Universal Foundation Models for Atomistic Simulation

This direction is merged into another direction: cross-domain molecular generalization. Its last version stays readable and citable here.

Version 1 of 163 papers71 conceptsfrozen 2026-10-04

For decades, computational chemistry has been trapped in a trade-off between the "gold standard" accuracy of quantum mechanics and the speed of classical simulations. While Density Functional Theory (DFT) provides the precision needed to understand chemical reactions, its cubic scaling with the number of atoms makes simulating large proteins or complex materials prohibitively slow.

The recent emergence of Large Atomistic Models (LAMs)—often called "foundation models for chemistry"—is breaking this bottleneck. Instead of training a new model for every specific molecule or crystal, researchers are building universal architectures that learn the fundamental physics of atomic interactions across the entire periodic table.

From Specialized Potentials to Generalist Backbones

Traditionally, machine learning in chemistry followed a "one model, one system" paradigm. If you wanted to simulate a specific alloy, you generated DFT data for that alloy and trained a specialized potential. However, these models were notoriously brittle; a model trained on liquid water would fail if it encountered a single solute it hadn't seen before.

The shift toward foundation models, such as DPA-2 and UMA, treats atomistic simulation as a large-scale pre-training task. By exposing a single model to hundreds of millions of diverse atomic environments—ranging from small organic molecules to crystalline semiconductors and metallic catalysts—the model internalizes the universal rules of the potential energy surface (PES).

As Nima Shoghi and Adeesh Kolluru demonstrate with the Joint Multi-domain Pre-training (JMP) framework, training on diverse data allows models to generalize across domains. A model that understands how carbon bonds in a small molecule can transfer that knowledge to understand carbon in a bulk diamond or a graphene sheet.

Why It Works: The Universality of Local Environments

The core intuition behind these models is that chemistry is fundamentally local. An atom's energy and the forces acting upon it are primarily determined by its immediate neighbors. While there are infinite possible molecules, the number of ways atoms can be locally arranged is finite and recurring.

By using equivariant architectures—which ensure that the model’s predictions don't change just because a molecule is rotated or translated—models like DeepPot-SE by Linfeng Zhang 0002 and Han Wang 0006 can map these local "environments" to energy contributions.

When a model is trained on a massive scale, it stops memorizing specific structures and starts learning a "chemical grammar." For example, CHGNet incorporates electronic degrees of freedom like oxidation states, allowing it to represent the "charge-aware" physics that governs battery materials and minerals. This allows for zero-shot generalization: the ability to predict the properties of a completely new material without any system-specific training data.

Scaling Laws and Cross-Domain Synergy

A striking finding across this research is that "more is different." Just as Large Language Models gain new capabilities with scale, atomistic models exhibit clear scaling laws. Duo Zhang and colleagues showed in DPA3 that as model size and data diversity increase, the error in energy and force predictions drops consistently.

Furthermore, there is a powerful synergy between disparate chemical domains. Xiangzhe Kong and Wenbing Huang found that training a Generalist Equivariant Transformer (GET) on both protein-protein and protein-ligand interactions improved performance on both tasks. The model learns the underlying physical forces—van der Waals, electrostatic, and covalent—that are common to all matter, regardless of whether the system is biological or inorganic.

Accelerating the "Computational Microscope"

The ultimate goal of these foundation models is to accelerate molecular dynamics (MD). Traditional MD takes tiny steps (femtoseconds) to ensure stability. New frameworks like Timewarp and UniSim use these pretrained representations to take massive "leaps" in time, accelerating simulations by 25x to 1,000,000x.

By bypassing the Kohn-Sham equations and directly learning the mapping from atomic positions to forces, models like PFP enable the simulation of complex, real-world systems—like the growth of nanoparticles or the diffusion of ions in a solid-state battery—at the speed of classical physics but with the "ground truth" accuracy of quantum mechanics.

Papers in this version

Showing 60 of 63 papers.

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_universal_foundation_models_for_atomisti_2d26a454,
  title = {Universal Foundation Models for Atomistic Simulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/universal-foundation-models-for-atomistic-simulation/rdr_2d26a454ecd34426a164fd5f007e9b5f},
  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_universal_foundation_models_for_atomisti_2d26a454_v1,
  title = {Universal Foundation Models for Atomistic Simulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/universal-foundation-models-for-atomistic-simulation/rdr_2d26a454ecd34426a164fd5f007e9b5f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}