Source-linked AI summary
Incorporating long-range physics in atomic-scale machine learning
Andrea Grisafi, Michele Ceriotti
TL;DR
Local atom-centered models struggle with long-range, non-local effects because finite environments cannot capture slowly decaying electrostatics and related physics. The paper introduces locally defined, O(3)-equivariant features encoding non-local information, and shows that this approach accurately models long-range electrostatic and dielectric properties while retaining additive-model transferability.
Problem
Finite-cutoff atom-centered representations limit accuracy for long-range electrostatics and other intrinsically non-local properties.
Method
The paper remaps non-local system representations into locally defined, symmetry-equivariant atom-centered features with electrostatic far-field asymptotics, exemplified by the LODE framework.
Results
The framework successfully reproduces Coulomb interactions, charged-molecular-dimer binding curves, and liquid-water dielectric response within additive machine-learning models.
Takeaways & Limitations
Combining long-range-sensitive features with atom-centered additive models provides a general framework for incorporating non-local physics into atomistic machine learning.
Takeaways & Limitations
Existing approaches that capture long-range effects without target-specific assumptions often use global representations and are difficult to transfer across systems with different physical character.
Abstract
from arXiv · showhide
The most successful and popular machine learning models of atomic-scale properties derive their transferability from a locality ansatz. The properties of a large molecule or a bulk material are written as a sum over contributions that depend on the configurations within finite atom-centered environments. The obvious downside of this approach is that it cannot capture non-local, non-additive effects such as those arising due to long-range electrostatics or quantum interference. We propose a solution to this problem by introducing non-local representations of the system that are remapped as feature vectors that are defined locally and are equivariant in O(3). We consider in particular one form that has the same asymptotic behavior as the electrostatic potential. We demonstrate that this framework can capture non-local, long-range physics by building a model for the electrostatic energy of randomly distributed point-charges, for the unrelaxed binding curves of charged organic molecular dimers, and for the electronic dielectric response of liquid water. By combining a representation of the system that is sensitive to long-range correlations with the transferability of an atom-centered additive model, this method outperforms current state-of-the-art machine-learning schemes, and provides a conceptual framework to incorporate non-local physics into atomistic machine learning.
INTRODUCTION
Atom-centered local models are transferable and efficient, but finite cutoffs limit their ability to represent long-range, non-local effects. Existing remedies often impose system-dependent physical assumptions or use global representations.
- Local atomistic models decompose properties into contributions from finite-radius atom-centered environments, supporting transferability across atomic arrangements.
- Coulomb interactions decay as ∼1/r, making convergence with local machine-learning schemes virtually impossible and limiting achievable accuracy.
- Long-range electronic energies are commonly handled by separating local many-body contributions from classical electrostatics modeled with Ewald-like energies, learned charges, multipoles, or charge equilibration.
- Dielectric response is intrinsically non-local because it depends on far-field electrostatics and the topological quantum nature of macroscopic polarization.
- Existing approaches are often system dependent and difficult to transfer to charge-transfer or near-metallic polarizability settings, while explicitly non-local methods commonly use global representations.
- The paper proposes symmetry-equivariant non-local information embedded in atom-centered features, with electrostatic asymptotics, and demonstrates it on three long-range prediction tasks.
LONG-DISTANCE EQUIVARIANT REPRESENTATION
The LODE framework transforms an atom density into a non-local potential representation, then symmetrizes it into locally defined equivariant descriptors. For p = 1, the representation has electrostatic asymptotic behavior and can encode interactions beyond the local cutoff.
- The starting representation is a decorated atom density built from localized functions centered at atomic positions and abstract species vectors.
- Applying an atom-density potential transformation makes each local representation depend on all atoms, with far-field contributions decaying as |r − r_i|^-p.
- Translation, rotation, and inversion symmetrization produce atom-centered descriptors that retain long-range information while supporting an additive property model.
- The LODE framework focuses on p = 1, which corresponds to electrostatic interactions and gives the atom-density potential the asymptotic form of an electrostatic potential.
- In the Dirac-δ first-invariant limit, the representation sums 1/r_ij terms from atoms outside the local region, equivalent to a fixed point-charge electrostatic model.
- Higher-order invariants increase structural information; the ν = 2 rotationally invariant form is equivalent to the SOAP power spectrum and is used in applications.
- Higher spatial-correlation orders and rotationally covariant tensor representations are straightforward extensions, while combinations of different potential exponents remain future work.
Efficient evaluation of the LODE representation
For bulk systems, LODE’s long-range integral is evaluated efficiently by expanding the potential in a plane-wave basis. This separates local geometric descriptors from global, system-dependent potential information.
