Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Misspecification error bars catch every quantum reference in tungsten

desk verdict Solid empirical validation of a heuristic UQ ansatz; the advertised theory has a real gap, but the paper earns a serious referee. read the letter →

arxiv 2502.07104 v2 pith:3T74X7B4 submitted 2025-02-10 cond-mat.mtrl-sci

Danny Perez , Aparna P. A. Subramanyam , Ivan Maliyov , Thomas D. Swinburne This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords uncertaintyquantificationmisspecificationmachine-learnedinteratomicpotentialsPOPSBayesianregressiontungstenfoundationmodelserrorbounds
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 the largest uncertainty in modern machine-learned interatomic potentials is model misspecification: no single set of parameters can exactly reproduce all reference quantum calculations. Standard Bayesian and loss-based uncertainty estimates are shown to ignore this source of error and therefore to be overconfident. The paper applies a misspecification-aware scheme, POPS, that builds an ensemble of plausible parameter sets and propagates it through atomistic simulations of tungsten. Its central result is that the ensemble's error bars contain the reference values in every property tested, from pointwise energies and forces to crystal stabilities, phonons, defect energies, and migration barriers. The same scheme, added as a linear corrector to a universal neural-network potential, also bounds its energy errors across a diverse database.

What carries the argument

The load-bearing object is the POPS-hypercube ansatz. For each training point $X$, the pointwise optimal parameter set (POPS) is the set of all parameters $\Theta$ with $M(X;\Theta)=E(X)$, and the paper shows that any posterior minimizing the generalization error must put mass in every such set. Because misspecification makes the mutual intersection of these sets empty, the true parameter uncertainty is finite and cannot be captured by a delta function at the global loss minimizer. POPS therefore collects, for each training point, the parameter vector that minimizes the global loss while lying in that point's POPS, then takes the posterior as uniform over the minimal axis-aligned hypercube containing all of them. This reduces uncertainty quantification to resampling from a hypercube at $O(P)$ cost for $P$ parameters, and the ensemble spread carries the misspecification signal. For stationary properties, an implicit differentiation of the energy minimum with respect to parameters propagates the ensemble without re-minimizing each sampled potential.

What would settle it

Train the same qSNAP model with POPS for a different element or with a different descriptor set, then compute a property far outside the training distribution—for example, a grain-boundary formation energy or a high-temperature free energy—and compare the ensemble's maximum and minimum bounds to converged reference quantum calculations; if the reference falls outside the bounds in more than the observed ~1–3% pointwise envelope-violation rate, the coverage claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a misspecification-aware posterior built from pointwise optimal parameter sets yields calibrated, conservative uncertainty bounds for the large class of interatomic potentials that are linear in their descriptors, and for linear correctors on nonlinear foundation models. In the regime of abundant, near-deterministic training data, the dominant error is not aleatoric noise or finite-data epistemic uncertainty but the inability of any parameter choice to fit all data exactly. POPS defines, for each training point, the set of parameters that reproduce that point exactly, and then places a uniform posterior over the smallest hypercube containing the corresponding constrained loss minimizers. Propagating this ensemble through resampled simulations gives prediction spreads whose standard deviation tracks the true error, and whose maximum and minimum bracket the reference value in all 13 crystal structures, the phonon spectra, five self-interstitial configurations, eleven surfaces, and the vacancy migration barrier, with only 2.1% and 3.3% pointwise envelope violations for energies and forces. Applied to the MACE-MPA-0 foundation model as a linear corrector, the same procedure bounds test energy errors with a 1% envelope violation.

Load-bearing premise

The method assumes that the true misspecification posterior is well represented by a uniform distribution over the smallest hypercube containing all pointwise-optimal parameter sets, and that the manually chosen energy and force weights in Eq. (9) do not bias this ensemble.

