Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Extracting Atomic Environments for Machine Learning Interatomic Potentials

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A cubic cut followed by deleting colliding atoms reproduces DFT forces as well as or better than generative and annealing methods, making it the preferred way to prepare MLIP training data.

desk verdict Useful benchmark and a pleasantly simple winner, but the 'optimal' claim outruns the statistics. read the letter →

arxiv 2607.26018 v2 pith:QY7WVOXT submitted 2026-07-28 cond-mat.mtrl-sci

Jared C. Stimac , Fei Zhou , Kyle Bushick , Bo Lei , Sebastien Hamel , Amit Samanta , Vincenzo Lordi This is my paper
classification cond-mat.mtrl-sci
keywords atomicenvironmentextractionmachinelearninginteratomicpotentialsperiodicboundaryconditionsDFTtrainingdataactivediffusionmodeldeletionsmethodamorphousSiO2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tackles a practical bottleneck in building machine-learned interatomic potentials: carving a small, DFT-tractable cell out of a large simulation box without corrupting the forces of the atoms one cares about. The authors benchmark six extraction strategies across three materials — amorphous silica, dislocated tantalum, and molten carbon — and find that the most low-tech option wins. The 'deletions' method simply cuts a cubic cell around an atom of interest, applies periodic boundary conditions, and sequentially removes atoms whose MLIP-evaluated forces exceed a threshold, eliminating collisions across the new boundary. On force accuracy, deletions is best or tied for best in every material; its advantage over the runner-up cubic cut is not statistically significant, while its advantage over diffusion and annealing methods is. Because it needs no bespoke AI model and no long annealing runs, the paper argues deletions should be the default first choice for extraction, even though it leaves residual boundary defects and slightly inflated energies in tantalum.

What carries the argument

The deletions algorithm: (1) choose a fixed spherical core of radius r_core around a central atom; (2) cut out the entire cube of side L_cell from the source; (3) apply periodic boundary conditions; (4) using a pretrained IAP, compute per-atom forces, and repeatedly remove the outside-core atom with the largest |F| until max |F| < F_tol. The workhorse insight is that boundary collisions manifest as outlier forces, so force magnitude is a sufficient collision detector; no relaxation, no generation, and no stoichiometry fixing is needed.

What would settle it

Run the deletions protocol on a material with intrinsically large internal forces (e.g., a covalently bonded glass) using the paper's F_tol rule, and compare DFT forces of the central atom against the source; if the RMSE exceeds that of cubic extract by more than the paper's measured gaps, the claim of universality fails. A cheaper check: scan F_tol over a range for SiO2 and show whether the optimal threshold lands at the paper's chosen value or at a value that depends on the arbitrary source frame.

Watch

Extended reading notes

Core claim

The central claim is that, for bulk systems, the atomic forces of a preserved environment are best reproduced when the extraction method keeps the central region's local geometry exactly as it was in the source, even at the cost of structural inconsistency at the newly imposed periodic boundaries. The deletions method achieves this: it extracts all atoms in a cube around the central atom, applies PBCs, and iteratively deletes the atom with the largest force outside the fixed core, as judged by a cheap interatomic potential, until the maximum force falls below a per-material tolerance. Across SiO2, Ta, and C, deletions matched or beat every alternative — including a diffusion-model-based gene

Load-bearing premise

The whole recommendation rests on the assumption that a single per-material force threshold F_tol, set just above the maximum DFT force in the source configuration, reliably identifies the atoms whose removal fixes boundary collisions without disturbing the preserved core — an assumption the Ta results already strain, since residual collisions remain and the per-atom energy is still inflated.

