Pith. sign in

REVIEW 2 major objections 4 minor 39 references

Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection of ASSYST Structure Pools

T0 review · 2 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Label-free leverage sampling cuts ACE training cost two- to threefold while preserving defect accuracy.

desk verdict A careful empirical study of leverage-based subset selection for linear ACE, but the cell-size confound in Eq. (21) and missing code/data keep it from being fully convincing. read the letter →

arxiv 2607.18524 v1 pith:EEDQ5BR6 submitted 2026-07-20 cond-mat.mtrl-sci

Aynour Khosravi , Marvin Poul , Jörg Neugebauer , Chad Sinclair This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords linearACEpotentialsleveragescoresCURmatrixsamplingASSYSTstructurepoolsDFTdatasetefficiencyinteratomicpotentialtransferabilitysubsetselectionlabeledfraction
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 argues that for linear ACE interatomic potentials trained on systematically generated ASSYST structure pools, the bottleneck of DFT labeling can be cut by choosing training configurations with statistical leverage scores that depend only on atomic geometry and the ACE basis, not on any energy or force labels. In elemental Al, leverage-guided subsets reach the same plateau accuracy as random sampling at about 30–40% labeled fraction instead of 75–80%, an effective two- to threefold reduction in DFT cost, with similar behavior for Cu and Al–Cu. The paper further shows that leverage-selected subsets preserve or improve defect-level accuracy—vacancy formation, grain-boundary, and dilute-alloy energetics—so data efficiency and transferability are not in conflict. A sympathetic reader cares because DFT labeling is the dominant expense in building machine-learned potentials, and this offers a practical, offline way to reduce it before any quantum-mechanical calculation.

What carries the argument

Regularized leverage scores are the diagonal entries of the hat matrix H_λ = Ã(βÃᵀà + Σ₀⁻¹)⁻¹Ãᵀ, where à is the weighted ACE design matrix (one row per energy, force component, or stress observable) and Σ₀⁻¹ is the smoothness-prior precision from Bayesian linear regression. These scores are label-free: they are computed from the geometry and the chosen basis before any DFT labels exist. Configuration-level leverage L_R = Σ_{i∈I_R} h_i aggregates the row scores over all observables of a configuration, and CUR/block-CUR sample configurations with probability proportional to L_R (or its energy-only analog), realizing a geometry-driven approximation to optimal experimental design without explici

What would settle it

Compute configuration-level leverage scores for the ASSYST Al pool and check whether the highest-ranked configurations are disproportionately the largest cells (8–10 atoms). If so, repeat the training comparison with a cell-size-stratified pool that contains equal numbers of each cell size; if the leverage advantage over random sampling disappears or shrinks, the reported gains are a cell-size artifact rather than a descriptor-coverage effect.

Watch

Extended reading notes

Core claim

The central claim is that configuration-level regularized leverage scores, defined as the sum over all energy and force rows of the diagonal of the Tikhonov-regularized hat matrix of the weighted ACE design matrix, provide a label-free ranking of ASSYST configurations by information content. Sampling without replacement proportional to these scores (CUR for energy-only, block-CUR for energy+force) yields training subsets that reach the test-error plateau at f_lab ≈ 0.30–0.40 for Al, whereas random sampling requires f_lab ≈ 0.75–0.80 — a two- to threefold reduction in DFT-labeled configurations. In Cu and Al–Cu, the same selection converges faster and, notably, produces lower vacancy-formatio

Load-bearing premise

The selection ranking assumes that summing per-observable leverage scores over a configuration is unbiased because ASSYST cells have comparable atom counts, but the pools actually span 1 to 10 atoms per element, so configurations with more atoms contribute more rows and may dominate the ranking purely by size.

