🔬 materials science

Extracting Atomic Environments for Machine Learning Interatomic Potentials

This paper benchmarks various techniques for extracting atomic environments from large-scale simulations for Density Functional Theory calculations, demonstrating that a simple "deletions" method outperforms alternative approaches across diverse material systems.

Original authors: Jared C. Stimac, Fei Zhou, Kyle Bushick, Bo Lei, Sebastien Hamel, Amit Samanta, Vincenzo Lordi

Published 2026-07-29

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Jared C. Stimac, Fei Zhou, Kyle Bushick, Bo Lei, Sebastien Hamel, Amit Samanta, Vincenzo Lordi

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

Imagine you are trying to understand how a massive, bustling city behaves during a traffic jam. To do this, you might build a giant, detailed model of the entire city. But here's the catch: the super-computers you use to calculate the physics of every single car, pedestrian, and streetlight are incredibly powerful, yet they can only handle a tiny neighborhood at a time. If you try to simulate the whole city, the computer crashes. So, scientists have a clever workaround: they zoom in on a small, interesting neighborhood (like a busy intersection) and try to study just that piece.

The problem is, a neighborhood doesn't exist in a vacuum. It's connected to the rest of the city. If you just cut out a square block and put it in a test tube, the edges look weird. The cars on the edge might crash into invisible walls, or the pedestrians might vanish into thin air. To fix this, scientists use a trick called "periodic boundary conditions," which is like wrapping the neighborhood in a magical bubble where if you walk out the front door, you instantly reappear at the back door. This makes the small piece feel like it's part of a huge, endless city. However, figuring out exactly how to cut that neighborhood out of the giant city, and how to wrap it in a bubble without breaking the laws of physics, is surprisingly tricky. If you get the cutting and wrapping wrong, your tiny model might look like a neighborhood, but the cars inside will drive like they're on Mars. This paper is about finding the best way to cut and wrap these tiny atomic neighborhoods so that scientists can study huge materials without needing a super-computer the size of a planet.

The Atomic Scissors: Cutting and Wrapping the Micro-World

In the world of materials science, researchers are obsessed with understanding how atoms behave to predict how materials like steel, glass, or even molten carbon will act in the real world. To do this, they use a high-tech microscope called "Density Functional Theory" (DFT). Think of DFT as a super-accurate calculator that can tell you exactly how every single atom in a material is pushing and pulling on its neighbors. The catch? This calculator is so hungry for power that it can only crunch numbers for a few hundred or a few thousand atoms at a time. But the real world is huge. Sometimes, to see a cool phenomenon like a crack forming in a bridge or a screw dislocation (a tiny twist in a crystal) moving through metal, you need to simulate millions of atoms.

So, scientists try to cheat. They take a giant simulation of millions of atoms, pick a small, interesting spot, and try to "extract" it into a tiny box to run their DFT calculator. But this is like trying to take a photo of a specific person in a crowded stadium and then pasting them onto a blank white background. If you just cut them out, the edges look fake. If you try to wrap them in a repeating pattern (so the person on the left edge is the same as the person on the right edge), you might accidentally paste their head into their own shoulder.

The authors of this paper, a team from Lawrence Livermore National Laboratory, decided to put six different "cutting and wrapping" techniques to the test. They wanted to see which method could take a tiny chunk of a giant material, put it in a small box, and make sure the atoms inside still felt exactly like they did in the giant original. They tested this on three very different materials: a glassy version of sand (amorphous SiO₂), a metal with a twisted defect (Tantalum with screw dislocations), and hot, melted carbon (molten C).

