Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

MP-ALOE: An r2SCAN dataset for universal machine learning interatomic potentials

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Universal MLIPs trained on off-equilibrium r2SCAN data survive extreme pressure better.

desk verdict A genuinely useful public r2SCAN off-equilibrium dataset with plausible benchmark gains, but the headline comparisons lack statistical backing. read the letter →

arxiv 2507.05559 v1 pith:U76Q2PQE submitted 2025-07-08 cond-mat.mtrl-sci cs.AIphysics.comp-ph

Matthew C. Kuner , Aaron D. Kaplan , Kristin A. Persson , Mark Asta , Daryl C. Chrzan This is my paper
classification cond-mat.mtrl-scics.AIphysics.comp-ph
keywords machinelearninginteratomicpotentialsr2SCANdatasetactiveoff-equilibriumstructuresmoleculardynamicsstabilityextremepressureuniversaldensityfunctionaltheory
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

MP-ALOE is a public dataset of 909,792 r2SCAN density-functional-theory frames spanning 89 elements, deliberately weighted toward off-equilibrium geometries, high force magnitudes, and extreme hydrostatic pressures. The paper claims that a machine learning interatomic potential trained on this distribution is markedly more stable under extreme molecular dynamics conditions and more physical under extreme static deformations than one trained on the earlier near-equilibrium MatPES dataset, while staying competitive on equilibrium thermochemistry. If correct, this means that careful active-learning sampling of the potential energy surface, rather than only model architecture or total data volume, is what buys robustness at high pressure and temperature, and that the new dataset is a drop-in complement to MatPES for training universal potentials.

What carries the argument

The load-bearing mechanism is an active-learning loop. Roughly 100 million candidate structures are generated by substituting 89 elements into 817 small prototype crystals (binaries and ternaries, 2 to 8 atoms); a committee of pretrained universal potentials predicts each structure's energy, forces, and stress, and any structure on which the committee disagrees beyond thresholds of 100 meV/atom, 100 meV/Å, or 100 meV/ų is kept. The kept set of about 500,000 structures is downsampled to about 125,000 diverse structures using the DIRECT stratified-sampling procedure, then each structure is run through three ionic steps of r2SCAN DFT. This produces 909,792 frames whose pressure distribution reaches roughly ±100 GPa, versus about ±30 GPa for MatPES, and after each loop three fresh MACE graph-neural-network potentials are trained and become the next committee.

What would settle it

Retrain each potential several times with different random number seeds and rerun the energy-volume scan and NPT benchmarks; if the spread of failure rates across seeds overlaps the reported gaps, such as an MP-ALOE seed with more than 5% failures or a MatPES seed with less than 10%, the claimed advantage is not established.

Watch

Extended reading notes

Core claim

The central discovery is that concentrating DFT data on the regions where an ensemble of potentials disagrees, mostly high-energy, high-force, high-pressure hypothetical structures, materially changes where a trained potential fails. On the energy-volume scan benchmark, the MP-ALOE-trained MACE model fails the physicality check on 2.5% of structures versus 14.8% for the MatPES-trained model, and its MD completion rate at 0 to 100 GPa NPT rises from 83.7% to 90.6%. The combined MP-ALOE plus MatPES training set does best on every benchmark, with 0.8% energy-volume failures and 93.2% NPT completion, so the off-equilibrium and equilibrium data are complementary rather than redundant.

Load-bearing premise

The comparison between datasets rests on a single training run per model and a single set of benchmark trajectories, so the reported gaps could in principle be noise rather than a real difference.

Editorial extensions

If this is right

  • MP-ALOE-trained potentials reduce energy-volume scan failures from 14.8% to 2.5%, and combining with MatPES pushes that to 0.8%.
  • NPT molecular dynamics completion at 0 to 100 GPa rises from 83.7% with MatPES to 90.6% with MP-ALOE, and to 93.2% for the combined dataset.
  • At the 500 GPa version of the NPT test, MP-ALOE completes about twice as many scheduled timesteps as MatPES, 69.9% versus 34.1%.
  • All three models predict far-from-equilibrium forces without the systematic softening previously reported for universal potentials.
  • Because MP-ALOE and MatPES use compatible DFT settings, the two datasets can be merged into a single training set with better equilibrium and off-equilibrium behavior than either alone.