Editorial extensions

If this is right

  • For elemental Al and Cu, training on leverage-selected subsets can reach the accuracy plateau with roughly 30–50% of the labeled configurations that random sampling would require, directly reducing DFT workload by a factor of two to three.
  • Because leverage scores require no labels, workflows can be reordered to filter-then-label: generate a large ASSYST pool, rank by leverage, and run DFT only on the selected subset.
  • Defect-level transferability is preserved at reduced training size: leverage-selected Cu models cut vacancy-formation error from about 0.25 eV to 0.04 eV and grain-boundary error by about a third compared with baselines.
  • For ordered Al–Cu intermetallics, leverage subsets at 25% labeled data match 50% baselines on antisite and vacancy errors; in dilute Cu-in-Al, CUR at 25% achieves vacancy-formation RMSE below the 50% baselines.
  • Leverage-based sampling offers an offline, non-iterative alternative to online active learning, avoiding repeated retraining and DFT-in-the-loop while still concentrating labels in informative regions of descriptor space.

Reading between the lines

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

  • The configuration-level sum of row leverages may be biased by cell size: ASSYST pools contain cells from 1 to 10 atoms, so a 10-atom cell contributes 31 rows while a 1-atom cell contributes 4 rows; normalizing by row count or atoms would test whether the ranking advantage reflects descriptor coverage or simply larger cells.
  • The same leverage machinery extends naturally to any linear or linearized MLIP with a design matrix and a regularized hat matrix (e.g., linear GAP or MTP), and to the last-layer features of neural network potentials, though these extensions are untested.
  • A hybrid scheme that uses leverage scores to build the initial diverse subset and then switches to active-learning uncertainty sampling might yield further reductions, since leverage covers the pool while active learning targets the current model's weak spots.
  • The reported 2–3x compression is measured at the accuracy plateau within one ACE architecture (ν=3, td=17, r_cut=6.5 Å); sweeping body order and basis size could reveal whether the compression factor grows or shrinks with model complexity.
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

2 major / 4 minor

Summary. The paper proposes using regularized leverage scores computed from the ACE design matrix to select subsets of ASSYST-generated configurations for DFT labeling, without access to energy or force labels. It compares two leverage-guided strategies, CUR and block-CUR, against random, energy-based, and force-based selection under an iterative labeling protocol for elemental Al, Cu, and Al-Cu alloys. The central claim is that leverage-guided selection reaches plateau-level accuracy at labeled fractions of roughly 30-40% for Al, versus 75-80% for random sampling, corresponding to a 2-3x reduction in DFT-labeled configurations, while preserving defect-level transferability.

Significance. If the central claim holds, this is a practically valuable result: it offers a label-free, geometry-only criterion for reducing DFT labeling cost in linear ACE model training, which is directly relevant to MLIP workflow design. The study has several strengths: a fixed ACE architecture across all methods, five-seed repeated runs, hold-out test sets, an independent 360-config DFT validation set for Al, and label-driven baselines that have strictly more information than the label-free methods. The Al learning-curve numbers (30-40% vs 75-80%) are striking. However, the mechanistic interpretation and the workload-reduction claim depend on the unnormalized configuration-level leverage definition in Eq. (21), which is not adequately justified and requires an additional control experiment before the central claim is secure.

