Equivariant representations for molecular Hamiltonians and N-center atomic-scale properties

Jigyasa Nigam, Michael J. Willatt, Michele CeriottiThe Journal of Chemical Physics 2021 · DOI · arXiv

Abstract

Symmetry considerations are at the core of the major frameworks used to provide an effective mathematical representation of atomic configurations that is then used in machine-learning models to predict the properties associated with each structure. In most cases, the models rely on a description of atom-centered environments and are suitable to learn atomic properties or global observables that can be decomposed into atomic contributions. Many quantities that are relevant for quantum mechanical calculations, however—most notably the single-particle Hamiltonian matrix when written in an atomic orbital basis—are not associated with a single center, but with two (or more) atoms in the structure. We discuss a family of structural descriptors that generalize the very successful atom-centered density correlation features to the N-center case and show, in particular, how this construction can be applied to efficiently learn the matrix elements of the (effective) single-particle Hamiltonian written in an atom-centered orbital basis. These N-center features are fully equivariant—not only in terms of translations and rotations but also in terms of permutations of the indices associated with the atoms—and are suitable to construct symmetry-adapted machine-learning models of new classes of properties of molecules and materials.

Introduction

In the last decade, machine learning (ML) has revolutionized chemistry by predicting molecular energies and interatomic potentials with high accuracy. Most current models rely on "atom-centered" descriptors—mathematical snapshots of the environment immediately surrounding a single atom. While this approach is highly effective for scalar properties like total energy, which can be decomposed into atomic contributions, it struggles with properties that are intrinsically multi-center.

A prominent example is the Hamiltonian matrix ($H$). In quantum mechanics, the elements of this matrix describe interactions between orbitals on different atoms. To predict $H$ accurately, an ML model must understand the relationship between two or more centers. Furthermore, the model must be equivariant: if the molecule is rotated or the labels of identical atoms are swapped, the predicted Hamiltonian matrix must transform in a mathematically consistent way. This work introduces a framework to address these challenges using $N$-center structural descriptors.

The N-Center Representation

The authors introduce a family of $N$-center structural descriptors that build on the successful "density correlation" framework (such as SOAP) but extend it to multiple tagged centers. By "tagging" $N$ specific atoms, they define a property's frame of reference that accounts for the relative positions and environments of all involved centers.

Schematic construction of the N-center equivariant representation showing the tensor product of atomic orbitals and density correlations.

Using the N-body Iterative Contraction (NICE) framework, the authors systematically construct features that are equivariant to translations, rotations ($O(3)$), and permutations of the atom indices. This mathematical generalization shows that $N$-center features are not just a new tool, but a formal extension of standard 1-center descriptors; summing over the indices of an $N$-center representation collapses it back into a $(N-1)$-center representation of higher body-order.

Symmetry-Adapted Learning

Because the Hamiltonian is a matrix of tensors, it can be decomposed into irreducible representations (irreps) of the rotation group. The framework trains separate, simple linear or kernel regression models for different "blocks" of the matrix, such as the interaction between an Oxygen $s$-orbital and a Hydrogen $p$-orbital.

Decomposition of the molecular Hamiltonian into symmetry-adapted diagonal, off-diagonal, and cross-species blocks.

To manage the complexity of large molecules, the authors utilize a Symmetry-Adapted Projected Hamiltonian (SAPH) approach. By projecting the electronic structure onto a minimal, physically motivated atomic basis, the number of required ML models remains manageable without losing essential physical information. This allows the model to treat symmetry as a "data multiplier," requiring significantly less training data to achieve high accuracy.

Key Results and Insights

The framework was validated on datasets including water, ethanol, and the QM7-CHNO database. Key findings include:

  • High Accuracy: On the water dataset, simple linear ridge regression achieved a Root Mean Square Error (RMSE) of less than 10 meV for Hamiltonian elements, outperforming many complex deep-learning architectures.
  • Eigenvalue Prediction: The models accurately reproduce energy levels. For water, the error in the HOMO-LUMO gap was approximately 1 meV. For the diverse QM7 dataset, the eigenvalue RMSE reached 0.14 eV when predicting the difference from an initial guess Hamiltonian.
  • Zero-Shot Symmetry: In a striking demonstration, a model trained on a single random benzene configuration correctly predicted the orbital symmetries and degeneracies required by group theory, proving the robustness of the equivariant construction.

Conclusion

While the framework provides a rigorous path toward learning multi-center quantum properties, challenges remain. For larger molecules like ethanol, the accuracy of 3-body features eventually plateaus, suggesting that higher-order correlations may be necessary for subtle electronic effects. Future work will focus on automating the management of orbital blocks and exploring 3-center integrals to further advance quantum simulations.

Summary updated

Abstract

Symmetry considerations are at the core of the major frameworks used to provide an effective mathematical representation of atomic configurations that is then used in machine-learning models to predict the properties associated with each structure. In most cases, the models rely on a description of atom-centered environments and are suitable to learn atomic properties or global observables that can be decomposed into atomic contributions. Many quantities that are relevant for quantum mechanical calculations, however-most notably the single-particle Hamiltonian matrix when written in an atomic orbital basis-are not associated with a single center, but with two (or more) atoms in the structure. We discuss a family of structural descriptors that generalize the very successful atom-centered density correlation features to the N-center case and show, in particular, how this construction can be applied to efficiently learn the matrix elements of the (effective) single-particle Hamiltonian written in an atom-centered orbital basis. These N-center features are fully equivariant-not only in terms of translations and rotations but also in terms of permutations of the indices associated with the atoms-and are suitable to construct symmetry-adapted machine-learning models of new classes of properties of molecules and materials.
BibTeX
@article{nigam2021equivariant,
  author = {Jigyasa Nigam and Michael J. Willatt and Michele Ceriotti},
  title = {{Equivariant representations for molecular Hamiltonians and <i>N</i>-center atomic-scale properties}},
  journal = {The Journal of Chemical Physics},
  year = {2021},
  doi = {10.1063/5.0072784},
}
Figure