Editorial extensions

If this is right

  • For any potential expressible as a linear combination of fixed nonlinear features, such as SNAP or ACE, POPS adds uncertainty quantification at negligible extra cost over the least-squares fit.
  • Error bars can be propagated to energy-minimized properties such as defect energies, volumes, and barriers with less than 4% error using implicit differentiation, avoiding brute-force resampling of every ensemble member.
  • Users can assess whether predicted rankings of properties are meaningful by comparing the MLE's rank correlation to the ensemble's distribution of rank correlations, without needing reference data.
  • The same machinery bounds the energy errors of a universal message-passing neural network potential when applied as a linear corrector, pointing toward uncertainty quantification for foundation models.
  • The ensemble's maximum and minimum predictions act as conservative worst-case bounds that can be passed downstream in multiscale modeling workflows.

Reading between the lines

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

  • Inference: Because the hypercube posterior is axis-aligned and uniform, strongly correlated parameters could inflate the worst-case bounds; testing a covariance-aware posterior shape on the same tungsten suite would show whether the observed conservatism is an artifact of the ansatz.
  • Inference: The manual energy and force weights in Eq. (9) shape which training points dominate the POPS ensemble, so perturbing these weights by factors of two and checking whether envelope coverage survives would isolate how much of the result depends on those choices.
  • Inference: If the bounds remain conservative across elements and properties, the ensemble's spread could serve as an acquisition score for active learning, flagging configurations where the misspecification band is widest and reducing the number of reference calculations needed.
  • Inference: A head-to-head comparison with query-by-committee ensembles and conformal prediction on the same validation suite would quantify how much tighter misspecification-aware bounds are than calibrated alternatives.
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

4 major / 6 minor

Summary. The paper presents a framework, POPS-hypercube, for quantifying misspecification uncertainty in machine-learned interatomic potentials that are linear in their parameters (or, for foundation models, in a linear corrector layer). The method constructs a posterior by taking the minimal hypercube that contains a set of pointwise-optimal parameter solutions (POPS) and sampling uniformly from it. The authors apply this to a qSNAP tungsten model and to a linear corrector for MACE-MPA-0, validating ensemble spread and min/max envelopes against DFT reference data for pointwise energies/forces, crystal structures, phonons, defects, surfaces, and energy barriers. They report that the ensemble bounds bracket the DFT reference in all property-level tests, with small pointwise envelope violation rates (2.1% energies, 3.3% forces), and they demonstrate fast propagation via implicit differentiation. The central claim is that propagated misspecification uncertainties robustly envelope errors to ab initio calculations outside the training dataset.

Significance. If the envelope-coverage results hold beyond the specific test cases, the method has genuine practical value: it offers a cheap, resampling-based UQ procedure for a broad class of linear MLIAPs and for linear correctors on modern universal potentials. The paper's strengths include a diverse validation suite against external DFT data (so the coverage claim is not circular), an open-source implementation, and an honest reporting of pointwise envelope violations rather than overclaiming perfect coverage. The fast implicit-differentiation propagation in Sec. III F is a useful contribution to the multiscale-UQ toolkit. However, the theoretical justification for the hypercube posterior is internally inconsistent with the stated optimality condition (see Major Comment 1), and the application to foundation models is narrower than the abstract suggests. The significance is therefore contingent on reframing the method as a heuristic with strong empirical support rather than as a principled minimizer of the generalization error.