Reading between the lines

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

  • A direct ablation would be to remove the high-pressure frames from MP-ALOE and retrain; if NPT completion drops back to MatPES levels, pressure diversity is the active ingredient, not off-equilibrium sampling in general.
  • The committee-disagreement thresholds and the three-ionic-steps rule are heuristics; tuning them could make the dataset smaller without losing robustness, since only a small fraction of the 100 million candidates is labeled.
  • Because MP-ALOE mostly samples small bulk crystals, the paper's own limitations section notes that defect, surface, and amorphous environments would likely need fine-tuning; a testable extension is fine-tuning an MP-ALOE-trained potential on such bespoke data.
  • If seed-to-seed variability is small, the active-learning protocol itself becomes a reusable recipe for upgrading any future DFT functional to a universal potential.
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 introduces MP-ALOE, a publicly released dataset of approximately 910,000 r2SCAN DFT frames covering 89 elements, generated by elemental substitution into prototype structures, random position/lattice scrambling, and query-by-committee active learning with DIRECT downsampling. The authors train MACE potentials on MP-ALOE, on the existing MatPES r2SCAN dataset, and on their union, and benchmark these models on equilibrium relaxation (WBM-derived), far-from-equilibrium force prediction, energy-volume scans under extreme uniform strain, and NVT/NPT molecular dynamics stability using MLIP Arena tasks. The reported results show that the MP-ALOE-trained model is competitive with the MatPES-trained model on equilibrium and force benchmarks and substantially better on EV-scan physicality and MD survival; the combined dataset performs best overall.

Significance. If the conclusions hold, MP-ALOE is a valuable public resource: it is the largest r2SCAN UMLIP dataset to date, with broader energy, force, and pressure distributions than MatPES, and it is constructed to be directly compatible with MatPES. The use of external test sets (WBM, RM24) avoids circularity, since the models are trained on independent DFT labels and evaluated on benchmark structures not used in training. The release of the dataset, the trained potentials, and supporting data through a DOI is a concrete reproducibility strength. However, the headline comparative claims about improved MD stability and PES physicality rest on single-run, single-seed training evaluations without error bars or significance tests; these claims need additional statistical support before they can be regarded as established.

