Pith. sign in

REVIEW 4 major objections 4 minor 68 references

NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an automated pipeline, NepTrain with NepTrainKit, produces an accurate neuroevolution potential for CsPbI3 from just 200 perturbed configurations, bypassing the usual manual dataset screening.

desk verdict Useful NEP toolkit, but the '200 structures' claim is not actually established because the final training-set size is unreported. read the letter →

arxiv 2506.01868 v1 pith:TEKTDKSL submitted 2025-06-02 cs.LG cond-mat.mtrl-sci

Chengbing Chen , Yutong Li , Rui Zhao , Zhoulin Liu , Zheyong Fan , Gang Tang , Zhiyong Wang This is my paper
classification cs.LGcond-mat.mtrl-sci
keywords neuroevolutionpotentialsactivelearninginteratomictrainingdatasetgenerationfarthest-pointsamplingbondlengthfilteringCsPbI3moleculardynamics
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

The paper introduces NepTrain and NepTrainKit, an open-source Python toolkit and a companion GUI for building training datasets and training neuroevolution potentials (NEPs). It argues that the main bottleneck for NEP use, time-consuming and manual construction and screening of training data, can be automated. Using CsPbI3 as a demonstration, it shows that a NEP model trained automatically from only 200 perturbed configurations, plus an active-learning loop that selects new structures from its own simulations, reaches energy, force, and virial errors comparable to a published NEP built from 510 structures. It also reproduces the expected phase transitions of CsPbI3 within tens of kelvin. The authors intend the toolkit to make NEP workflows accessible beyond specialist groups, and state the workflow should transfer to other material systems.

What carries the argument

The load-bearing machinery is the automated active-learning loop, in which farthest-point sampling in descriptor space selects diverse structures from a large pool, a bond-length criterion rejects non-physical geometries, and each iteration retrains the NEP on a growing set of DFT-labeled configurations. Farthest-point sampling works by repeatedly adding the unselected point whose shortest distance to the already-selected set is largest, thereby keeping the training set spread across descriptor space. The bond-length filter flags any atomic pair shorter than a coefficient times the sum of the two covalent radii, and deletes the whole structure containing it.

What would settle it

Label energies and forces for a held-out set of CsPbI3 configurations generated independently of the tested NEP, for example from DFT molecular dynamics or from experimental structures, and compare RMSEs; if errors are substantially larger than the reported 1.9 meV/atom energy and 46.0 meV/Å force errors, the acceptable-accuracy claim depends on self-sampling.

Watch

Extended reading notes

Core claim

The central claim is that data preparation, not model training, is the real obstacle to making neuroevolution potentials accurate, and that NepTrain removes that obstacle. The package combines the pieces a practitioner would otherwise assemble by hand: it perturbs initial crystal structures, filters out physically impossible geometries with a bond-length criterion, selects a small diverse subset by farthest-point sampling in descriptor space, trains a NEP, runs molecular dynamics with the resulting model, labels newly visited configurations with DFT single-point calculations, and iterates. The CsPbI3 case study shows that after this automated loop, a model trained on 200 configurations has RMSEs below 1.9 meV/atom for energy, 46.0 meV/Å for force, and 17.0 meV/atom for virial, with test errors roughly half the training errors, and reproduces the orthorhombic to tetragonal to cubic phase transitions at about 309 K and 404 K, close to the reference values of 297 K and 438 K.

Load-bearing premise

In Section 5.2, the validation test set of 100 structures is drawn from molecular dynamics trajectories generated by the very NEP model under test, and the claim assumes that these structures represent the model's true error, which is optimistic for regions the model never visits.

Editorial extensions

If this is right

  • NEP practitioners can replace manual dataset curation with a command-line driven active-learning loop, lowering the barrier for smaller research groups.
  • The 200-structure model reaches RMSEs comparable to a published 510-structure model, indicating that a small, well-selected training set can be competitive.
  • Automated error plots, descriptor projections, and selection records generated at each iteration make the training-data construction reproducible and auditable.
  • Because datasets are stored in standard extended XYZ form with descriptors, the pipeline's outputs can be repurposed to train other machine-learned interatomic potentials, not just NEP.