major comments (4)
  1. [Sec. II B] The POPS-hypercube ansatz contradicts the theoretical condition used to justify it. The text states that any posterior minimizing the generalization error G must have mass in every pointwise-optimal parameter set (POPS), since otherwise the integrand exp(-||M(X;Theta)-E(X)||^2/epsilon^2) cannot remain O(1) as epsilon -> 0. The paper then defines pi_H as a uniform distribution over the minimal hypercube containing the POPS-constrained loss minimizers. For a linear model with P parameters, each POPS is generically a hyperplane of codimension at least 1, hence a Lebesgue-null set, so a uniform density assigns it zero probability. Consequently the integral defining G[pi_H] is O(epsilon) and diverges as epsilon -> 0, meaning pi_H is not an approximate minimizer of G in the sense used in the derivation. This internal inconsistency removes the advertised theoretical guarantee for the 'robust envelope' claim. The authors should either prove a precise relaxed sense in which pi_H is near-optimal (e.g., with a finite-epsilon analysis) or explicitly state that the hypercube choice is a heuristic ansatz, and then soften the corresponding optimality language throughout the paper.
  2. [Sec. III G, Eq. (10)] The application to MACE-MPA-0 is a linear corrector of the form E_DFT - E_MACE - Theta·D_i, not an uncertainty quantification of the foundation model itself. The POPS posterior is defined over the corrector parameters only, while the MACE backbone is frozen. The abstract and introduction state that the framework 'demonstrate[s] application to recent foundational machine learning interatomic potentials' and 'bound[s] errors in MACE-MPA-0 energy predictions'; this overstates the scope. The results show that one can envelope the residual error of a frozen backbone with a linear corrector, which is useful, but it is not the same as quantifying the uncertainty of the foundation model's own predictions. Please qualify the claims in the abstract and Sec. III G accordingly.
  3. [Secs. III B–III E, Tables I–V] The central coverage claim rests on the min/max envelope of 500 ensemble samples. This is a finite-sample statistic and will, with high probability, underestimate the true support of the hypercube posterior, especially for a high-dimensional parameter space. The statement in Sec. III B that 'in all cases, the extreme values predicted by the ensemble bound the actual reference result, providing strong guarantees' is not backed by any statistical guarantee or by an analysis of how the envelope depends on ensemble size. Moreover, the weights in Eq. (9) are hand-chosen (a = 2 eV, b = 50 eV/Å) and the ensemble/hypercube inherits this choice. The paper should provide a sensitivity analysis with respect to these weights and to the ensemble size, or at minimum explicitly acknowledge that the envelope property is an empirical finding for this particular model and weighting, not a certified bound.
  4. [Sec. III A] The reported pointwise envelope violation rates of 2.1% for energies and 3.3% for forces are non-negligible, yet the paper's abstract and conclusions describe the propagated uncertainties as 'robustly envelope[ing]' errors. For a method presented as providing worst-case bounds, violation rates of several percent deserve a dedicated discussion: are these expected finite-sample effects of using the max/min over 500 draws, or systematic underestimation of the hypercube tails? The paper should either provide a calibration analysis that accounts for the number of ensemble members or temper the language from 'bound' to 'likely bracketing' in the pointwise setting.
minor comments (6)
  1. [Eq. (1)] The Gaussian density is typeset incorrectly: the displayed formula has a missing minus sign and an unbalanced parenthesis; it should read exp(-||E(X)-Y||^2/epsilon^2)/sqrt(pi epsilon^2).
  2. [Sec. II B] The sentence 'the POPS-hypercube posterior can then be resampled for only O(P) computational effort' is ambiguous: it should state whether O(P) is the cost per sample or the total cost to generate the ensemble of N samples.
  3. [Sec. II C] The solution obtained by minimizing the weighted least-squares objective with weights (9) is referred to as the 'MLE solution', but this is not the maximum-likelihood estimator under the Gaussian likelihood (1) once the arbitrary weights are introduced. Please clarify the terminology.
  4. [Sec. III B] There is a typo: 'the actual error incured by the MLE MLIAP' should read 'incurred'.
  5. [Fig. 11] The caption or text states 'within less then 4% of error'; the word 'then' should be 'than'.
  6. [Sec. II D] The paper does not state how many random seeds were used for the uniform resampling of the 500-model hypercube ensemble, nor the statistical uncertainty on the reported envelope violation rates. Reporting a confidence interval would strengthen the empirical claims.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the POPS method is self-cited but the uncertainty-envelope claims are validated against external DFT benchmarks.