Editorial extensions

If this is right

  • For bulk MLIP training-set construction, the deletions method gives DFT-force accuracy on par with or better than generative and annealing approaches, at a fraction of the setup cost.
  • Because deletions requires only an approximate force evaluator, it could be run with a simple distance cutoff instead of an MLIP; the paper explicitly suggests this.
  • The per-atom energy penalty of cubic extract (hundreds of eV in Ta) is removed by deletions, making the latter usable for energy-sensitive training.
  • Force accuracy degrades with distance from the central atom in all methods; deletions and cubic extract degrade most gradually, so a single DFT cell can provide trustworthy forces for more atoms.
  • The method is directly useful for active-learning and on-the-fly training workflows where large-scale MD configurations must be repeatedly reduced to DFT-sized cells.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Since the superiority of deletions over cubic extract is not statistically distinguishable on force error alone, the decisive advantage is energy cleanliness; the conclusion would flip if a way to neutralize cubic-extract collisions (e.g., short relaxation) produced identical forces.
  • The per-material F_tol tuning rule (just above max source force) is a heuristic; testing a fixed universal threshold across materials would clarify whether the method is effectively parameter-free in practice.
  • If force-magnitude thresholds are transferable, the same recipe could be applied to other embeddings (spherical clusters in vacuum, slab cells) without re-tuning, but that remains untested.
  • The finding suggests that 'physical plausibility' of the boundary region (as produced by diffusion or annealing) matters less than exact preservation of near-neighbor geometry; this invites a test of how much boundary disorder a trained MLIP actually tolerates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper benchmarks six methods for extracting atom-centered environments from large bulk configurations into smaller periodic cells suitable for DFT, with the goal of reproducing the DFT forces of the extracted atoms. The methods include the previously reported anneal (AME) approach and several new ones: spherical extract, cubic extract, deletions, deletions+relax, and a generative diffusion-model-based approach. Three materials are tested: amorphous SiO2, BCC Ta with screw dislocations, and molten C. The evaluation uses independent DFT reference forces on atoms in the source configurations and compares the forces in the extracted cells against those references. The authors report that the simple deletions method—which removes high-force atoms across periodic boundaries after a cubic extract—yields the lowest or tied-lowest force RMSE and maximum error for the central atoms among all methods, and they assert that deletions is the optimal extraction method because cubic extract, while competitive on forces, produces unphysically large per-atom energies.

Significance. If the central claim were fully established, the paper would provide a valuable practical recommendation: a trivial boundary-cleanup rule after cubic extraction can match or exceed more sophisticated generative or annealing-based methods for preparing MLIP training data, at a fraction of the computational cost. The study is carefully executed on three materially diverse systems, with explicit DFT convergence testing and paired t-tests on the force errors, which is a methodological improvement over prior anecdotal comparisons. The deletions method is also simple enough to be immediately adopted by practitioners. However, the force-accuracy advantage of deletions over cubic extract is not statistically significant for SiO2 and is only marginal for Ta relative to generative and deletions+relax, so the paper's headline claim of optimality rests heavily on a per-atom energy comparison that the authors themselves acknowledge is not rigorous and that also shows deletions leaves residual close contacts in Ta. The practical insight is still useful, but the strength of the conclusion exceeds what the evidence supports.

