First-hand information for everyone
This paper studies how to cut small, DFT-ready chunks from very large atomistic simulations so that accurate quantum calculations can be done on regions of interest. The authors benchmark six practical ways to extract an atom-centered environment from a bulk structure with periodic boundaries and embed it into a smaller cell suitable for Density Functional Theory (DFT). Their main finding is that a straightforward procedure they call “deletions” performs better than a variety of alternative methods across the systems they tested.
The team applied all six extraction methods to three different material cases: amorphous silicon dioxide (SiO2), body-centered cubic tantalum (Ta) containing screw dislocations, and molten carbon (C). For each case they started from large source configurations (more than a thousand atoms), extracted many fixed-radius atom-centered cores, and embedded those cores in smaller periodic cells. They then ran DFT to compute atomic forces in the small cells and compared those forces to the original DFT forces from the large configuration. The authors use the match of atomic forces as the main objective metric because forces reflect the local interactions they aim to preserve.
The six approaches are: spherical extract, which places the core in vacuum inside a small cell; generative, which first uses the spherical extract then fills the empty region with atoms produced by a diffusion-based generative model; cubic extract, which copies the cubic region around the atom and reimposes periodic boundaries; deletions, which starts from the cubic extract then iteratively removes atoms that create large spurious forces across the new boundaries until a force tolerance is met; deletions+relax, which follows deletions with a cheap energy minimization while preventing atoms from moving into the core; and anneal, a variant of the reported Amorphous Matrix Embedding (AME) method that imposes new periodic boundaries then anneals and quenches atoms outside the core to remove artificial boundary effects. The deletions procedure uses a low-cost interatomic potential to detect and remove problematic atoms and stops when remaining forces are below a chosen tolerance.
This work matters because modern machine-learned interatomic potentials (MLIPs) and active-learning workflows often need DFT labels for atoms that appear only inside very large simulations. DFT calculations are expensive and limited to a few hundred or thousand atoms, so extracting representative small environments is a practical necessity. By comparing methods on realistic and diverse materials and by using DFT forces as the yardstick, the paper gives guidance on which extraction strategies are most likely to preserve the local physics needed for training and validation.
Important caveats are noted. Successful extraction depends on the locality of interactions—how far away atoms influence the forces—and this assumption is weaker for systems with strong long-range electrostatics. The benchmark covers three representative systems but is not an exhaustive survey of all materials or defect types, so users should validate methods for their own problems. Some approaches, like annealing, can be computationally costly because they require molecular dynamics steps, and generative models sometimes placed atoms inside the fixed core and needed correction. Finally, the deletions method requires choosing a force tolerance and a cheap interatomic potential to screen atoms; those choices affect results and must be tuned. The authors report that deletions gave superior performance in their tests, but they recommend careful testing before applying the recipe to other materials or conditions.