major comments (2)
  1. [Sec. II.D.1, Eq. (21), Table I] The definition L_R = Σ_{i∈I_R} h_i is justified by the statement that 'variations in |I_R| are limited'. Table I contradicts this: ASSYST cells contain 1–10 atoms per element, so |I_R| = 1 + 3N_R spans 4 to 31, an eight-fold variation. Since block-CUR samples with probability proportional to L_R (Eq. 23), it will systematically favor configurations with more force rows unless per-row leverages are strongly anti-correlated with N_R. No such anti-correlation is shown. Consequently, the block-CUR plateau advantage in Fig. 2 and the defect-energy gains in Table II could be, at least in part, a cell-size selection effect rather than evidence of descriptor-space coverage. A size-matched random baseline (random sampling conditioned on the same cell-size distribution as the leverage-selected set) or normalization of L_R by |I_R|, together with a report of selected cell-size distributions, is nee
  2. [Sec. III.E, Abstract] The reported 2–3x reduction is stated in terms of the number of DFT-labeled configurations (f_lab). DFT cost, however, scales with the number of atoms in the cell, not simply with configuration count. If leverage selection enriches for larger cells (see major comment above), the reduction in total DFT workload may be smaller than claimed. The paper should report, at the plateau for each method, the total number of atoms labeled or an estimated DFT cost, not only the configuration fraction. This is load-bearing for the practical claim that leverage-guided selection reduces DFT labeling cost by 2–3x.
minor comments (4)
  1. [Sec. III.A/B] The hyperparameter scan uses a fixed 20% validation split to select the ACE architecture; the learning-curve protocol then reserves 20% of the data as an 'independent' test set. Please clarify whether these are the same 20% or disjoint splits. If the same split is reused for hyperparameter selection and testing, the absolute error levels may be optimistic, although the relative comparison between methods would remain valid.
  2. [Sec. III.B, Eq. (25)] The text says force-based selection 'selecting configurations solely by large force magnitudes does not efficiently span descriptor space', but Eq. (25) ranks by force residual, not by force magnitude. Please align the wording with the actual selection criterion.
  3. [Sec. V] Code availability states that a link 'will be added upon publication'. For reproducibility of the specific leverage computations and the iterative protocol, it would be helpful to make the code available with the submitted version or to include sufficient pseudocode in an appendix.
  4. [Table II] Minor typographical/formatting issues: 'm eV/atom' and 'm eV/Å' should be 'meV/atom' and 'meV/Å'. Also, 'Global E' and 'Global F' columns are not explicitly defined in the table caption; please clarify the metric.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: leverage scores are label-free and the central comparisons are evaluated on held-out and independent DFT data.

full rationale

The derivation chain is not circular. Regularized leverage scores (Eq. 19) depend only on the weighted ACE design matrix A~ = W A and the prior, not on the DFT targets y; CUR/block-CUR selection probabilities (Eqs. 22-23) are built from these label-free scores. The claimed efficiency gain—plateau accuracy at f_lab ≈ 0.30–0.40 versus f_lab ≈ 0.75–0.80 for random sampling—is measured on a held-out 20% split of the ASSYST pool and on an independent 360-configuration DFT validation set (Sec. III.D), so the 'prediction' is not equivalent by construction to a fitted input. The ASSYST pools do come from prior work by two of the authors (Refs. 35,36), but this is input data rather than load-bearing proof; all selection strategies share the same pool, and the DFT reference values used for validation are recomputed in this paper. The one real caveat is mechanistic, not circular: Eq. (21) sums row leverage scores without normalizing by |I_R|, and the paper's own Table I lists 1–10 atoms per cell (so |I_R| = 1+3N ranges from 4 to 31), making the sentence 'variations in |I_R| are limited and do not introduce a systematic bias in the ranking' questionable. But that is a potential confound or alternative explanation (cell-size selection), not a definitional reduction: the selection score is not defined in terms of, nor fitted to, the target test errors, and no self-citation chain is used to force the conclusion. Therefore no significant circularity is present.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities. Its free parameters are the ACE architecture and fit/regularization hyperparameters, several of which (lambda, beta, Sigma_0) are not reported. The central assumptions are the representational adequacy of linear ACE, the richness of ASSYST pools, and the correctness of the configuration-level leverage aggregation, whose stated justification conflicts with the pool's stated cell-size range.