major comments (3)
  1. [§IV and §III.A] The central claim—that deletions is the optimal extraction method—is not supported by the paper's own primary metric. The paired t-tests in §III.A give p=0.80 for the SiO2 force-error difference between deletions and cubic extract, and p=0.08 and p=0.16 for Ta versus generative and deletions+relax, respectively, none of which reach the p≤0.05 significance criterion. For C, the RMSEs are reported as equivalent. Thus the force data show deletions is tied with cubic extract, not superior. The conclusion in §IV overstates the evidence; the claim should be softened to 'among the best and simplest' unless additional evidence is provided.
  2. [§III.D and Fig. 8] The energy-based tie-breaker used to dismiss cubic extract is not internally consistent. The authors state in §III.D that comparing DFT total energies of extracted cells to the source is 'not rigorous,' yet they use the per-atom energy excess of cubic extract (especially for Ta) as the decisive reason to prefer deletions. The same metric shows deletions itself leaves residual close contacts in Ta, with energies ~8 eV/atom above the source (Fig. 8b, inset). By the paper's own criterion, deletions is also energetically contaminated for Ta. If the energy metric is allowed to rule out cubic extract, it equally undermines deletions for Ta. The conclusion should either rely on force-only metrics (which do not distinguish deletions from cubic extract) or justify a rigorous energy-decomposition framework.
  3. [Table I and §II.A (deletions)] The deletions method's performance depends on the per-material force threshold F_tol, which is chosen as 'close to, but greater than, the maximum DFT force in the respective source configurations' (Table I). This means the method requires prior knowledge of the DFT force distribution of the large source configuration—exactly the quantity the extraction procedure is meant to avoid computing. No sensitivity analysis is provided to show that the results are robust to reasonable variations in F_tol, nor is an a priori selection rule given. Since F_tol is effectively tuned to each source, the claim that deletions is broadly superior may not transfer to new systems where the source DFT forces are unknown. A sensitivity study or a transferable prescription for setting F_tol is needed to support the general recommendation.
minor comments (5)
  1. [Appendix C] The paired t-test is described only in the main text; the appendix shows figures but not the numerical p-values for all pairs. Including a table of p-values for all method pairs and materials would make the statistical analysis more transparent and reproducible.
  2. [Table captions (A3–A6)] Several table captions in Appendix A are mislabeled: Tables A3 and A4 both describe Ta, but A4 is captioned 'k-points mesh size' while its column is h [Å], and A5/A6 similarly mix k-mesh and real-space mesh descriptions. Correct the captions to avoid confusion.
  3. [Abstract] The abstract says 'We demonstrated a notably simple procedure'—change to 'We demonstrate' for consistency with the rest of the paper's tense.
  4. [§II.A (anneal)] The anneal method description mentions a margin δ_margin=0.5 Å and removal of atoms to achieve stoichiometry for SiO2, but the subsequent density analysis in the SI shows the anneal configurations have densities 33% lower than the source. A brief comment on how the margin and stoichiometry adjustments affect density would help the reader interpret the structural results.
  5. [Fig. 5] The RDF plot for cubic extract is said to contain an extraneous peak near r=0, but this is difficult to see in the main figure. The zoomed-in panels are in the SI; consider adding a zoomed inset in the main figure or citing the SI more prominently at first mention.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the benchmark uses independent DFT force labels; minor parameter calibration and one non-load-bearing self-citation do not reduce the central claim to its inputs.

full rationale

The paper's central claim is an empirical ranking of extraction methods, not a derivation from first principles. The reference labels are DFT forces computed on the source configurations and on each extracted cell; the extraction procedures are not defined in terms of the force-error metric on which they are scored. The deletions threshold F_tol is 'chosen to be close to, but greater than, the maximum DFT force in the respective source configurations' (Table I), which is a heuristic calibration using the reference force scale, but it does not define the reported RMSE and the deletions configurations still require independent DFT evaluation. The energy-based tie-breaker used to prefer deletions over cubic extract is explicitly admitted to be 'not rigorous' (Sec. III.D), and the Ta generative model was trained on frames from the same compression trajectory as the source configuration; these are limitations in the evidence for the optimality claim, not definitional circularity. The diffusion-model approach cites Kwon et al. (Ref. 12), which includes two present authors, but that citation is not load-bearing for the main conclusion, since the generative method is only one of several comparators and the conclusion does not depend on it. No equation reduces to its own input, and no fitted parameter is renamed as a prediction. The paper's own p-values (e.g., p=0.80 for SiO2 against cubic extract) weaken the strength of the optimality claim, but that is an evidentiary issue, not circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the locality assumption, on careful DFT benchmarking, and on a set of hand-chosen extraction parameters (notably F_tol, r_core, L_cell). No new physical entities are introduced. The IAPs and diffusion model are borrowed from prior work and are treated as reliable tools.