Reading between the lines

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

  • If the reported accuracy depends on the self-sampling of the test set, a fairer benchmark would label configurations from independent DFT-based MD; this would likely raise the measured errors for unseen regions of configuration space.
  • The demonstration on CsPbI3, a material with simple chemistry and well-known phases, may not transfer to chemically heterogeneous systems, where 200 configurations could be insufficient; the automated loop would then simply need more iterations.
  • The pre-trained 89-element model embedded in the GUI could let users pre-screen structures and estimate properties before any DFT runs, turning the toolkit into a general-purpose exploratory materials tool.
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 / 4 minor

Summary. The paper presents NepTrain, a Python package for automated construction and iterative active-learning refinement of training datasets for neuroevolution potentials (NEP), and NepTrainKit, a GUI for visualizing and editing such datasets. The toolkit combines cell/atomic perturbation, bond-length filtering, farthest-point sampling, automated NEP training, GPUMD MD sampling, and VASP single-point calculations. The central case study is CsPbI3: starting from 200 perturbed structures selected from a larger perturbation pool, the authors run an automated active-learning loop (MD at 0–620 K with stages up to 5 ns), train NEP models, and validate them by reporting energy/force/virial RMSEs, equation-of-state curves, and phase-transition temperatures. The paper claims that this workflow yields accurate NEP models with minimal manual data preparation and a small training set.

Significance. If the central claims are substantiated, the software contribution is genuinely useful: NepTrain and NepTrainKit address a real bottleneck in NEP modeling, are open-source, and are accompanied by a concrete CsPbI3 case study with reproducible commands and a public repository. The reported RMSEs (energy <1.9 meV/atom, force <46.0 meV/Å, virial <17.0 meV/atom) and the EOS/phase-transition agreement are plausible and align with earlier NEP studies. However, the evidence for the headline 'only 200 configurations' advantage is currently incomplete, and the validation protocol has a self-selection bias that needs to be addressed before the claims can be accepted as stated.

major comments (4)
  1. [Sections 5.1 and 5.2] The final training-set size is never reported, and the text is internally inconsistent about it. Section 5.1 describes an active-learning loop in which the train module runs nep, gpumd, select, and vasp iteratively and 'the training set is updated based on the selected data,' with MD stages of 100 ps, 500 ps, 1 ns, and 5 ns. This necessarily adds structures to the initial 200-structure seed. Yet Section 5.2 states that 'the training set consists of only 200 representative structures' and the conclusion repeats that NepTrain generates 'training sets containing only 200 perturbed configurations.' The number of configurations added in each active-learning iteration is never given. Because the claimed small-data advantage and the comparison with the 510-structure literature model in Section 5.2 depend directly on the actual final dataset size, the authors must report the per-iteration training-set sizes and reconcile the wording. As written, the central low-data claim is unverifiable.
  2. [Section 5.2 (test-set validation)] The test set is not independent of the model under evaluation. The 100-structure test set is obtained 'by performing MD simulations using GPUMD' for orthorhombic CsPbI3 with a total simulation time of 10 ns, and GPUMD MD uses the NEP model being validated. Consequently, the sampled configurations are ones the model itself generates, and the reported RMSEs in Figure 4 do not probe regions of configuration space that the model does not visit. This can make both the RMSE values and the phase-transition temperatures in Figure 7 optimistic. The authors should add an independent test set, for example DFT-labeled configurations from MD runs driven by a different potential or from a broader set of thermal distortions, or otherwise quantify how much of the relevant configuration space is covered by the GPUMD-generated test set.
  3. [Section 5.2 (comparison with literature model)] The comparison with the 510-structure CsPbI3 model of Wiktor et al. is presented as evidence that 'only 200 perturbed configurations' suffice, but the comparison is confounded by at least two missing quantities: the final training-set size after active learning and the test protocol. Even setting the size issue aside, the literature model was trained with its own data selection and hyperparameters, so equal RMSEs do not by themselves establish that NepTrain's workflow achieves the same accuracy with fewer labeled configurations. A controlled comparison (same test set, same descriptor/hyperparameter settings, and documented final training sizes) is needed to support the claim.
  4. [Section 4.2, Eq. (2)] The 'Maximum Error Point Selection' criterion computes the sum of absolute differences between a 'NEP-predicted structure descriptor' and a 'DFT reference descriptor,' but the paper does not define how a DFT reference descriptor is obtained or why disagreement in descriptor space is the appropriate error measure for identifying anomalous structures. Since this tool is one of the four key data-editing functions of NepTrainKit, its definition should be clarified or the criterion should be justified.
minor comments (4)
  1. [Section 5.1] The sentence 'We selected 200 representative structures from the dataset with 0.001 Å as the maximum atomic perturbation distance' is confusing: in the select command, -d is presumably the minimum distance for farthest-point sampling, whereas Section 3.1 defines -d as the atomic position perturbation magnitude. Please use a distinct parameter name or clarify the meaning in this context.
  2. [Section 3.4] There is a typo: 'makingit applicable' should read 'making it applicable.'
  3. [Figures 3, 5, and 6] The figure captions and surrounding text contain sequences such as '/uni00000013/uni00000011/...' that are not valid image paths or meaningful text; this appears to be a rendering artifact and should be cleaned in the final version.
  4. [Section 3.5] The farthest-point-sampling pseudocode would be clearer if the descriptor space in which distances are computed were stated explicitly, especially since the select module can use either NEP_CPU or SOAP descriptors.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the validation rests on external DFT labels and independent reference benchmarks.