major comments (4)
  1. [Results: Molecular Dynamics Stability; Methods: Model Training] The central comparative claim that the MP-ALOE-trained model 'demonstrates improved stability in MD runs and physicality of the PES' rests entirely on single-run evaluations. In Methods ('Model Training'), each potential is trained once with one 90/5/5 split and no random seeds are reported, so the EV-scan failure rates in Table I (2.5% vs 14.8%), the NPT completion rates in Fig. 5b (90.6% vs 83.7%), and the 500 GPa rates in Fig. S5 (69.9% vs 34.1%) are point estimates with no uncertainty. Because these are discrete counts (e.g., ~25 vs ~148 failures out of 1000 structures; 90.6 vs 83.7 surviving runs out of 100), the differences could lie within seed-to-seed or split-to-split variability. I request repeated training with at least three to five seeds per dataset, reporting mean and standard error (or confidence intervals) for each metric, and ideally a bootstrap or permutation test over structures/trajectories. Without this, the word 'significantly' in the Discussion is not supported.
  2. [Methods: Density Functional Theory Calculation Details] The workflow converged for 82% of structures, meaning 18% of the QBC-selected candidates are omitted from the dataset. The paper does not analyze whether the non-converged structures are systematically different from the converged ones. Since the dataset is explicitly intended to cover far-from-equilibrium and high-pressure regions, a systematic failure to converge on the most extreme inputs would bias the dataset away from precisely the regimes claimed to be improved. I ask for a comparison of the pre-DFT descriptors (e.g., scrambling magnitude, lattice strain, QBC disagreement values, predicted-volume ratio) between converged and non-converged structures, together with a discussion of how any differences affect the dataset's coverage claims.
  3. [Benchmarking / Molecular Dynamics Stability] The comparison between MP-ALOE and MatPES is confounded by dataset size and generation protocol: MP-ALOE contains 909,792 frames versus 387,897 for MatPES (Table S1), and the two datasets were constructed by different sampling pipelines. The improved EV-scan and MD-stability metrics could therefore reflect the larger training set rather than the off-equilibrium sampling strategy per se. To support the attribution that off-equilibrium sampling improves robustness, the authors should either train on a matched-size random subset of MP-ALOE (e.g., ~388k frames) and on a comparable-size MatPES sample, or present an ablation that varies data volume while holding sampling protocol fixed. At minimum, the confounding effect of size should be acknowledged.
  4. [Methods: Query By Committee] The first active-learning iteration selects structures using an ensemble of PBE-trained models (MACE-MP-0, CHGNet, M3GNet), while the dataset labels are r2SCAN. The disagreement thresholds (100 meV/atom, 100 meV/Å, 100 meV/Å3) were chosen to approximate MACE-MP-0's errors on its own test set, which is a PBE test set. Since PBE and r2SCAN uncertainties can differ, the committee may systematically miss regions where r2SCAN is uncertain but PBE is confident, or vice versa. I am not claiming circularity, because the DFT labels are independent, but the sampling bias is a correctness risk for the active-learning pipeline. Please quantify the overlap between structures selected by the PBE committee and those selected by later r2SCAN-trained committees (e.g., fraction of re-selected structures per iteration), or otherwise demonstrate that the first-iteration selection does not dominate the final dataset's composition.
minor comments (6)
  1. [Abstract; The Dataset] The abstract states 'nearly 1 million DFT calculations', but the text in 'The Dataset' says '909,792 frames of DFT data (from 303,264 structure relaxations)'; since each relaxation contributes three ionic steps, please define 'frame' as one ionic step so the count is not conflated with the number of separate DFT relaxations.
  2. [Fig. 2a] The horizontal axis label in Fig. 2a appears to read '-8 4 0' instead of '-8 -4 0'; please check the sign of the middle tick.
  3. [Discussion] The phrases 'MP-ALOE significantly outperforms MatPES' and 'clearly achieves the best performance' are not supported by the reported statistics, which contain no error bars or significance tests; please soften these claims or add the statistical support requested in the major comments.
  4. [Methods: Query By Committee] The 'heuristic factor' used to increase the threshold for noble gases and f-block elements is not specified; please state the factor values or at least the range used, since this is a tunable parameter that affects the final dataset composition.
  5. [Results: Molecular Dynamics Stability] The sentence 'Note that we modify the original NPT benchmark from MLIP Arena to reduce the maximum pressure to 100 GPa' appears only in the Results; this modification should be stated in the Benchmarks section so readers know the protocol before seeing the results.
  6. [Table I] The 'Failures percentage' column is a binomial proportion and would benefit from a simple confidence interval (e.g., Wilson interval), which would make the 2.5% vs 14.8% comparison more interpretable even before re-training with multiple seeds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MP-ALOE dataset is benchmarked on external r2SCAN data and independent MLIP Arena tasks; the central claims are empirical comparisons, not derivations from fitted inputs.

