🔬 materials science

Atoms that Matter: Largest Materials Datasets Are 99% Redundant, and Local Chemistry Can Guide the Required 1%

This study demonstrates that training machine-learning force fields on less than 1% of the OMat24 dataset by selecting structurally and chemically distinct local environments yields accuracy comparable to models trained on the full 100-million-structure set, proving that chemical diversity rather than sheer data volume is the critical factor for general-purpose materials modeling.

Original authors: Evgeny Moerman, Igor Poltavsky, Anton Charkin-Gorbulin, Alexandre Tkatchenko

Published 2026-10-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Evgeny Moerman, Igor Poltavsky, Anton Charkin-Gorbulin, Alexandre Tkatchenko

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to understand how the physical world holds together, scientists have long relied on a powerful but expensive tool: quantum mechanics. This branch of physics allows researchers to calculate how atoms interact with one another, predicting everything from the strength of a new metal to the way a battery stores energy. However, these calculations are so computationally heavy that simulating even a small piece of material for a fraction of a second can take days or weeks on the world's most powerful supercomputers. To overcome this bottleneck, a new generation of artificial intelligence has emerged. These machine-learning models act as a shortcut; they are trained on vast libraries of quantum calculations and then learn to predict atomic behavior almost instantly, offering the accuracy of the heavy calculations with the speed of a simple guess. The prevailing belief in the field has been that to make these AI models truly reliable and able to handle any material, they must be fed an enormous amount of data, often numbering in the hundreds of millions of atomic configurations.

A team of researchers at the University of Luxembourg has challenged this assumption, suggesting that the sheer size of these data libraries may be an illusion of necessity. They examined the Open Materials 2024 dataset, a massive collection containing over 110 million structures, which has served as the foundation for the most advanced AI force fields to date. The researchers asked a simple question: does the model actually need all 110 million examples to learn, or is most of that information just a repetition of the same chemical lessons? To find out, they trained new AI models on tiny slices of this massive dataset—specifically, subsets containing less than one percent of the total atoms, or roughly 6.5 million atoms. They tested three different ways to choose these slices: picking atoms completely at random, using a sophisticated clustering method to find distinct groups, and employing a new technique they developed that counts the unique local neighborhoods around each atom.

The results were striking. The models trained on these tiny, carefully selected subsets performed just as well as the model trained on the entire 110-million-structure dataset. In tests measuring how accurately the AI could predict the forces holding atoms together, the small-subset models were nearly identical to the giant one. They also matched the full model in predicting key physical properties like how much a material expands when heated, how stiff it is, and how easily lithium ions move through it. Perhaps most surprisingly, when the researchers pushed these models to their limits by simulating extreme conditions—ramping up temperatures to 3000 Kelvin and pressures to 500 gigapascals, conditions found deep inside planets—the models trained on the small subsets actually proved more stable and reliable than the one trained on the full dataset. The full dataset model frequently crashed under these extreme stresses, while the streamlined versions kept running, suggesting that the extra 99 percent of the data did not add robustness but rather introduced redundancy.

The researchers discovered that the key to building a better model was not volume, but variety. They found that the massive dataset contained many structures that shared the exact same local chemical environments, essentially repeating the same information over and over. To solve this, they introduced a method called EnvTok, which acts like a cataloger for atomic neighborhoods. Instead of looking at the whole structure, it breaks the material down into the immediate surroundings of each atom, noting what elements are present, how they are arranged, and how much they are squeezed or stretched. By using this catalog to select only the structures that offer a new, unique neighborhood, they could build a training set that covered the entire chemical landscape without the repetition. This approach allowed them to identify the essential 1 percent of data that carries the real information, leaving the rest behind.

This finding suggests a major shift in how scientists should approach training artificial intelligence for materials science. Rather than spending immense computational resources to label and process billions of redundant structures, researchers can now focus on identifying and labeling only the distinct, chemically unique configurations. This efficiency opens the door to using even more accurate, higher-level physics calculations to train these models, a step that was previously too expensive to consider for such large datasets. By freeing up the computational budget, scientists can now explore a wider range of chemical environments and push the boundaries of what these models can predict, from designing new super-hard materials to understanding the interiors of distant planets. The study confirms that in the world of atomic simulation, quality of information matters far more than the quantity of data.

