Unified theory of atom-centered representations and graph convolutional machine-learning schemes

Jigyasa Nigam, Guillaume Fraux, Michele CeriottiarXiv 2022 · arXiv

Introduction

In machine learning for chemistry and materials science, two primary paradigms have emerged: atom-centered representations and message-passing schemes. Atom-centered descriptors, such as SOAP or ACE, describe the local environment of an atom through systematic expansions of density. While mathematically transparent and physically symmetric, they are inherently local. Conversely, message-passing neural networks (MPNNs), like SchNet or DimeNet, treat molecules as graphs where information flows between nodes. These are highly performant but often criticized as "black boxes."

This paper introduces a unified mathematical framework that bridges these two approaches. By extending the Atom-Centered Density Correlation (ACDC) framework, the authors show that message-passing is essentially a computationally efficient way to build multi-centered, high-body-order representations.

Methods: The ACDC Framework

The core of the proposed theory is the representation of atomic arrangements as a sum of localized functions. To manage the exponential cost of many-body correlations, the authors employ a "density trick," first computing local density and then taking tensor products.

Conceptual unification of local atom-centered environments and message-passing schemes.

The key innovation is the "decoration" of these densities. Instead of merely considering the immediate neighbors of an atom, the framework incorporates information about the neighbors' own environments recursively. Using Clebsch-Gordan (CG) iterations, the authors ensure that these features remain equivariant under rotation and reflection. This recursive construction demonstrates that the "messages" in a graph neural network can be mapped directly to specific contractions of decorated ACDC features, providing a complete linear basis for symmetric functions of atomic coordinates.

Key Results and Discussion

The framework was validated across several datasets, including random methane ($CH_4$), bulk sodium chloride ($NaCl$), and the QM7/QM9 organic molecule sets.

Convergence and Efficiency

A primary finding is that message-passing features converge significantly faster than single-center features for a fixed basis size. In tests on $CH_4$, multi-center message-passing features achieved a 7% error, whereas standard 3-body descriptors (like the bispectrum) saturated at 16%.

Learning curves showing the convergence of different body-order representations.

The "Long-Range" Myth

The study challenges the common belief that message-passing is the optimal solution for long-range physics. For pure electrostatic interactions in $NaCl$, the authors found that simply increasing the local cutoff radius was more effective than adding more message-passing steps. This suggests that while MPNNs capture multi-center correlations well, they are not a silver bullet for long-range decay.

Extrapolation and Covariance

The linear MP-ACDC model trained on the QM7 dataset showed superior extrapolative accuracy on the larger QM9 molecules compared to previous state-of-the-art non-linear models. Furthermore, the framework naturally handles covariant properties, such as dipole moments, by tracking irreducible representations during the CG iterations.

Conclusion

By unifying atom-centered and graph-convolutional schemes, this work provides a rigorous foundation for designing more principled machine-learning architectures. It clarifies that scalar non-linearities in neural networks serve as a "cheap" way to reach high body-order, but only in angularly blind channels. Future work will focus on optimizing the balance between angularly-resolved linear features and these non-linearities, as well as addressing the computational complexity of full tensor products through feature selection techniques like PCA.

Summary updated

Abstract

Data-driven schemes that associate molecular and crystal structures with their microscopic properties share the need for a concise, effective description of the arrangement of their atomic constituents. Many types of models rely on descriptions of atom-centered environments, that are associated with an atomic property or with an atomic contribution to an extensive macroscopic quantity. Frameworks in this class can be understood in terms of atom-centered density correlations (ACDC), that are used as a basis for a body-ordered, symmetry-adapted expansion of the targets. Several other schemes, that gather information on the relationship between neighboring atoms using "message-passing" ideas, cannot be directly mapped to correlations centered around a single atom. We generalize the ACDC framework to include multi-centered information, generating representations that provide a complete linear basis to regress symmetric functions of atomic coordinates, and provides a coherent foundation to systematize our understanding of both atom-centered and message-passing, invariant and equivariant machine-learning schemes.
BibTeX
@misc{nigam2022unified,
  author = {Jigyasa Nigam and Guillaume Fraux and Michele Ceriotti},
  title = {{Unified theory of atom-centered representations and graph convolutional machine-learning schemes}},
  year = {2022},
  eprint = {2202.01566},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.01566},
}
Figure