full rationale

The paper's central claims are that MP-ALOE, an r2SCAN dataset of off-equilibrium structures, improves the robustness of universal MLIPs, with evidence from equilibrium relaxation benchmarks, off-equilibrium force predictions, EV-scan physicality tests, and MD stability runs. None of these claims reduce to the dataset's construction by definition. The QBC active-learning step uses PBE-trained models (MACE-MP-0, CHGNet, M3GNet) to select candidate structures, but the DFT labels themselves are computed independently with r2SCAN, and the benchmarks use external structures from WBM and MLIP Arena, so the selection bias is a sampling property, not a circular argument. The benchmark protocols are inherited from MatPES (Ref. [27], which shares an author with this paper) and MLIP Arena (Ref. [41]), but these are external, reproducible benchmark tasks with independent ground-truth DFT values; the self-citation is not load-bearing because the target results are not asserted by the cited work but measured here against external data. The only substantive weakness is that each model is trained once with a single 90/5/5 split and the headline gaps (EV failures 2.5% vs 14.8%; NPT completion 90.6% vs 83.7%) are point estimates without error bars or significance tests. That is a statistical robustness concern, not a circularity concern: it does not make the prediction equivalent to a fitted parameter or to the input data by construction. No equation in the paper defines a predicted quantity in terms of the benchmark target, and no fitted parameter is renamed as a prediction. Accordingly, the paper is self-contained against external benchmarks and warrants a circularity score of 0.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The dataset generation depends on a set of hand-set parameters (scrambling magnitudes, QBC thresholds, number of ionic steps) and domain assumptions about r2SCAN fidelity and the representativeness of prototype-derived structures. No new physical entities are introduced.

free parameters (8)
  • Atomic position scrambling standard deviation = 2% of lattice vector length
    Chosen so that median force in the dataset is approximately 1 eV/Å; selected by hand (SI 'Scrambling Structures').
  • Lattice strain standard deviation = 2% across six strain modes
    Chosen to increase diversity of stresses; no principled justification (SI 'Scrambling Structures').
  • QBC disagreement thresholds = 100 meV/atom, 100 meV/Å, 100 meV/Å3
    Set to roughly correspond to MACE-MP-0 test errors; heuristic (Methods, Query By Committee).
  • Heuristic factor for noble gas and f-block thresholds = not specified
    Applied to avoid over-representation of noble gases, technetium, and f-block elements; value not given (Methods, Query By Committee).
  • Number of r2SCAN ionic steps = 3
    Chosen because 90% of test calculations included both positive and negative pressure in at least one ionic step; heuristic (Methods, DFT Calculation Details).
  • Downsampling target per active learning cycle = about 125,000 from about 500,000
    Fixed number selected per BIRCH cluster; no sensitivity analysis (Methods, Downsampling).
  • Model training epochs = 100
    Empirically selected based on when learning appeared to plateau (Methods, Model Training).
  • Lattice parameter estimation radii = pymatgen atomic_radius with minor modifications
    Selected after testing several radius sets for closeness to DFT-relaxed geometries (SI, Guessing Lattice Parameters).
assumptions (5)
  • domain assumption r2SCAN is an adequate reference level of theory for training universal MLIPs
    Central premise of dataset value; the paper cites evidence that r2SCAN improves over PBE for various material classes (Introduction).
  • domain assumption PBE static then r2SCAN relaxation workflow yields valid r2SCAN energies and forces
    The WAVECAR from a PBE static calculation initializes the r2SCAN relaxation; assumes this accelerates convergence without biasing the r2SCAN result (Methods, DFT Calculation Details).
  • ad hoc to paper Three ionic steps of r2SCAN relaxation produce physically meaningful off-equilibrium configurations
    The dataset is defined by these partial relaxations; no test of convergence of the PES labels is provided (Methods, DFT Calculation Details).
  • domain assumption The WBM and RM24 benchmark structures are representative of intended UMLIP applications
    All benchmark claims are based on these external sets; representativeness is asserted but not systematically tested (Benchmarking sections).
  • ad hoc to paper The omitted 18% non-converged structures do not bias the dataset
    Structures that failed to converge are excluded without characterizing their chemistry or configurational similarity to the kept ones (Methods, DFT Calculation Details).

how reviews work

0 comments
Cite this review

Pith. "Pith review of MP-ALOE: An r2SCAN dataset for universal machine learning interatomic potentials." pith.science (2026). https://pith.science/paper/U76Q2PQE