The Contenders: Six Ways to Slice the Cake

The team tried six different recipes for extracting these atomic neighborhoods:

Spherical Extract: They cut out a perfect ball of atoms and put it in a box with empty space (vacuum) around it. It's like putting a snowball in a box; the air around it is just empty.
Generative: They took the snowball and used a fancy AI (a "diffusion model") to invent new atoms to fill the empty space, trying to make the whole box look like a continuous city.
Cubic Extract: They cut out a perfect cube of atoms from the giant simulation and just slapped it into a box. They didn't worry if atoms on the edges crashed into each other when the box wrapped around.
Deletions: They started with the "Cubic Extract" cube, but then they played a game of "whack-a-mole." They checked which atoms were crashing into each other across the invisible boundaries and simply deleted the ones causing the most trouble until the crashes stopped.
Deletions + Relax: They did the "Deletions" game, but then they let the remaining atoms wiggle and settle down into a comfortable position, like letting a shaken-up soda can sit until it stops fizzing.
Anneal: They took a cube, left a small empty gap around the edges, heated it up to a scorching 4000 K (hotter than lava!), shook it around, and then slowly cooled it down to see if it settled into a nice shape.

The Results: The Simple Winner

The team ran their super-accurate DFT calculator on all these tiny boxes and compared the forces (the pushes and pulls) on the atoms to the forces in the original giant simulation. They were looking for the method that made the tiny box feel exactly like the big one.

The results were a bit surprising. The fancy methods, like the AI "Generative" approach or the high-heat "Anneal" method, didn't win. In fact, the "Anneal" method created tiny holes (pores) in the metal because the cooling process was a bit too aggressive, and the AI sometimes added atoms that didn't quite fit the neighborhood's style.

The winner was the Deletions method. It was the simplest approach: just cut a cube, wrap it, and delete the atoms that were crashing into each other. It turned out that this "rough and ready" method actually preserved the local environment of the atoms better than the complex, high-tech alternatives. The forces on the atoms in the "Deletions" boxes matched the original giant simulation almost perfectly, with an average error of just 0.08 eV/Å. The "Anneal" method, by comparison, had a much higher error of 0.23 eV/Å.

Even the "Cubic Extract" method (which didn't delete anything and let atoms crash into each other) did a decent job with the forces, but it had a major flaw: the crashing atoms created huge, unrealistic energy spikes. It's like having a car crash in your model; the physics of the crash is real, but it ruins the simulation of the rest of the traffic. The "Deletions" method avoided these crashes without needing to run expensive simulations or train complex AI models.

Why This Matters

The authors suggest that this simple "Deletions" technique is a game-changer for building training data for Machine Learning Interatomic Potentials (MLIPs). These are the "smart" models that scientists train to predict material behavior without running the super-expensive DFT calculations every time. To train these smart models, you need lots of examples of how atoms behave. By using the "Deletions" method, scientists can grab tiny, accurate snapshots from massive simulations and feed them to their AI, making the AI smarter and more reliable.

While the paper doesn't claim this solves every problem in materials science, it strongly suggests that sometimes the simplest solution—cutting a cube and deleting the messy bits—is better than trying to be too clever with AI or high-temperature cooking. It's a reminder that in the chaotic world of atoms, keeping things local and avoiding collisions might be the key to unlocking the secrets of huge, complex materials.

Technical Summary: Extracting Atomic Environments for Machine Learning Interatomic Potentials

Problem Statement
Developing Machine-Learned Interatomic Potentials (MLIPs) requires training datasets that span the configuration space of large-scale atomistic simulations (e.g., Molecular Dynamics with millions of atoms). However, generating the ground-truth labels (energies and forces) for these structures via Density Functional Theory (DFT) is computationally limited to systems of roughly $10^2$ to $10^3$ atoms. A critical bottleneck arises when critical atomic structures emerge only in large-scale simulations that exceed DFT capabilities. To address this, researchers must extract small clusters of atoms from large "source" configurations and reconstruct them into smaller "destination" cells suitable for DFT calculations with Periodic Boundary Conditions (PBC). While the principle of locality (nearsightedness) suggests this is feasible, there is no systematic analysis or established benchmark for the optimal methods to select the shape, size, and passivation of these extracted sets, particularly regarding how to handle PBCs without introducing nonphysical artifacts.