full rationale

The paper is a software toolkit description rather than a derivation chain. The NEP model is trained on DFT-labeled configurations and evaluated against DFT single-point energies, forces, and virials, as well as independent literature phase-transition temperatures and EOS curves from Fransson et al. [20] and Wiktor et al. [21]. No fitted parameter is renamed as a prediction, and no central claim reduces to a self-citation: the NEP/GPUMD citations are standard method attributions, not load-bearing uniqueness arguments. The active-learning loop uses the model to propose configurations that are then labeled by VASP/DFT, which is a standard data-generation strategy and not a circularity. One methodological weakness is that the 100-structure test set is sampled from GPUMD MD trajectories produced by the model under evaluation, so it may under-represent regions the model does not visit; however, the DFT labels remain external and the reported RMSEs are not forced by construction. Separately, the manuscript asserts that the final training set 'consists of only 200 representative structures' without reporting how many configurations were added during active learning; this is a missing-support issue in the small-dataset claim, not a circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper is a software toolkit contribution, so the 'free parameters' are the user-adjustable knobs in the workflow. Each parameter affects the composition of the training set and hence the final model. The domain assumptions are the unvalidated heuristics that the toolkit's filtering and selection strategies actually improve model quality. No new physical entities are introduced.

free parameters (5)
  • Coeff (covalent radius coefficient) = not specified in paper
    Used in Eq. (3) to decide whether an atomic pair has non-physical bond length; its value changes the filtering outcome.
  • -d (maximum atomic perturbation distance) = 0.001 Å in example
    Controls how close selected structures are to existing ones during farthest-point sampling; affects diversity of the training set.
  • Perturbation magnitudes (-c and -d) = 4% strain, 0.3 Å displacement
    Set the spread of initial perturbed structures; directly determines initial dataset coverage.
  • Number of perturbed and selected structures = 5000 generated, 200 selected
    Dataset size and density; no systematic study of how these numbers affect final model accuracy.
  • FPS maximum retention number and minimum distance in NepTrainKit = user-specified
    Controls dataset sparsification; default values not reported in the paper.
assumptions (4)
  • domain assumption Bond-length filtering with the covalent-radii criterion removes non-physical structures without discarding valid ones.
    Invoked in Section 4.2 Eq. (3) and used in Section 5.1; no verification that the threshold does not exclude physical configurations.
  • domain assumption Farthest-point sampling in descriptor space yields representative, diverse training subsets.
    Used in Sections 3.5 and 4.2; the paper provides the algorithm but no evidence that it outperforms random or other sampling for NEP.
  • domain assumption The NEP descriptors and training framework from prior work are reliable enough that errors on sampled structures reflect model quality.
    Section 2 relies on NEP/GPUMD as given; no new theoretical support is provided.
  • domain assumption The test set generated by MD with the trained NEP model is representative of the model's accuracy.
    Section 5.2 uses GPUMD MD with the model to generate test structures; if the model has unexplored error regions, RMSEs are optimistic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials." pith.science (2026). https://pith.science/paper/TEKTDKSL