full rationale

The paper's central claims are empirical: the POPS-hypercube ensemble, built from the qSNAP tungsten training data and a MACE linear corrector, is shown to produce uncertainty bounds that contain direct DFT reference values on held-out properties. The validation targets (crystal phase formation energies, volumes, bulk moduli, phonons, defect formation energies, migration barriers, MACE test-set energies) are external to the ensemble construction; the weights in Eq. (9) are fixed rather than fitted to these targets, and Sec. II C states the potential was not fine-tuned. The only material self-citation is Ref. [30] (Swinburne and Perez), which supplies the POPS methodology, and Ref. [32] (Maliyov et al.) for implicit differentiation; these provide the method but do not by themselves establish the reported envelope coverage, which is independently benchmarked against DFT. I therefore find no reduction of a claimed prediction to an input by construction. A separate rigor caveat, not a circularity: Sec. II B imports the theorem that a generalization-error minimizer must have mass in every POPS, whereas the uniform hypercube posterior assigns zero Lebesgue measure to each POPS hyperplane for linear models, leaving the theoretical justification of the ansatz incomplete despite the successful empirical demonstration.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. It relies on the training-data assumptions, the misspecification regime, and the POPS hypercube ansatz as the main load-bearing premises. The only free parameters are the manual weighting scales and the ensemble size, which affect the posterior but are not fitted to the validation targets.

free parameters (3)
  • energy weighting scale a = 2 eV
    Manual choice in Eq. (9) for weighting near-equilibrium energies in the loss; not optimized, but shapes the POPS ensemble and thus the propagated bounds.
  • force weighting scale b = 50 eV/Angstrom
    Manual choice in Eq. (9) for weighting force residuals; as above, affects the ensemble and the resulting uncertainty estimates.
  • hypercube resampling ensemble size = 500
    Number of parameter vectors uniformly sampled from the minimal hypercube; chosen for computational cost, and determines the width and stability of the reported envelope bounds.
assumptions (4)
  • domain assumption DFT training data is near-deterministic with vanishing aleatoric error.
    Sec. II A, condition 1. Justifies the misspecification-dominated regime; if the reference data were noisy, the method would need modification.
  • domain assumption The qSNAP model is misspecified and underparameterized relative to the training set.
    Sec. II A, conditions 2 and 3. Ensures the pointwise-optimal parameter sets are nonempty and the enclosing hypercube is finite.
  • ad hoc to paper The uniform distribution over the minimal hypercube containing all POPS-constrained loss minimizers approximates the posterior that minimizes generalization error.
    Sec. II B. This is the central ansatz, asserted from Ref. [30] and validated empirically in this paper; it is not derived here.
  • domain assumption The validation properties are not represented in the training set and are representative of genuinely unseen configurations.
    Sec. II E. If the validation configurations were close to training points, the envelope coverage would be less informative about extrapolative uncertainty.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials." pith.science (2026). https://pith.science/paper/3T74X7B4

@misc{pith2026250207104,
  author       = {Pith},
  title        = {Pith review of: Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3T74X7B4}},
  note         = {Machine review of arXiv:2502.07104}
}
read the original abstract

The use of high-dimensional regression techniques from machine learning has significantly improved the quantitative accuracy of interatomic potentials. Atomic simulations can now plausibly target quantitative predictions in a variety of settings, which has brought renewed interest in robust means to quantify uncertainties on simulation results. In many practical settings, encompassing both classical and a large class of machine learning potentials, the dominant form of uncertainty is currently not due to lack of training data but to misspecification, namely the inability of any one choice of model parameters to exactly match all ab initio training data. However, Bayesian inference, the most common formal tool used to quantify uncertainty, is known to ignore misspecification and thus significantly underestimates parameter uncertainties. Here, we employ a recent misspecification-aware regression technique to quantify parameter uncertainties, which is then propagated to a broad range of phase and defect properties in tungsten via brute force resampling or implicit differentiation. The propagated misspecification uncertainties robustly envelope errors to direct \textit{ab initio} calculation of material properties outside of the training dataset, an essential requirement for any quantitative multi-scale modeling scheme. Finally, we demonstrate application to recent foundational machine learning interatomic potentials, accurately predicting and bounding errors in MACE-MPA-0 energy predictions across the diverse materials project database. Perspectives for the approach in multiscale simulation workflows are discussed.

Figures

Figures reproduced from arXiv: 2502.07104 by the authors.