@misc{pith2026250705559,
  author       = {Pith},
  title        = {Pith review of: MP-ALOE: An r2SCAN dataset for universal machine learning interatomic potentials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U76Q2PQE}},
  note         = {Machine review of arXiv:2507.05559}
}
read the original abstract

We present MP-ALOE, a dataset of nearly 1 million DFT calculations using the accurate r2SCAN meta-generalized gradient approximation. Covering 89 elements, MP-ALOE was created using active learning and primarily consists of off-equilibrium structures. We benchmark a machine learning interatomic potential trained on MP-ALOE, and evaluate its performance on a series of benchmarks, including predicting the thermochemical properties of equilibrium structures; predicting forces of far-from-equilibrium structures; maintaining physical soundness under static extreme deformations; and molecular dynamic stability under extreme temperatures and pressures. MP-ALOE shows strong performance on all of these benchmarks, and is made public for the broader community to utilize.

Figures

Figures reproduced from arXiv: 2507.05559 by the authors.

Figure 1
Figure 1. FIG. 1: Overview of the MP-ALOE dataset [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Describing the MP-ALOE dataset [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Benchmarking performance of UMLIPs at equilibrium [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Predictions of off-equilibrium forces (eV/Å) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5: Stability during molecular dynamics simulations [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in 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. VASP Plugins: Linking the Vienna ab-initio Simulation Package with Python

    cond-mat.mtrl-sci 2026-07 accept novelty 5.5 of 10

    A C++/pybind11 shared-memory plugin layer exposes VASP SCF and ionic data as NumPy arrays so Python can modify structure, forces, local potential, and occupancies in place.

Reference graph

Works this paper leans on

63 extracted references · 50 canonical work pages · cited by 1 Pith paper

  1. [1]

    & Sham, L

    Kohn, W. & Sham, L. J. Self-consistent equations including exchange and correlation effects. Phys. Rev.140, A1133–A1138 (1965)

    work page 1965

  2. [2]

    D., Levy, M

    Kaplan, A. D., Levy, M. & Perdew, J. P. The predictive power of exact constraints and appropriate norms in density functional theory.Annu. Rev. Phys. Chem.74, 193–218 (2023)

    work page 2023

  3. [3]

    & Blaha, P

    Tran, F., Stelzl, J. & Blaha, P. Rungs 1 to 4 of DFT jacob’s ladder: Extensive test on the lattice constant, bulk modulus, and cohesive energy of solids. J. Chem. Phys. 144, 204120 (2016)

    work page 2016

  4. [4]

    Blöchl, P. E. Projector augmented-wave method.Phys. Rev. B50, 17953–17979 (1994)

    1994

  5. [5]

    Daw, M. S. & Baskes, M. I. Embedded-atom method: Derivation and application to impurities, surfaces, and other defects in metals.Phys. Rev. B29, 6443–6453 (1984)

    work page 1984

  6. [6]

    Baskes, M. I. Modified embedded-atom potentials for cubic materials and impurities.Phys. Rev. B46, 2727–2742 (1992)

    work page 1992

  7. [7]

    & Parrinello, M

    Behler, J. & Parrinello, M. Generalized neural-network representation of high-dimensional potential-energy surfaces. Phys. Rev. Lett.98, 146401 (2007)

    work page 2007

  8. [8]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry (2017). Preprint at http://arxiv.org/abs/1704.01212, 1704.01212

    arXiv 2017

Show all 63 references
  1. [9]

    Schütt, K. T. et al. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions (2017). Preprint at http://arxiv.org/abs/1706.08566, 1706.08566

    2017 arXiv

  2. [10]

    E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials

    Batzner, S.et al. E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials. Nat. Commun.13, 2453 (2022)

    2022

  3. [11]

    Atomic cluster expansion for accurate and transferable interatomic potentials.Phys

    Drautz, R. Atomic cluster expansion for accurate and transferable interatomic potentials.Phys. Rev. B99, 014104 (2019)

    2019

  4. [12]

