystems while maintaining covariance.
• Describe how weights W^ell are shared and learned, while the phi^ell_m fragments are formed from tensor products of child states and relative positions.

Experimental results

Research questions

• RQ1 How can one design neural networks for many-body physical systems that are exactly covariant to spatial rotations (SO(3))?
• RQ2 What aggregation rules (Phi) preserve SO(3) covariance when combining subsystems at multiple hierarchical levels?
• RQ3 How can representation theory (irreps, Clebsch–Gordan decompositions) be integrated into neural network architectures to enable efficient, covariant learning of atomic potentials?
• RQ4 What are practical, scalable gating mechanisms (zeroth and first order) that capture local and pairwise interactions while respecting symmetry constraints?

Key findings

• Proposes SO(3)-covariant N-body networks where each node represents a physical subsystem with position and an internal state transforming under SO(3).
• Shows that aggregations must decompose into irreducible fragments and that mixing is restricted to fragments of the same angular momentum level (ell) for covariance (Proposition 1).
• Demonstrates that nonlinearities can be realized via tensor products in Fourier space using Clebsch–Gordan transforms, eliminating time-domain nonlinearities as a bottleneck.
• Introduces practical gate types (zeroth and first order) to model multiscale interactions without sacrificing rotational covariance.
• Provides a formal framework connecting covariant neural architectures with group representation theory, enabling learnable, multiscale representations of atomic environments.