Modeling Long Range Interactions in Atomistic Machine Learning

This direction is merged into another direction: long-range interaction modeling. Its last version stays readable and citable here.

Version 1 of 119 papers23 conceptsfrozen 2026-10-04

In the quest to simulate molecules and materials with quantum-mechanical accuracy, researchers have hit a fundamental wall: the "locality assumption." Most modern machine learning interatomic potentials (MLIPs) operate like a nearsighted observer, only considering the immediate neighborhood of an atom—typically within a 5 to 10 Ångström cutoff. While this makes simulations computationally efficient, it ignores the "invisible threads" that hold complex systems together: long-range electrostatics, van der Waals forces, and London dispersion.

The research direction emerging across these papers focuses on breaking this locality barrier to capture the global physics of large-scale molecular assemblies.

The Locality Bottleneck

The success of early ML models for chemistry relied on the principle of "nearsightedness"—the idea that the energy of an atom is primarily determined by its closest neighbors. As Michael J. Willatt and Gábor Cśanyi demonstrate in Machine Learning of Atomic-Scale Properties Based on Physical Principles, using local descriptors like Smooth Overlap of Atomic Positions (SOAP) allows models to respect physical symmetries like rotation and translation invariance.

However, this nearsightedness is a double-edged sword. Marcel F. Langer and Matthias Rupp point out in their comprehensive review that local representations inherently fail to account for non-local physical effects. When simulating a large protein or a complex crystal surface, an atom is influenced by the charge distribution of atoms dozens of Ångströms away. Without these long-range forces, models cannot accurately predict the behavior of large-scale biological systems or the stability of complex materials.

Why Local Models Fail the Global Test

The problem is not just a lack of data, but a structural limitation of the architectures. As noted by John A. Keith and Alexandre Tkatchenko in Combining Machine Learning and Computational Chemistry for Predictive Insights Into Chemical Systems, local descriptors overlook the "dispersion" forces that govern how molecules stack and interact.

In the realm of crystal growth, Yuan Gao 0004 and Jianfeng Lu 0001 show that long-range elastic interactions create dependencies where the motion of a single atomic step is influenced by every other step on the surface. This creates a "non-local" dependency that traditional partial differential equations struggle to aggregate into a single field operator.

Strategies for Non-Local Integration

To solve this, researchers are moving toward architectures that can "see" further without losing the efficiency of local models. Several distinct strategies have emerged:

  1. Iterated Equivariant Representations: Instead of simple message passing, which can be slow to propagate information across a system, the Allegro architecture by Albert Musaelian and Simon Batzner uses iterated tensor products of local representations. This allows the model to scale to 100 million atoms while maintaining the precision needed for atomistic dynamics.
  2. Global Receptive Fields: In macroscopic simulations like cloth modeling, Ruochen Chen and colleagues utilize Fourier Neural Operators to create global receptive fields. This allows the model to account for constraints that affect the entire structure simultaneously, ensuring stability in stiff simulations.
  3. Multi-scale Kernels and Non-local Operators: For surface tension and phase transitions, Juan Pablo Borthagaray and Ricardo H. Nochetto have developed finite element frameworks for nonlocal minimal graphs. These frameworks replace local curvature with energy functionals that depend on interactions between distant points.

The Path Forward: Beyond the Cutoff

The consensus across recent work, including the benchmarking of E(3)-equivariant potentials by Ilyes Batatia and Gabor Csanyi, is that the next generation of ML potentials must integrate specialized modules for long-range interactions. Whether through "learning-on-the-fly" as proposed by David Holzmüller or transfer learning from "gold standard" quantum data as explored by Ben Nebgen, the goal is the same: a unified model that respects both the immediate chemical bond and the distant electrostatic pull.

By bridging the gap between the local and the global, these researchers are enabling the simulation of complex, multicomponent systems—from biological membranes reshaped by proteins to the large-scale evolution of metal alloys—with a level of physical fidelity previously thought impossible.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_modeling_long_range_interactions_in_atom_9a6b22c5,
  title = {Modeling Long Range Interactions in Atomistic Machine Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-long-range-interactions-in-atomistic-machine-learning/rdr_9a6b22c5acba4fcba37ca9973a44429d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_modeling_long_range_interactions_in_atom_9a6b22c5_v1,
  title = {Modeling Long Range Interactions in Atomistic Machine Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/modeling-long-range-interactions-in-atomistic-machine-learning/rdr_9a6b22c5acba4fcba37ca9973a44429d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}