    P., Simm, G., Ortner, C

    Batatia, I., Kovacs, D. P., Simm, G., Ortner, C. & Csanyi, G. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields.Adv. Neural Inf. Process. Syst. 35, 11423–11436 (2022)

    2022

  5. [13]

    & Ong, S

    Chen, C. & Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nat. Comput. Sci.2, 718–728 (2022). 19

    2022

  6. [14]

    Deng, B.et al.CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nat. Mach. Intell.5, 1031–1041 (2023)

    2023

  7. [15]

    Batatia, I. et al. A foundation model for atomistic materials chemistry (2023). Preprint at http://arxiv.org/abs/2401.00096, 2401.00096

    2023 arXiv

  8. [16]

    & Han, S

    Park, Y., Kim, J., Hwang, S. & Han, S. Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations.J. Chem. Theory Comput.20, 4857– 4868 (2024)

    2024

  9. [17]

    Preprint at http://arxiv.org/abs/2405.04967, 2405.04967

    Yang, H.et al.MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures. Preprint at http://arxiv.org/abs/2405.04967, 2405.04967

    arXiv

  10. [18]

    Barroso-Luque, L. et al. Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models (2024). Preprint at http://arxiv.org/abs/2410.12771, 2410.12771

    2024 arXiv

  11. [19]

    Systematic softening in universal machine learning interatomic potentials.npj Comput

    Deng, B.et al. Systematic softening in universal machine learning interatomic potentials.npj Comput. Mater.11, 1–9 (2025)

    2025

  12. [20]

    Riebesell, J. et al. Matbench Discovery – A framework to evaluate machine learning crystal stability predictions (2024). Preprint at http://arxiv.org/abs/2308.14920, 2308.14920

    2024 arXiv

  13. [21]

    P., Burke, K

    Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett.77, 3865–3868 (1996)

    1996

  14. [22]

    Jain, A. et al. Commentary: The materials project: A materials genome approach to acceler- ating materials innovation.APL Mater.1, 011002 (2013)

    2013

  15. [23]

    K.et al.Accelerated data-driven materials science with the materials project.Nat

    Horton, M. K.et al.Accelerated data-driven materials science with the materials project.Nat. Mater. 1–11 (2025)

    2025

  16. [24]

    & Marques, M

    Schmidt, J., Pettersson, L., Verdozzi, C., Botti, S. & Marques, M. A. L. Crystal graph attention networks for the prediction of stable materials.Sci. Adv.7, eabi7948 (2021)

    2021

  17. [25]

    Schmidt, J.et al.Machine-learning-assisted determination of the global zero-temperature phase diagram of materials.Advanced Materials35, 2210788 (2023)

    2023

  18. [26]

    & Neugebauer, J

    Nazarov, R., Hickel, T. & Neugebauer, J. Vacancy formation energies in fcc metals: Influence of exchange-correlation functionals and correction schemes.Physical Review B85 (2012)

    2012

  19. [27]

    Kaplan, A. D. et al. A Foundational Potential Energy Surface Dataset for Materials (2025). Preprint at http://arxiv.org/abs/2503.04070, 2503.04070. 20

    2025

  20. [28]

    W., Kaplan, A

    Furness, J. W., Kaplan, A. D., Ning, J., Perdew, J. P. & Sun, J. Accurate and numerically efficient r2SCAN meta-generalized gradient approximation.J. Phys. Chem. Lett.11, 8208–8215 (2020)

    2020

  21. [29]

    Performance comparison of r2SCAN and SCAN metaGGA density func- tionals for solid materials via an automated, high-throughput computational workflow.Phys

    Kingsbury, R.et al. Performance comparison of r2SCAN and SCAN metaGGA density func- tionals for solid materials via an automated, high-throughput computational workflow.Phys. Rev. Mater.6, 013801 (2022)

    2022

  22. [30]

    Testing the r2SCAN density functional for the thermodynamic stability of solids with and without a van der waals correction.ACS Mater

    Kothakonda, M.et al. Testing the r2SCAN density functional for the thermodynamic stability of solids with and without a van der waals correction.ACS Mater. Au3, 102–111 (2023)