Figure 1
Figure 1. FIG. 1. Distribution of regression coefficients over the loss-minimizing POPS ensemble [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Absolute value of the covariance of the regression coefficients over the loss-minimizing [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Characterization of the statistics of pointwise errors obtained from [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Left: Equilibrium formation energy for different crystal structures relative to the BCC [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Spearman rank correlation analysis for various quantities. Blue: histogram of rank [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Transformation paths between different crystal phases. MLE predictions are shown in [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Left: phonon dispersion in the BCC phase. Right: vibrational density of states. MLE [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Self-interstitial formation energies in the BCC phase. MLE predictions are shown by blue [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Surface formation energies in the BCC phase. MLE predictions are shown by blue bars, [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. Minimum energy pathway for a first-neighbor vacancy hop in BCC. MLE predictions [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11. Implicit derivative predictions vs true molecular statics minimization for the UQ ensemble [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12. Characterization of the statistics of pointwise energy errors of the MACE-MPA-0 [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Model Accuracy and Data Heterogeneity Shape Uncertainty Quantification in Machine Learning Interatomic Potentials

    cond-mat.mtrl-sci 2025-08 unverdicted novelty 5.0 of 10

    Higher model accuracy improves uncertainty-error correlation and novelty detection in MLIP UQ, and clustering-enhanced local D-optimality better detects novel environments on heterogeneous datasets.

Reference graph

Works this paper leans on

76 extracted references · 65 canonical work pages · cited by 1 Pith paper

  1. [30]

    I. R. Best, T. J. Sullivan, and J. R. Kermode, The Journal of Chemical Physics 161, 064112 (2024)

    work page 2024

  2. [1]

    The reference data (here DFT energies and forces) is near-deterministic, i.e., it exhibits vanishing aleatoric errors

    work page

  3. [2]

    The ML model is misspecified, i.e., no single choice of the free parameters can repro- duce all reference data exactly

    work page

  4. [3]

    The model is underparameterized, i.e., the amount of training data significantly ex- ceeds the number of trainable parameters In the context of MLIAPs, condition 1 reflects the near-deterministic nature of well- converged quantum calculations, where repeated calculations with the same inputs result in the same output. While some MLIAP formalism provide co...

    work page

  5. [4]

    A. M. Goryaeva, J. D´ er` es, C. Lapointe, P. Grigorev, T. D. Swinburne, J. R. Kermode, L. Ven- telon, J. Baima, and M.-C. Marinica, Phys. Rev. Materials 5, 103803 (2021)

    work page 2021

  6. [5]

    Behler, The Journal of chemical physics 145 (2016)

    J. Behler, The Journal of chemical physics 145 (2016)

    work page 2016

  7. [6]

    V. L. Deringer, M. A. Caro, and G. Cs´ anyi, Advanced Materials 31, 1902765 (2019)

    2019

  8. [7]

    Mishin, Acta Materialia 214, 116980 (2021)

    Y. Mishin, Acta Materialia 214, 116980 (2021)

    work page 2021

Show all 76 references
  1. [8]

    Stocker, J

    S. Stocker, J. Gasteiger, F. Becker, S. G¨ unnemann, and J. T. Margraf, Machine Learning: Science and Technology 3, 045010 (2022)

    2022

  2. [9]

    Lysogorskiy, C

    Y. Lysogorskiy, C. van der Oord, A. Bochkarev, S. Menon, M. Rinaldi, T. Hammerschmidt, M. Mrovec, A. Thompson, G. Cs´ anyi, C. Ortner, et al. , npj Computational Materials 7, 1 (2021)

    2021

  3. [10]

    Mortazavi, X

    B. Mortazavi, X. Zhuang, T. Rabczuk, and A. V. Shapeev, Materials Horizons 10, 1956 (2023)

    2023

  4. [11]

    Forces are not enough: Benchmark and critical evaluation for machine learning force fields with 29 molecular simulations,

    X. Fu, Z. Wu, W. Wang, T. Xie, S. Keten, R. Gomez-Bombarelli, and T. Jaakkola, “Forces are not enough: Benchmark and critical evaluation for machine learning force fields with 29 molecular simulations,” (2023), arXiv:2210.07237 [physics.comp-ph]

    2023 arXiv

  5. [12]

    S. R. Xie, M. Rupp, and R. G. Hennig, npj Computational Materials 9, 162 (2023)

    2023

  6. [13]