free parameters (7)
  • ACE body order nu = 3
    Chosen from the hyperparameter scan (Fig. 1) as lying in the low-error region; defines the descriptor/basis size that the leverage ranking operates on.
  • totaldegree truncation td = 17
    Total-degree truncation chosen from the scan; sets the basis size for each correlation order and hence the design matrix geometry.
  • radial cutoff r_cut = 6.5 A
    Chosen from the scan; larger cutoffs consistently reduced errors (Fig. 1), so this choice affects all reported accuracies.
  • angular/radial degree weight w_L = 1.5
    Fixed default in the mixed-degree criterion (Eq. 7); controls which descriptors exist and therefore the leverage geometry.
  • energy/force fit weights w_E, w_F = 1.0 / 0.1
    Chosen fit weights (Sec. II.C); the block-CUR vs CUR force advantage depends on this balance, and the weights enter the weighted design matrix used for leverage.
  • regularization strength lambda / prior precisions (beta, Sigma_0) = not reported
    Enter the regularized leverage scores (Eq. 19) and the fit (Eq. 11); values are not given in the paper, so the reader cannot reproduce the ranking.
  • leverage regularization beta (Eq. 19) = not reported
    Scales the data precision inside the regularized hat matrix; unspecified, despite being load-bearing for the selection criterion.
assumptions (6)
  • domain assumption ACE energy decomposition E = sum_i epsilon_i with symmetry-adapted body-ordered basis (Eqs. 1-7) is an adequate representation for Al, Cu, and Al-Cu within the chosen truncation.
    Adopted from ACEpotentials.jl / Refs. [12,37,38]; the entire leverage geometry is computed in this basis, so errors of the ACE representation bound the selection.
  • domain assumption ASSYST pools (Sec. II.A, Table I) are a rich, unbiased sample of configurational space such that a 30-40% subset can still cover the descriptor directions needed for bulk and defect prediction.
    If the pool lacks relevant environments (e.g., certain low-coordination or long-range environments), selection cannot recover them; this bounds the absolute accuracy of any subset strategy.
  • domain assumption Regularized leverage scores (Eqs. 19-20) rank information content of configurations (approximately maximizing posterior information gain).
    Standard optimal-design/hat-matrix interpretation carried from Refs. [25,34] into the Bayesian linear setting; stated, not proven, and used as the selection criterion.
  • ad hoc to paper Configuration-level leverage L_R = sum_i h_i (Eq. 21) is unbiased because 'variations in |I_R| are limited' (Sec. II.D.1).
    Table I states 1-10 atoms per element, so row counts span roughly 3-30; the assumption that sums are comparable appears inconsistent with the stated pool parameters. Load-bearing for the ranking.
  • standard math Gaussian prior/posterior (Eqs. 12-15) with posterior mean as deployed model; Tikhonov regularization on degree-based smoothness.
    Standard Bayesian linear regression; the prior is structural (degree-based), not fit to labels.
  • domain assumption Five independent runs with 80/20 splits give 'within statistical uncertainty' comparisons.
    Confidence intervals are read from 5 runs; statistical power to detect differences at the plateau is modest, and the 'within uncertainty' phrasing carries the efficiency claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection of ASSYST Structure Pools." pith.science (2026). https://pith.science/paper/EEDQ5BR6

@misc{pith2026260718524,
  author       = {Pith},
  title        = {Pith review of: Data-Efficient Training of Linear ACE Potentials through Leverage-Guided Subset Selection of ASSYST Structure Pools},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EEDQ5BR6}},
  note         = {Machine review of arXiv:2607.18524}
}
read the original abstract

The construction of machine-learned interatomic potentials (MLIPs) is often limited by the cost of generating large density-functional-theory (DFT) training datasets. For systematically generated structure pools such as ASSYST, a central practical question is how many configurations must be labeled to achieve reliable accuracy. Here we assess geometry-based, label-free subset selection for training linear Atomic Cluster Expansion (ACE) potentials. Using statistical leverage scores and CUR-type sampling, we compare leverage-guided selection against random, energy-based, and force-based baselines under controlled iterative protocols. Elemental Al provides the primary benchmark, with Cu and Al-Cu alloys used for transfer validation. Leverage-guided subsets recover plateau-level energy and force accuracy using substantially smaller labeled fractions (approximately 30-40%) than random sampling, corresponding to an effective 2-3x reduction in DFT labeling for the systems studied. In alloy tests, defect energetics remain comparable across strategies once sufficient chemical diversity is included, while leverage selection maintains competitive accuracy at reduced training size. These results demonstrate that descriptor-space-guided, label-free subsampling can significantly reduce DFT workload for linear ACE models trained on ASSYST structure pools without degrading defect-level fidelity.

Figures

Figures reproduced from arXiv: 2607.18524 by the authors.

