Scaling Molecular Data for Universal Atomic Models

This direction is split into other directions: chemical space coverage. Its last version stays readable and citable here.

Version 1 of 1147 papers158 conceptsfrozen 2026-10-04

For decades, computational chemistry was a trade-off between speed and accuracy. You could have the "gold standard" precision of Coupled-Cluster theory, but only for a handful of atoms. Or you could use faster methods like Density Functional Theory (DFT), which still struggled to scale to the millions of molecules required for modern drug discovery or materials design.

The recent shift toward machine learning potentials promised a third way: the speed of classical force fields with the accuracy of quantum mechanics. However, researchers quickly hit a "data drought." Models trained on small, equilibrium datasets of simple organic molecules failed when they encountered the messy reality of chemical reactions, transition states, or the broader periodic table.

The current research direction focuses on bridging this "coverage gap" by scaling molecular data across three dimensions: chemical diversity (more elements), configurational diversity (off-equilibrium states), and structural complexity (larger, more varied frameworks).

The Chemical Coverage Gap

Most early molecular datasets focused on the "basic" elements of organic chemistry: carbon, hydrogen, nitrogen, and oxygen. While useful, this narrow focus created models that were blind to the rest of the periodic table.

As Guangyong Chen and colleagues argued in the Alchemy project, standard benchmarks lacked the structural complexity found in real-world medicinal chemistry, particularly elements like sulfur and chlorine. Without representative data for these atoms, models like ALPHA-ML showed diminished performance when predicting properties like molecular polarizability.

To solve this, researchers began generating massive, high-fidelity datasets that span the periodic table. Toshiyuki Maeda released the PubChemQC dataset, providing electronic properties for 86 million molecules. Similarly, the MSR-ACC/TAE25 collection from Microsoft Research provides 73,040 high-precision atomization energies, specifically including inorganic molecules to ensure models can navigate a broader chemical space.

Moving Beyond Equilibrium

A molecule is not a static object; it is a vibrating, reacting system. If a model only learns from "equilibrium" geometries—the lowest-energy, most stable shapes—it becomes useless for simulating molecular dynamics or chemical reactions.

Justin S. Smith and Adrián E. Roitberg addressed this with ANI-1, a dataset of 20 million "off-equilibrium" calculations. By slightly distorting molecular structures, they taught models the "forces" that push atoms back toward stability.

This concept was pushed further by Arghya Bhowmik and the team behind Transition1x. Instead of just looking at stable states, they focused on the "mountain passes" of chemical reactions—the transition states. By training on 9.6 million calculations of these high-energy configurations, they enabled machine learning potentials to accurately predict reaction barriers, a prerequisite for autonomous catalyst design.

The Rise of Large Atomic Models

The culmination of this scaling effort is the "Large Atomic Model" (LAM). Much like Large Language Models (LLMs) learn the structure of human language by reading the entire internet, LAMs learn the "language of atoms" by training on diverse datasets simultaneously.

Duo Zhang and collaborators developed DPA-1 and its successor DPA-2. These models use multi-task pre-training across dozens of elements. A fascinating emergent property of this approach is that the model's latent representations of atoms naturally organize into a structure that mirrors the periodic table. This universal representation allows the model to generalize to new chemical environments with significantly less data than traditional methods.

Mining the Literature and Synthetic Data

When high-throughput simulations are too expensive, researchers are turning to the vast archives of existing chemical knowledge. Aditya Nandy and Heather J. Kulik developed MOFSimplify, using natural language processing to mine decades of literature for experimental stability data on metal–organic frameworks.

In cases where even the literature is sparse, researchers like Yuhan Chen are using LLMs to generate "artificially real" molecule-description pairs. By pre-training on this pseudo-data, models can learn the underlying relationships between structure and function before ever seeing a real experimental data point. This "data-centric" approach, exemplified by models like ChemLLM, suggests that the path to universal chemical intelligence lies not just in better algorithms, but in the massive, diverse aggregation of everything we know about the molecular world.

Papers in this version

Showing 60 of 147 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_scaling_molecular_data_for_universal_ato_f0fd3cd8,
  title = {Scaling Molecular Data for Universal Atomic Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-molecular-data-for-universal-atomic-models/rdr_f0fd3cd8e91040db8d9f72e6ef515027},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_scaling_molecular_data_for_universal_ato_f0fd3cd8_v1,
  title = {Scaling Molecular Data for Universal Atomic Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-molecular-data-for-universal-atomic-models/rdr_f0fd3cd8e91040db8d9f72e6ef515027/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}