    Montes de Oca Zapiain, M

    D. Montes de Oca Zapiain, M. A. Wood, N. Lubbers, C. Z. Pereyra, A. P. Thompson, and D. Perez, npj Computational Materials 8 (2022), 10.1038/s41524-022-00872-x

    2022 doi

  7. [14]

    Alexander, A

    F. Alexander, A. Almgren, J. Bell, A. Bhattacharjee, J. Chen, P. Colella, D. Daniel, J. DeSlippe, L. Diachin, E. Draeger, et al., Philosophical Transactions of the Royal Society A 378, 20190056 (2020)

    2020

  8. [15]

    Gavini, S

    V. Gavini, S. Baroni, V. Blum, D. R. Bowler, A. Buccheri, J. R. Chelikowsky, S. Das, W. Daw- son, P. Delugas, M. Dogan, et al. , Modelling and Simulation in Materials Science and Engi- neering 31, 063301 (2023)

    2023

  9. [16]

    J. S. Smith, B. Nebgen, N. Mathew, J. Chen, N. Lubbers, L. Burakovsky, S. Tretiak, H. A. Nam, T. Germann, S. Fensin, and K. Barros, Nature Communications 12 (2021), 10.1038/s41467-021-21376-0

    2021 doi

  10. [17]

    Hegde, E

    A. Hegde, E. Weiss, W. Windl, H. N. Najm, and C. Safta, International Journal for Uncer- tainty Quantification (2024)

    2024

  11. [18]

    E. V. Podryabinkin and A. V. Shapeev, Computational Materials Science 140, 171 (2017)

    2017

  12. [19]

    Zaverkin, D

    V. Zaverkin, D. Holzm¨ uller, H. Christiansen, F. Errica, F. Alesiani, M. Takamoto, M. Niepert, and J. K¨ astner, npj Computational Materials10, 83 (2024)

    2024

  13. [20]

    Longbottom and P

    S. Longbottom and P. Brommer, Modelling and Simulation in Materials Science and Engi- neering 27, 044001 (2019)

    2019

  14. [21]

    Kulichenko, K

    M. Kulichenko, K. Barros, N. Lubbers, Y. W. Li, R. Messerly, S. Tretiak, J. S. Smith, and B. Nebgen, Nature Computational Science 3, 230 (2023)

    2023

  15. [22]

    Kurniawan, C

    Y. Kurniawan, C. L. Petrie, K. J. Williams, M. K. Transtrum, E. B. Tadmor, R. S. Elliott, D. S. Karls, and M. Wen, The Journal of Chemical Physics 156 (2022)

    2022

  16. [23]

    Hegde, E

    A. Hegde, E. Weiss, W. Windl, H. Najm, and C. Safta, Computational Materials Science 214, 111660 (2022)

    2022

  17. [24]

    J. Busk, M. N. Schmidt, O. Winther, T. Vegge, and P. B. Jørgensen, Physical Chemistry Chemical Physics 25, 25828 (2023)

    2023

  18. [25]

    A. Zhu, S. Batzner, A. Musaelian, and B. Kozinsky, The Journal of Chemical Physics 158, 164111 (2023). 30

    2023

  19. [26]

    A. R. Tan, S. Urata, S. Goldman, J. C. Dietschreit, and R. G´ omez-Bombarelli, npj Compu- tational Materials 9, 225 (2023)

    2023

  20. [27]

    Y. Hu, J. Musielewicz, Z. W. Ulissi, and A. J. Medford, Machine Learning: Science and Technology 3, 045028 (2022)

    2022

  21. [28]

    A. F. Psaros, X. Meng, Z. Zou, L. Guo, and G. E. Karniadakis, Journal of Computational Physics 477, 111902 (2023)

    2023

  22. [29]

    A. P. Bart´ ok, J. Kermode,et al. , arXiv preprint arXiv:2206.08744 (2022)

    2022 arXiv

  23. [31]

    Lahlou, M

    S. Lahlou, M. Jain, H. Nekoei, V. I. Butoi, P. Bertin, J. Rector-Brooks, M. Korablyov, and Y. Bengio, arXiv preprint arXiv:2102.08501 (2021)

    2021 arXiv

  24. [32]