Figure 1
Figure 1. FIG. 1: Hyperparameter sensitivity of linear ACE models for Al (top) and Cu (bottom), evaluated using held-out [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Test energy and force learning curves for Al. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Energy and force learning curves on Cu, [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: FIG. 5: Error distributions on the independent DFT validation set for the different subset-selection strategies. The [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6: RMSE summary on the independent DFT [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 3 canonical work pages

  1. [1]

    Thus, the stacked vectorycollects all observables from all configurations in the candidate pool

    Configuration-level leverage scores Within the linear ACE framework, fitting the model reduces to a weighted linear regression problem of the form y≈Ac, where each row of the global design matrixAcontains the ACE basis functions evaluated for a single observable (an energy, force component, or stress component), andc denotes the vector of ACE coefficients...

  2. [2]

    block-CUR

    Leverage-guided CUR and block-CUR subsampling We now use these configuration-level leverage scores to construct compact, informative training subsets. Our strategy is inspired by CUR-type decompositions and op- timal experimental design for high-dimensional feature matrices [28, 34]: training-set design is viewed as anof- flinesubset-selection problem on ...

  3. [3]

    label–then–filter

    Baseline selection strategies To benchmark the leverage-based, label-free methods introduced below, we consider three reference selection strategies. All methods operate within the same incremental la- belling protocol. For each run, the training set is ini- tialized by randomly sampling 10% of the available con- figurations from the ASSYST pool. A provis...

  4. [4]

    Mueller, A

    T. Mueller, A. Hernandez, and C. Wang, J. Chem. Phys. 152, 050902 (2020)

    2020

  5. [5]

    Behler, Angewandte Chemie International Edition56, 12828–12840 (2017)

    J. Behler, Angewandte Chemie International Edition56, 12828–12840 (2017)

    2017

  6. [6]

    V. L. Deringer, M. A. Caro, and G. Cs´ anyi, Advanced Materials31(2019), 10.1002/adma.201902765

    doi:10.1002/adma.201902765 2019

  7. [7]

    No´ e, A

    F. No´ e, A. Tkatchenko, K.-R. M¨ uller, and C. Clementi, Annual Review of Physical Chemistry71, 361–390 (2020)

    2020

  8. [8]

    Behler, The Journal of Chemical Physics145(2016), 10.1063/1.4966192

    J. Behler, The Journal of Chemical Physics145(2016), 10.1063/1.4966192

    doi:10.1063/1.4966192 2016

Show all 39 references
  1. [9]

    O. T. Unke, S. Chmiela, H. E. Sauceda, M. Gastegger, I. Poltavsky, K. T. Sch¨ utt, A. Tkatchenko, and K.-R. M¨ uller, Chemical Reviews121, 10142–10186 (2021)

    2021

  2. [10]

    Friederich, F

    P. Friederich, F. H¨ ase, J. Proppe, and A. Aspuru-Guzik, Nature Materials20, 750–761 (2021)

    2021

  3. [11]

    Behler and M

    J. Behler and M. Parrinello, Physical Review Letters98 (2007), 10.1103/physrevlett.98.146401

    2007 doi

  4. [13]

    A. V. Shapeev, Multiscale Modeling & Simulation 14, 1153–1173 (2016)

    2016

  5. [14]

    I. S. Novikov, K. Gubaev, E. V. Podryabinkin, and A. V. Shapeev, Machine Learning: Science and Technology2, 025002 (2021)

    2021

  6. [16]

    Dusson, M

    G. Dusson, M. Bachmayr, G. Cs´ anyi, R. Drautz, S. Etter, C. van der Oord, and C. Ortner, Journal of Computa- tional Physics454, 110946 (2022)

    2022

  7. [17]

    Thompson, L

    A. Thompson, L. Swiler, C. Trott, S. Foiles, and G. Tucker, Journal of Computational Physics285, 316–330 (2015)

    2015

  8. [18]

    M. A. Wood and A. P. Thompson, The Journal of Chem- ical Physics148(2018), 10.1063/1.5017641

    2018 doi

  9. [19]

    Batatia, D

    I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Csanyi, inAdvances in Neural Information Process- ing Systems, Vol. 35, edited by S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Curran Associates, Inc., 2022) pp. 11423–11436

    2022

  10. [20]

    O. T. Unke and M. Meuwly, Journal of Chemical Theory and Computation15, 3678–3693 (2019)

    2019

  11. [21]