free parameters (7)
  • F_tol (deletion force threshold) = SiO2: 7.5 eV/Å, Ta: 10 eV/Å, C: 25 eV/Å
    Chosen 'close to, but greater than, the maximum DFT force in the respective source configuration' (Sec. II A); tunes how aggressively atoms are deleted across periodic boundaries. A larger value would retain collisions; a smaller one would overly distort the extracted environment.
  • r_core (fixed core radius) = SiO2: 5.0 Å, Ta: 6.0 Å, C: 6.0 Å
    Defines the preserved environment around each central atom; affects how many atoms are kept and how much of the source environment is reproduced. Chosen per material, not varied.
  • L_cell (destination cell side) = SiO2: 14.4 Å, Ta/C: 14.0 Å
    Sets the new periodic cell size; implies vacuum gaps of 4.4, 2, and 2 Å for the spherical extract. Fixed per material to ensure fair comparison, but affects all results.
  • Repulsive wall parameters (epsilon, sigma) = epsilon=1 eV, sigma=0.5 Å
    Used in deletions+relax and anneal to keep outer atoms out of the core (SI Sec. I).
  • Relaxation stopping criterion = 0.15 eV/Å max force
    BFGS stopping criterion for deletions+relax; unusually loose, chosen by the authors.
  • t_anneal and quench rate = 50 ps (SiO2, Ta), 10 ps (C); quench 3.7e12 K/s
    Annealing schedule for the anneal (AME) method; affects pore formation and density.
  • Diffusion generation selection threshold = top 10 by RDF comparison
    Rule to pick generated structures; heuristic, not a systematic selection criterion.
assumptions (5)
  • domain assumption Nearsightedness / locality of electronic structure (Kohn-Prodan)
    Justifies that DFT forces in the extracted cell should match the source if the local environment is preserved; invoked in the Introduction (Refs 1–3). The central premise of the entire benchmark.
  • domain assumption DFT (PBE, SPARC) is an adequate ground truth for both source and extracted forces
    All comparisons use DFT as reference; convergence tests in Appendix A support the numerical settings, but exchange-correlation functional error is not discussed.
  • domain assumption The preselected IAPs (ACE for SiO2, EAM for Ta, ChIMES for C) are accurate enough to detect collisions and drive relaxations
    Deletions relies on IAP force magnitudes to decide what to remove; the anneal and relax methods rely on IAP dynamics. Fig. B1 shows DFT-vs-IAP RMSEs up to 1.17 eV/Å, so this proxy is imperfect but adequate.
  • standard math Paired Student t-test assumptions on squared residuals
    Used in Sec. III A and Appendix C to declare significance; requires independence and approximate normality, which is plausible but not verified.
  • domain assumption Diffusion model (Kwon et al., including LLNL co-authors) generates physically reasonable periodic structures
    The generative method's behavior is inherited from the score model trained on selected data; the paper's conclusions about generative underperformance may depend on training details. Ref. 12 has overlapping authors (Lordi, Zhou).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extracting Atomic Environments for Machine Learning Interatomic Potentials." pith.science (2026). https://pith.science/paper/QY7WVOXT

@misc{pith2026260726018,
  author       = {Pith},
  title        = {Pith review of: Extracting Atomic Environments for Machine Learning Interatomic Potentials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QY7WVOXT}},
  note         = {Machine review of arXiv:2607.26018}
}
abstract

In order to appropriately capture large-scale material features and emergent phenomena via atomistic simulations, such as Molecular Dynamics (MD), the system scale can range up to hundreds of millions of atoms. However, the force-field models that drive those simulations are generally trained with Density Functional Theory (DFT) reference data, limited to relatively small configurations on the order of 100s or 1000s of atoms. To compute DFT forces on atoms in regions of interest, for example for active-learning or on-the-fly training of interatomic potentials, one needs to extract a small set of atoms from the larger simulation box, and typically work with periodic boundary conditions for DFT. However, methods to select the shape and size of this extracted set of atoms, as well as to generate a potentially necessary passivating envelope, have not been systematically analyzed. In this work, we benchmark several techniques, including a generative diffusion-based artificial intelligence (AI) approach, for extracting atomic environments from large, bulk configurations and embedding them into smaller configurations suitable for DFT calculations with periodic boundary conditions. We test with a diverse set of material systems, which includes amorphous $\mathrm{SiO_2}$, Ta with screw dislocations, and molten C. We demonstrated a notably simple procedure, a method we refer to as deletions, yields superior performance over an array of alternative extraction methods.