    2023

  23. [31]

    & Sai Gautam, G

    Swathilakshmi, S., Devi, R. & Sai Gautam, G. Performance of the r 2SCAN functional in transition metal oxides.J. Chem. Theory Comput.19, 4202–4215 (2023)

    2023

  24. [32]

    I., He, J

    Liu, M., Gopakumar, A., Hegde, V. I., He, J. & Wolverton, C. High-throughput hybrid- functional dft calculations of bandgaps and formation energies and multifidelity learning with uncertainty quantification.Phys. Rev. Mater.8, 043803 (2024)

    2024

  25. [33]

    W., Wood, B

    Qi, J., Ko, T. W., Wood, B. C., Pham, T. A. & Ong, S. P. Robust training of machine learning interatomic potentials with dimensionality reduction and stratified sampling. npj Comput. Mater. 10, 1–11 (2024)

    2024

  26. [34]

    & Drautz, R

    Lysogorskiy, Y., Bochkarev, A., Mrovec, M. & Drautz, R. Active learning strategies for atomic cluster expansion models.Phys. Rev. Mater.7, 043801 (2023)

    2023

  27. [35]

    S., Opper, M

    Seung, H. S., Opper, M. & Sompolinsky, H. Query by committee. In Proceedings of the fifth annual workshop on Computational learning theory, COLT ’92, 287–294 (Association for Computing Machinery, 1992)

    1992

  28. [36]

    & Rehme, S

    Zagorac, D., Müller, H., Ruehl, S., Zagorac, J. & Rehme, S. Recent developments in the inorganic crystal structure database: theoretical crystal structure data and related features.J. Appl. Crystallogr.52, 918–925 (2019)

    2019

  29. [37]

    & Baird, S

    Riebesell, J., Yang, H., Goodall, R. & Baird, S. G. Pymatviz: visualization toolkit for materials informatics (2022). URL https://github.com/janosh/pymatviz

    2022

  30. [38]

    & Marques, M

    Wang, H.-C., Botti, S. & Marques, M. A. L. Predicting stable crystalline compounds using chemical similarity.npj Comput. Mater.7, 1–9 (2021)

    2021

  31. [39]

    Zimmermann, N. E. R. & Jain, A. Local structure order parameters and site fingerprints for quantification of coordination environment and crystal structure similarity.RSC Adv.10, 21 6063–6081 (2020)

    2020

  32. [40]

    Lindemann, V. F. A. The calculation of molecular eigen-frequencies. Phys. Z. 11, 609–614 (1910)

    1910

  33. [41]

    Chiang, Y. et al. Mlip arena: Advancing fairness and transparency in machine learning interatomic potentials through an open and accessible benchmark platform (2025). URL https://openreview.net/forum?id=ysKfIavYQE

    2025

  34. [42]

    B., Perdew, J

    Alchagirov, A. B., Perdew, J. P., Boettger, J. C., Albers, R. C. & Fiolhais, C. Energy and pressure versus volume: Equations of state motivated by the stabilized jellium model.Phys. Rev. B63, 224115 (2001)

    2001

  35. [43]

    D., Clark, S

    Kaplan, A. D., Clark, S. J., Burke, K. & Perdew, J. P. Calculation and interpretation of classical turning surfaces in solids.npj Computational Materials7 (2021). See especially Figs. 6 and 7, as well as the Supplementary Tables

    2021

  36. [44]

    & Persson, K

    Grimvall, G., Magyari-Köpe, B., Ozolin,š, V. & Persson, K. A. Lattice instabilities in metallic elements. Rev. Mod. Phys.84, 945–986 (2012)

    2012

  37. [45]

    Martínez, L., Andrade, R., Birgin, E. G. & Martínez, J. M. PACKMOL: A package for building initial configurations for molecular dynamics simulations. J. Comput. Chem. 30, 2157–2164 (2009)

    2009

  38. [46]