- Direct evaluation of the long-range integral is prohibitive for periodically repeated bulk systems, paralleling the difficulty of condensed-phase electrostatics.
- A plane-wave auxiliary basis enables efficient evaluation of local potential projections by inserting a plane-wave identity resolution.
- The basis projections can be computed analytically for Gaussian-type radial functions, while Fourier components of the full-system Gaussian potential are also analytically accessible.
- The resulting factorization separates the geometric local representation from its system-dependent global character.
RESULTS
The study evaluates LODE for scalar electrostatic properties using Gaussian-process regression and compares it with SOAP and a SOAP–LODE range-separated model. Figure 1 specifically reports learning curves for a random point-charge gas.
- LODE is tested with Gaussian-process regression and simple polynomial kernels so feature quality, rather than regression complexity, is emphasized.
- SOAP provides the baseline because of its close relation to LODE and strong prior performance, while SOAP–LODE combines short- and long-range representations.
- Figure 1 compares local SOAP learning curves at 3, 6, and 9 Å cutoffs with LODE configurations using 2–3 Å cutoffs and different smearing or invariant orders.
- The point-charge benchmark uses 1500 training configurations and 500 independent test configurations.
A gas of point charges
The point-charge experiment shows that local representations struggle with electrostatic energy, while LODE features efficiently capture the system’s long-range interactions.
- A local model is inefficient for learning electrostatic energy dominated by long-range effects.
- 20% RMSE: SOAP with rcut=9 Å barely reaches this accuracy using the maximum training set.
- Below 1% error: linear LODE(ν = 1) achieves this performance with a handful of training points.
- LODE(ν = 2) initially performs worse but eventually reaches and surpasses the accuracy of linear LODE(ν = 1), σ = 1 Å.
- Gaussian smearing prevents the LODE(ν = 1) error from converging to zero, while halving σ dramatically reduces the error.
Binding curves of charged dimers
For charged organic dimers, SOAP captures short-range interactions but loses accuracy at larger separations, whereas SOAP+LODE(ν = 2) predicts binding curves across the full distance range.
- 661 charged organic dimers containing H, C, N and O were extracted from the BioFragment Database.
- The reference DFT energies were baselined against monomer energies so the model learned interaction energies between fragments.
- ∼20% RMSE: optimal SOAP performance, versus ∼4% for a suitable SOAP+LODE(ν = 2) combination.
- SOAP accurately captures short-range interactions but becomes ineffective when intermolecular distances exceed its environment cutoff.
- SOAP+LODE(ν = 2) accurately predicts binding curves across distances, including monopole, polarization, and charge-dipole interactions.
- The scheme does not transparently represent polarization or charge transfer because distant fields lack explicit dependence on neighboring structure.
Dielectric response of liquid water
For liquid water’s dielectric response, combining local SOAP with non-local LODE captures the multiscale character of the isotropic response more effectively than either description alone.
- The study learns the isotropic component ε0 = Tr[ε∞], using 800 training structures and 200 independent test configurations.
- LODE performs much better than SOAP at rcut=3 Å for the isotropic dielectric response.
- SOAP improves substantially as its cutoff increases, eventually surpassing LODE at rcut=6 Å.
- Combining SOAP and LODE at rcut=3 Å gives optimal predictions by incorporating fine-grained local and coarse-grained non-local information.
- The results highlight that ε0 has a multiscale character requiring both local many-body information and long-range electrostatic effects.
CONCLUSIONS
The paper extends transferable atom-centered machine-learning models with LODE features that incorporate global structural information into local, symmetry-consistent representations. Across three long-range tasks, LODE captures physics that local models miss, while the authors identify several directions requiring further investigation.
- CONCLUSIONS: LODE folds global information about system structure and composition into local representations with physically motivated asymptotic behavior and symmetry-consistent computation.The representation is designed to retain the transferability of additive atom-centered models while incorporating long-range physics.
- CONCLUSIONS: LODE, alone or combined with SOAP, outperforms local machine-learning methods for point-charge electrostatic energies, charged-organic-dimer binding curves, and bulk-water dielectric constants.These applications test long-range physics across electrostatic energy, molecular binding, and condensed-phase dielectric response.
- CONCLUSIONS: The reported examples establish an efficient assay of long-range information incorporation, but systematic links to long-range interatomic potentials and extensions such as dispersion and polarizability remain open.The authors also propose investigating alternative exponents, combined equivariant expansions, SOAP–LODE force fields, and computational efficiency.