    Maliyov, P

    I. Maliyov, P. Grigorev, and T. D. Swinburne, arXiv preprint arXiv:2407.02414 (2024)

    2024 arXiv

  25. [33]

    Masegosa, Advances in Neural Information Processing Systems 33, 5479 (2020)

    A. Masegosa, Advances in Neural Information Processing Systems 33, 5479 (2020)

    2020

  26. [34]

    Swinburne and D

    T. Swinburne and D. Perez, Machine Learning: Science and Technology (2025), 10.1088/2632- 2153/ad9fce

    2025 doi

  27. [35]

    Imbalzano, Y

    G. Imbalzano, Y. Zhuang, V. Kapil, K. Rossi, E. A. Engel, F. Grasselli, and M. Ceriotti, The Journal of Chemical Physics 154 (2021)

    2021

  28. [36]

    A. P. Bart´ ok, M. C. Payne, R. Kondor, and G. Cs´ anyi, Physical review letters 104, 136403 (2010)

    2010

  29. [37]

    Batatia, P

    I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kov´ acs, J. Riebesell, X. R. Advincula, M. Asta, W. J. Baldwin, N. Bernstein, et al. , arXiv preprint arXiv:2401.00096 (2023)

    2023 arXiv

  30. [38]

    B. Deng, P. Zhong, K. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder, Nature Machine Intelligence , 1–11 (2023)

    2023

  31. [39]

    Z. Li, J. R. Kermode, and A. De Vita, Phys. Rev. Lett. 114, 096405 (2015)

    2015

  32. [40]

    H. S. Seung, M. Opper, and H. Sompolinsky, in Proceedings of the fifth annual workshop on Computational learning theory (1992) pp. 287–294

    1992

  33. [41]

    Vandermause, S

    J. Vandermause, S. B. Torrisi, S. Batzner, Y. Xie, L. Sun, A. M. Kolpak, and B. Kozinsky, npj Computational Materials 6, 20 (2020)

    2020

  34. [42]

    On the uncer- tainty estimates of equivariant-neural-network-ensembles interatomic potentials,

    S. Lu, L. M. Ghiringhelli, C. Carbogno, J. Wang, and M. Scheffler, “On the uncer- tainty estimates of equivariant-neural-network-ensembles interatomic potentials,” (2023), arXiv:2309.00195 [cond-mat.mtrl-sci]. 31

    2023 arXiv

  35. [43]

    Blondel, Q

    M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-L´ opez, F. Pedregosa, and J.-P. Vert, Advances in neural information processing systems 35, 5230 (2022)

    2022

  36. [44]

    J. S. Smith, B. Nebgen, N. Lubbers, O. Isayev, and A. E. Roitberg, The Journal of chemical physics 148 (2018)

    2018

  37. [45]

    Artrith and J

    N. Artrith and J. Behler, Physical Review B—Condensed Matter and Materials Physics 85, 045439 (2012)

    2012

  38. [46]

    Zaverkin and J

    V. Zaverkin and J. K¨ astner, Machine Learning: Science and Technology 2, 035009 (2021)

    2021

  39. [47]

    Schran, K

    C. Schran, K. Brezina, and O. Marsalek, The Journal of Chemical Physics 153 (2020)

    2020

  40. [48]

    Breiman, Machine learning 24, 123 (1996)

    L. Breiman, Machine learning 24, 123 (1996)

    1996

  41. [49]

    Kellner and M

    M. Kellner and M. Ceriotti, Machine Learning: Science and Technology (2024)

    2024

  42. [50]

    to quantify parametric uncertainties [13, 18, 19, 51, 52], which can be extended to Bayesian NN [53]. In our recent work[30], discussed in more detail below, we address a known shortcoming of all Bayesian regression which minimize some expected loss, irrespective of model arch...