Figures

Figures reproduced from arXiv: 2607.26018 by the authors.

Figure 1
Figure 1. FIG. 1. Flow diagram illustrating the six methods for atom-centered environment extractions [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 1
Figure 1. FIG. 1. Flow diagram illustrating the six methods for atom-centered environment extractions [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Source configurations for (a) SiO [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figures from the paper (8 more)
Figure 3
Figure 3. Figure 3: FIG. 3. Results for the atoms at the center of the extracted environments. (a) Maximum absolute [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]
Figure 3
Figure 3. Figure 3: FIG. 3. Results for the atoms at the center of the extracted environments. (a) Maximum absolute [PITH_FULL_IMAGE:figures/full_fig_p012_3.png]
Figure 4
Figure 4. Figure 4: FIG. 4. Accuracy of forces for non-central atoms within [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 4
Figure 4. Figure 4: FIG. 4. Accuracy of forces for non-central atoms within [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Radial distribution functions, [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Example Ta structure from the [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Stoichiometric ratio [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Difference between the average energy per atom of the extracted configurations and the [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 2 linked inside Pith

  1. [1]

    Journal of Physics: Condensed Matter , volume=

    Ask Hjorth Larsen and Jens Jørgen Mortensen and Jakob Blomqvist and Ivano E Castelli and Rune Christensen and Marcin Dułak and Jesper Friis and Michael N Groves and Bjørk Hammer and Cory Hargus and Eric D Hermes and Paul C Jennings and Peter Bjerre Jensen and James Kermode and John R Kitchin and Esben Leonhard Kolsbjerg and Joseph Kubal and Kristen Kaasbj...

    2017

  2. [2]

    Scientific Reports , volume=

    On machine learnability of local contributions to interatomic potentials from density functional theory calculations , author=. Scientific Reports , volume=. 2024 , publisher=

    2024

  3. [3]

    The new algorithm , author=

    The convergence of a class of double-rank minimization algorithms: 2. The new algorithm , author=. IMA journal of applied mathematics , volume=. 1970 , publisher=

    1970

  4. [4]

    Journal of Physics: Condensed Matter , volume=

    Recent progress in linearscaling ab initio electronic structure techniques , author=. Journal of Physics: Condensed Matter , volume=. 2002 , publisher=

    2002

  5. [5]

    Physical Review B , volume=

    Machine learning based interatomic potential for amorphous carbon , author=. Physical Review B , volume=. 2017 , publisher=

    2017

  6. [6]

    Physical Review B , volume=

    Atomic cluster expansion for accurate and transferable interatomic potentials , author=. Physical Review B , volume=. 2019 , publisher=

    2019

  7. [7]

    Nature Communications , volume=

    Modelling atomic and nanoscale structure in the silicon--oxygen system through active machine learning , author=. Nature Communications , volume=. 2024 , publisher=

    2024

  8. [8]

    Journal of Physics: Condensed Matter , volume=

    Generalized representative structures for atomistic systems , author=. Journal of Physics: Condensed Matter , volume=. 2024 , publisher=

    2024

Show all 27 references
  1. [9]

    A consistent and accurate ab initio parametrization of density functional dispersion correction (

    Grimme, Stefan and Antony, Jens and Ehrlich, Stephan and Krieg, Helge , journal=. A consistent and accurate ab initio parametrization of density functional dispersion correction (. 2010 , publisher=

    2010

  2. [10]

    Computer Physics Communications , volume=

    Atomsk: A tool for manipulating and converting atomic data files , author=. Computer Physics Communications , volume=. 2015 , publisher=

    2015

  3. [11]