    & Littmark, U

    Ziegler, J., Biersack, J. & Littmark, U. Empirical stopping powers for ions in solids 88–100 (1983)

    1983

  39. [47]

    Ziegler, J. F. & Biersack, J. P. The stopping and range of ions in matter. In Bromley, D. A. (ed.) Treatise on Heavy-Ion Science: Volume 6: Astrophysics, Chemistry, and Condensed Matter, 93–129 (Springer US, 1985)

    1985

  40. [48]

    P.et al.Python materials genomics (pymatgen): A robust, open-source python library for materials analysis.Comput

    Ong, S. P.et al.Python materials genomics (pymatgen): A robust, open-source python library for materials analysis.Comput. Mater. Sci.68, 314–319 (2013)

    2013

  41. [49]

    S., Nebgen, B., Lubbers, N., Isayev, O

    Smith, J. S., Nebgen, B., Lubbers, N., Isayev, O. & Roitberg, A. E. Less is more: Sampling chemical space with active learning.J. Chem. Phys.148, 241733 (2018)

    2018

  42. [50]

    A survey of deep active learning.ACM Comput

    Ren, P.et al. A survey of deep active learning.ACM Comput. Surv.54, 1–40 (2022)

    2022

  43. [51]

    W.et al.Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry (2025)

    Ko, T. W.et al.Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry (2025). Preprint at http://arxiv.org/abs/2503.03837. 22

    2025 arXiv

  44. [52]

    & Livny, M

    Zhang, T., Ramakrishnan, R. & Livny, M. BIRCH: an efficient data clustering method for very large databases. SIGMOD Rec.25, 103–114 (1996)

    1996

  45. [53]

    & Joubert, D

    Kresse, G. & Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B59, 1758–1775 (1999)

    1999

  46. [54]

    & Furthmüller, J

    Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set.Phys. Rev. B54, 11169–11186 (1996)

    1996

  47. [55]

    Ganose, A. et al. Atomate2: Modular workflows for materials science (2025). Preprint at https://chemrxiv.org/engage/chemrxiv/article-details/678e76a16dde43c9085c75e9

    2025

  48. [56]

    Jain, A. et al. FireWorks: a dynamic workflow system designed for high-throughput applica- tions. Concurr. Comput.27, 5037–5059 (2015)

    2015

  49. [57]

    Rosen, A. S.et al. Jobflow: Computational workflows made simple.J. Open Source Softw.9, 5995 (2024)

    2024

  50. [58]

    & Razinkovas, L

    Maciaszek, M., Žalandauskas, V., Silkinis, R., Alkauskas, A. & Razinkovas, L. The application of the scan density functional to color centers in diamond.The Journal of Chemical Physics 159 (2023)

    2023

  51. [59]

    Goerigk, L. et al. A look at the density functional theory zoo with the advanced gmtkn55 database for general main group thermochemistry, kinetics and noncovalent interactions.Phys- ical Chemistry Chemical Physics19, 32184–32215 (2017)

    2017

  52. [60]

    Bosoni, E. et al. How to verify the precision of density-functional-theory implementations via reproducible and universal workflows.Nature Reviews Physics6, 45–58 (2023)

    2023

  53. [61]

    Lane, C. et al. First-principles calculation of spin and orbital contributions to magnetically ordered moments insr2iro4. Phys. Rev. B101, 155110 (2020)

    2020

  54. [62]

    & Ong, S

    Chu, I.-H., Roychowdhury, S., Han, D., Jain, A. & Ong, S. P. Predicting the volumes of crystals. Comput. Mater. Sci.146, 184–192 (2018)

    2018

  55. [63]

    & Kresse, G

    Furthmüller, J., Käckell, P., Bechstedt, F. & Kresse, G. Extreme softening of vanderbilt pseu- dopotentials: General rules and case studies of first-row and d-electron elements.Phys. Rev. B 61, 4576–4587 (2000). 23 Supplemental Information: MP-ALOE: An r 2SCAN dataset for univ...

    2000

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.