  43. [51]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, Advances in neural information processing systems 30 (2017)

    2017

  44. [52]

    R. K. Tripathy and I. Bilionis, Journal of computational physics 375, 565 (2018)

    2018

  45. [53]

    Kahle and F

    L. Kahle and F. Zipoli, Physical Review E 105, 015311 (2022)

    2022

  46. [54]

    C. M. Bishop and M. E. Tipping, in Advances in learning theory: methods, models and appli- cations (IOS Press, 2003) pp. 267–285

    2003

  47. [55]

    S. L. Frederiksen, K. W. Jacobsen, K. S. Brown, and J. P. Sethna, Physical review letters 93, 165501 (2004)

    2004

  48. [56]

    Williams, K

    L. Williams, K. Sargsyan, A. Rohskopf, and H. N. Najm, Computational Materials Science 242, 113074 (2024)

    2024

  49. [57]

    Goan and C

    E. Goan and C. Fookes, Case Studies in Applied Bayesian Data Science: CIRM Jean-Morlet Chair, Fall 2018 , 45 (2020)

    2020

  50. [58]

    M. J. Bayarri, J. O. Berger, R. Paulo, J. Sacks, J. A. Cafeo, J. Cavendish, C.-H. Lin, and J. Tu, Technometrics 49, 138 (2007)

    2007

  51. [59]

    Sargsyan, X

    K. Sargsyan, X. Huan, and H. N. Najm, International Journal for Uncertainty Quantification 9 (2019)

    2019

  52. [60]

    Y. Kato, D. M. Tax, and M. Loog, arXiv preprint arXiv:2210.16938 (2022)

    2022 arXiv

  53. [61]

    W. R. Morningstar, A. Alemi, and J. V. Dillon, in International Conference on Artificial Intelligence and Statistics (PMLR, 2022) pp. 8270–8298. 32

    2022

  54. [62]

    B. J. K. Kleijn and A. W. van der Vaart, The Annals of Statistics 34, 837 (2006)

    2006

  55. [63]

    Kleijn and A

    B. Kleijn and A. van der Vaart, Electronic Journal of Statistics 6, 354 (2012)

    2012

  56. [64]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and ´E. Duchesnay, Journal of Machine Learning Research 12, 2825 (2011)

    2011

  57. [65]

    Germain, F

    P. Germain, F. Bach, A. Lacoste, and S. Lacoste-Julien, Advances in Neural Information Processing Systems 29 (2016)

    2016

  58. [66]

    Hoeffding, The collected works of Wassily Hoeffding , 409 (1994)

    W. Hoeffding, The collected works of Wassily Hoeffding , 409 (1994)

    1994

  59. [67]

    Alquier, arXiv preprint arXiv:2110.11216 (2021)

    P. Alquier, arXiv preprint arXiv:2110.11216 (2021)

    2021 arXiv

  60. [68]

    A. P. Thompson, L. P. Swiler, C. R. Trott, S. M. Foiles, and G. J. Tucker, Journal of Computational Physics 285, 316 (2015)

    2015

  61. [69]

    M. A. Wood and A. P. Thompson, The Journal of chemical physics 148 (2018)

    2018

  62. [70]

    Karabin and D

    M. Karabin and D. Perez, The Journal of Chemical Physics 153 (2020)

    2020

  63. [71]

    Montes de Oca Zapiain, M

    D. Montes de Oca Zapiain, M. A. Wood, N. Lubbers, C. Z. Pereyra, A. P. Thompson, and D. Perez, npj Computational Materials 8, 189 (2022)

    2022

  64. [72]

    Information-entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials,

    A. Subramanyam and D. Perez, “Information-entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials,” (2024), arXiv:2407.10361

    2024 arXiv

  65. [73]

    Batatia, D

    I. Batatia, D. P. Kov´ acs, G. N. Simm, C. Ortner, and G. Cs´ anyi, arXiv preprint arXiv:2206.07697 (2022)

    2022 arXiv

  66. [74]

    Bochkarev, Y

    A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Phys. Rev. X 14, 021036 (2024)

    2024

  67. [75]

    Schmidt, N

    J. Schmidt, N. Hoffmann, H.-C. Wang, P. Borlido, P. J. Carri¸ co, T. F. Cerqueira, S. Botti, and M. A. Marques, Advanced Materials 35, 2210788 (2023)

    2023

  68. [76]

    Drautz, Phys

    R. Drautz, Phys. Rev. B 99, 014104 (2019). 33

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.