Methodology
The authors benchmarked six distinct atom-centered extraction techniques using three diverse material systems: amorphous SiO $_2$ , body-centered cubic (BCC) Ta containing screw dislocations, and molten Carbon (C). In all cases, a "fixed core" (a central atom and its immediate neighbors within a predefined radius) was extracted from a large source configuration and embedded into a smaller cubic destination cell. The methods tested were:

Spherical Extract: The fixed core is placed in a vacuum-filled destination cell.
Generative: A spherical extract is augmented by a generative diffusion AI model that adds atoms to the vacuum region to form a coherent periodic structure.
Cubic Extract: The fixed core and all atoms within the cubic volume of the destination cell are extracted intact from the source, imposing new arbitrary PBCs without modification.
Deletions: Starting from the cubic extract, atoms colliding across PBCs are iteratively removed based on a low-cost Interatomic Potential (IAP) until forces on non-core atoms fall below a tolerance threshold ( $F_{tol}$ ).
Deletions + Relax: The "Deletions" configuration is further refined via energy minimization (BFGS) using an IAP, with a repulsive wall preventing atoms from entering the fixed core.
Anneal (AME): Following the Amorphous Matrix Embedding (AME) protocol, a margin is enforced, and atoms outside the core are annealed at high temperature (4000 K) and quenched to remove spurious boundary interactions.

The performance of these methods was evaluated by comparing DFT-calculated forces in the extracted destination cells against the forces in the original source configurations. The primary metric was the ability to reproduce the atomic forces of the extracted atoms, specifically focusing on the central atom and the surrounding fixed core. Structural fidelity was assessed via Radial Distribution Functions (RDFs) and stoichiometry, while energetic consistency was evaluated via per-atom average energies.

Key Results

Force Accuracy: The Deletions method yielded the lowest average Root-Mean-Squared Error (RMSE) for atomic forces (0.08 eV/Å) across all materials, outperforming the Anneal method (0.23 eV/Å). The Cubic Extract method performed nearly as well as Deletions in terms of force error but was statistically indistinguishable from it for SiO $_2$ and C.
Statistical Significance: Paired Student t-tests confirmed that the performance difference between Deletions and all other methods (except Cubic Extract and, in specific cases, Generative/Deletions+Relax for Ta) was statistically significant ( $p \le 0.05$ ).
Spatial Dependence: For atoms within the fixed core, Deletions and Cubic Extract maintained lower force errors at greater distances from the central atom compared to other methods. Errors for other methods increased sharply as atoms approached the new periodic boundaries.
Structural and Energetic Trade-offs:
- Cubic Extract: While accurate for forces, this method allowed atomic collisions across PBCs, resulting in unphysically high per-atom energies (hundreds of eV/atom for Ta), making it unsuitable for energy-based training.
- Anneal: This method produced structurally coherent configurations but introduced significant artifacts, including broadened RDF peaks and the formation of atomic-scale pores due to the enforced margin and quenching process.
- Generative: The diffusion model produced reasonable structures for Ta and C but showed broadened RDF peaks for SiO $_2$ , attributed to the model being trained on a broader dataset than the specific source configuration.
- Deletions: This method successfully removed most atomic collisions while preserving the local environment. Although it did not eliminate all collisions (resulting in moderately elevated energies for Ta), it avoided the severe structural distortions seen in the Anneal method.

Significance and Claims
The paper claims that the Deletions method is the optimal extraction strategy among those tested for the specific bulk materials examined (covalent SiO $_2$ , metallic Ta, and molten C). The authors emphasize that this method is "serendipitously simple," requiring no generative AI models or time-consuming MD simulations, yet it outperforms more sophisticated approaches in reproducing the target DFT forces.

The primary conclusion is that preserving the exact local environment of the central atoms (even if it results in structural inconsistencies at the periodic boundaries) is more critical for force accuracy than attempting to create a perfectly coherent atomic arrangement across the boundaries. The authors assert that while structural metrics (RDFs, stoichiometry) are important, they do not guarantee the reproduction of atomic forces, which is the primary goal for active learning and on-the-fly training of MLIPs. The work suggests that simple, deterministic extraction and pruning strategies may be superior to complex generative or relaxation-based embedding techniques for curating high-fidelity MLIP training data.

More like this