    Machine Learning: Science and Technology , volume=

    In operando active learning of interatomic interaction during large-scale simulations , author=. Machine Learning: Science and Technology , volume=. 2020 , publisher=

    2020

  4. [12]

    Physical Review Letters , volume=

    Density functional and density matrix method scaling linearly with the number of atoms , author=. Physical Review Letters , volume=. 1996 , publisher=

    1996

  5. [13]

    Physical Review B , volume=

    Embedded-atom-method tantalum potential developed by the force-matching method , author=. Physical Review B , volume=. 2003 , publisher=

    2003

  6. [14]

    Journal of chemical theory and computation , volume=

    Chimes: A force matched potential with explicit three-body interactions for molten carbon , author=. Journal of chemical theory and computation , volume=. 2017 , publisher=

    2017

  7. [15]

    Physical review letters , volume=

    Restoring the density-gradient expansion for exchange in solids and surfaces , author=. Physical review letters , volume=. 2008 , publisher=

    2008

  8. [16]

    Proceedings of the National Academy of Sciences , volume=

    Nearsightedness of electronic matter , author=. Proceedings of the National Academy of Sciences , volume=. 2005 , publisher=

    2005

  9. [17]

    Physical Review B—Condensed Matter and Materials Physics , volume=

    Nearsightedness of electronic matter in one dimension , author=. Physical Review B—Condensed Matter and Materials Physics , volume=. 2006 , publisher=

    2006

  10. [18]

    The Journal of chemical physics , volume=

    Achieving plane wave accuracy in linear-scaling density functional theory applied to periodic systems: A case study on crystalline silicon , author=. The Journal of chemical physics , volume=. 2007 , publisher=

    2007

  11. [19]

    Chemical Physics Letters , volume=

    On nearsightedness in metallic systems for O (N) Density Functional Theory calculations: A case study on aluminum , author=. Chemical Physics Letters , volume=. 2017 , publisher=

    2017

  12. [20]

    arXiv preprint arXiv:2110.06197 , year=

    Crystal diffusion variational autoencoder for periodic material generation , author=. arXiv preprint arXiv:2110.06197 , year=

    arXiv

  13. [21]

    doi:10.1038/s42256-024-00837-3 , journal =

    Zheng, Shuxin and He, Jiyan and Liu, Chang and Shi, Yu and Lu, Ziheng and Feng, Weitao and Ju, Fusong and Wang, Jiaxi and Zhu, Jianwei and Min, Yaosen and Zhang, He and Tang, Shidi and Hao, Hongxia and Jin, Peiran and Chen, Chi and No. doi:10.1038/s42256-024-00837-3 , journal =

    doi

  14. [22]

    Machine Learning: Science and Technology , month =

    Kwon, Hyuna and Hsu, Tim and Sun, Wenyu and Jeong, Wonseok and Aydin, Fikret and Chapman, James and Chen, Xiao and Lordi, Vincenzo and Carbone, Matthew R and Lu, Deyu and Zhou, Fei and. Machine Learning: Science and Technology , month =. doi:10.1088/2632-2153/ad8c10 , eprint =

    doi

  15. [23]

    Advances in Neural Information Processing Systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in Neural Information Processing Systems , volume=

  16. [24]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    2011 arXiv

  17. [25]

    International Conference on Machine Learning , pages=

    Deep unsupervised learning using nonequilibrium thermodynamics , author=. International Conference on Machine Learning , pages=. 2015 , organization=

    2015

  18. [26]

    The Journal of Chemical Physics , volume=

    A nearsighted force-training approach to systematically generate training data for the machine learning of large atomic structures , author=. The Journal of Chemical Physics , volume=. 2022 , publisher=

    2022

  19. [27]

    0.0: Spin-orbit coupling, dispersion interactions, and advanced exchange--correlation functionals , author=

    SPARC v2. 0.0: Spin-orbit coupling, dispersion interactions, and advanced exchange--correlation functionals , author=. Software Impacts , volume=. 2024 , publisher=

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.