Technical Summary: Atoms that Matter: Largest Materials Datasets Are 99% Redundant, and Local Chemistry Can Guide the Required 1%

Problem Statement
General-purpose machine-learning force fields (MLFFs) for materials have recently been trained on datasets containing on the order of $10^8$ structures, such as the Open Materials 2024 (OMat24) dataset, under the assumption that transferability requires massive data volume. However, training on such scales incurs prohibitive computational costs for both reference density-functional theory (DFT) calculations and model training. Furthermore, the generation of these datasets—often involving non-equilibrium configurations derived from a smaller set of parent structures via rattling and molecular dynamics—suggests substantial redundancy in local atomic environments. The fraction of data providing distinct, learnable information remains unknown, and it is unclear whether models trained on a small fraction of these datasets can match the performance of full-data models.

Methodology
The authors investigated the information content of OMat24 by training MACE (Message Passing Atomic Cluster Expansion) models on subsets containing approximately 6.5 million atoms (less than 1% of the f-block-free OMat24 dataset, which comprises ~1.2 billion atoms). Three distinct selection strategies were employed to construct these subsets:

Random Sampling: Served as a baseline to test if common environments are captured by chance.
DIRECT: A model-guided approach that performs principal component analysis (PCA) on the latent space of a pre-trained M3GNet model, clusters the structures, and samples from each cluster to ensure coverage of configuration space extrema.
EnvTok (Environment Token): A new, model-free method introduced in this work. EnvTok assigns each atom a discrete, interpretable 7-field token based on:
- Central element and magnetic moment.
- Coordinating elements (local composition).
- Coordination polyhedron type.
- Distortion level (symmetry measure).
- Bond compression (relative to covalent radii).
- Compression asymmetry.
  The subset selection is formulated as a weighted set-cover problem, aiming to find the smallest set of structures that collectively covers all sufficiently frequent environment classes defined by these tokens.

The resulting models were benchmarked against the full-data reference model (MACE-OMAT-0) across four categories: static atomic forces (MatPES), near-equilibrium thermodynamic and mechanical properties (MatCalc), molecular dynamics (MD) stability under extreme conditions (MLIP Arena), and Li $^+$ migration barriers (nebDFT2k).

Key Results

Redundancy: All three subset-trained models (random, DIRECT, and EnvTok) achieved performance comparable to the full-data MACE-OMAT-0 model. For static forces, they approached the reference within a few tens of meV/Å. For formation energies, elastic moduli, and heat capacities, they remained close to or exceeded the full-data model's accuracy.
Selection Strategy Impact:
- Near-Equilibrium Properties: Performance was largely independent of the selection strategy. Random sampling performed nearly as well as targeted methods for formation energies and bulk properties, suggesting that common environments are ubiquitous even in small random subsets.
- Extreme Conditions (MD Stability): The full-data model was the least stable, with only 75% of MD runs completing without physical divergence under ramped temperature (300–3000 K) and pressure (0–500 GPa). In contrast, the EnvTok and DIRECT models achieved 92–93% completion rates. This indicates that balanced coverage of specific local environments (particularly compressed and asymmetric ones) is more critical for stability than total dataset size.
- Extrapolation (Li $^+$ Migration): For Li $^+$ migration barriers (a defect structure absent from OMat24), targeted selection (EnvTok and DIRECT) significantly outperformed random sampling, reducing Mean Absolute Errors (MAE) by 15–21% and Root Mean Square Errors (RMSE) by 49–55% relative to the full-data model.
EnvTok Validation: The authors mapped EnvTok tokens to the latent space of the trained MACE-OMAT-0 model. They found that 98.4% of descriptors formed a single connected component in the latent graph, and 96.9% formed tighter clusters than random samples. This confirms that the model-free EnvTok descriptor captures chemically meaningful distinctions that align with the representations learned by the MLFF.

Significance and Claims
The paper claims that the information content of OMat24 is defined by chemical environment diversity rather than sheer data volume, with approximately 99% of the structures being largely redundant. The primary contribution is demonstrating that reference labels are only strictly necessary for structurally and chemically distinct configurations.

By using a model-free, interpretable descriptor like EnvTok, researchers can identify these distinct environments before any labeling or training occurs. This enables a "hundredfold reduction" in dataset size, making it computationally feasible to:

Relabel general-purpose training sets with higher-level theories (e.g., hybrid-DFT) rather than being restricted to semilocal functionals.
Free up computational budgets to extend environmental coverage (e.g., including low-dimensional structures) and pursue multi-fidelity learning.

The work argues that the route to robust, transferable MLFFs lies in broad sampling of chemical space to capture distinct environments, rather than simply accumulating massive volumes of redundant data.

More like this