    Zubatyuk, J

    R. Zubatyuk, J. S. Smith, J. Leszczynski, and O. Isayev, Science Advances5(2019), 10.1126/sciadv.aav6490

    2019 doi

  12. [22]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nature Communications13(2022), 10.1038/s41467-022-29939-5

    2022 doi

  13. [23]

    K. T. Sch¨ utt, O. T. Unke, and M. Gastegger, inInter- national Conference on Machine Learning(2021)

    2021

  14. [24]

    Zhang, J

    L. Zhang, J. Han, H. Wang, R. Car, and W. E, Physical Review Letters120(2018), 10.1103/phys- revlett.120.143001

    2018 doi

  15. [25]

    H. Wang, L. Zhang, J. Han, and W. E, Computer Physics Communications228, 178–184 (2018)

    2018

  16. [26]

    S. Mohr, L. E. Ratcliff, L. Genovese, D. Caliste, P. Boulanger, S. Goedecker, and T. Deutsch, Physical Chemistry Chemical Physics17, 31360–31370 (2015)

    2015

  17. [27]

    M. S. Daw and M. I. Baskes, Physical Review B29, 6443–6453 (1984)

    1984

  18. [28]

    Optimal design of exper- iments in the context of machine-learning inter-atomic potentials: improving the efficiency and transferability of kernel based methods,

    B. Barzdajn and C. P. Race, “Optimal design of exper- iments in the context of machine-learning inter-atomic potentials: improving the efficiency and transferability of kernel based methods,” (2024), arXiv:2405.08636 [cond- mat.mtrl-sci]

    2024 arXiv

  19. [29]

    T. B. Blank, S. D. Brown, A. W. Calhoun, and D. J. Doren, The Journal of Chemical Physics103, 4129–4137 (1995)

    1995

  20. [30]

    Allen and A

    C. Allen and A. P. Bart´ ok, Mach. Learn. Sci. Technol.3, 045031 (2022)

    2022

  21. [31]

    Rocke and J

    T. Rocke and J. R. Kermode, Model. Simul. Mat. Sci. Eng.33, 055020 (2025)

    2025

  22. [32]

    Y. Liu, J. D. Morrow, C. Ertural, N. L. Fragapane, J. L. A. Gardner, A. A. Naik, Y. Zhou, J. George, and V. L. Deringer, Nature Communications16(2025), 10.1038/s41467-025-62510-6

    2025 doi

  23. [33]

    PA Subramanyam and D

    A. PA Subramanyam and D. Perez, npj Computational Materials11, 218 (2025)

    2025

  24. [34]

    V. G. Fletcher, A. P. Bart´ ok, and L. B. P´ artay, npj Computational Materials (2025)

    2025

  25. [35]

    Baghishov, J

    I. Baghishov, J. Janssen, G. Henkelman, and D. Perez, Digital Discovery5, 332 (2025)

    2025

  26. [36]

    Active learning methods based on statistical leverage scores,

    C. Orhan and O. Tastan, “Active learning methods based on statistical leverage scores,” (2018), arXiv:1812.02497 [cs.LG]

    2018 arXiv

  27. [37]

    R. K. Cersonsky, B. A. Helfrecht, E. A. Engel, S. Kli- avinek, and M. Ceriotti, Mach. Learn. Sci. Technol.2, 035038 (2021)

    2021

  28. [38]

    M. Poul, L. Huber, E. Bitzek, and J. Neugebauer, Phys- ical Review B107(2023), 10.1103/physrevb.107.104103

    2023 doi

  29. [39]

    M. Poul, L. Huber, and J. Neugebauer, npj Computa- 16 tional Materials11(2025), 10.1038/s41524-025-01669-4

    2025 doi

  30. [40]

    W. C. Witt, C. van der Oord, E. Gelˇ zinyt˙ e, T. J¨ arvinen, A. Ross, J. P. Darby, C. H. Ho, W. J. Baldwin, M. Sachs, J. Kermode, N. Bernstein, G. Cs´ anyi, and C. Or- tner, The Journal of Chemical Physics159(2023), 10.1063/5.0158783

    2023 doi

  31. [41]

    Bochkarev, Y

    A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Phys. Rev. X14, 021036 (2024)

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.