@misc{pith2026250601868,
  author       = {Pith},
  title        = {Pith review of: NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TEKTDKSL}},
  note         = {Machine review of arXiv:2506.01868}
}
abstract

As a machine-learned potential, the neuroevolution potential (NEP) method features exceptional computational efficiency and has been successfully applied in materials science. Constructing high-quality training datasets is crucial for developing accurate NEP models. However, the preparation and screening of NEP training datasets remain a bottleneck for broader applications due to their time-consuming, labor-intensive, and resource-intensive nature. In this work, we have developed NepTrain and NepTrainKit, which are dedicated to initializing and managing training datasets to generate high-quality training sets while automating NEP model training. NepTrain is an open-source Python package that features a bond length filtering method to effectively identify and remove non-physical structures from molecular dynamics trajectories, thereby ensuring high-quality training datasets. NepTrainKit is a graphical user interface (GUI) software designed specifically for NEP training datasets, providing functionalities for data editing, visualization, and interactive exploration. It integrates key features such as outlier identification, farthest-point sampling, non-physical structure detection, and configuration type selection. The combination of these tools enables users to process datasets more efficiently and conveniently. Using $\rm CsPbI_3$ as a case study, we demonstrate the complete workflow for training NEP models with NepTrain and further validate the models through materials property predictions. We believe this toolkit will greatly benefit researchers working with machine learning interatomic potentials.

Figures

Figures reproduced from arXiv: 2506.01868 by the authors.

Figure 1
Figure 1. FIG. 1. Workflow of the automated training process in NepTrain. Starting from the prepared [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Interface of NepTrainKit, illustrating the three primary sections: Dataset Manipulation(1. Maximum error point [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. The projected schematic of structural descriptors for [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: FIG. 5. EOS curves for [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Schematic illustration of the evolution of the [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Lattice parameters of [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 52 canonical work pages

  1. [1]

    This advancement is driving the shift incomputationalmaterialssciencefromempiricalmodels to a data-driven paradigm

    Introduction The rapid development of machine-learned potentials (MLPs), or machine-learned force fields [1], has provided a revolutionary solution for molecular dynamics (MD) simulations, offering both the accuracy of density func- tionaltheory(DFT)calculationsandtheefficiencyofem- pirical potential. This advancement is driving the shift incomputationalm...

    arXiv 2025

  2. [2]

    Like many other neural network po- tentials [2], NEP is a many-body potential

    The NEP method and the GPUMD package The NEP method [12?, 13] is a neural network po- tential which is trained using the separable natural evo- lution strategy [38]. Like many other neural network po- tentials [2], NEP is a many-body potential. However, it is still a localized potential with well defined site en- ergyU i for each atomiin an extended syste...

    work page

  3. [3]

    It utilizes the Python package ASE [41] for read- ing and writing lattice structures

    NepTrain NepTrain invokes tools such as NEP, GPUMD and VASP to facilitate the automated training of NEP mod- els. It utilizes the Python package ASE [41] for read- ing and writing lattice structures. The software oper- ates using the formatNepTrain command argv, where commandspecifies the functional module, andargvpro- vides the corresponding parameters f...

    work page

  4. [4]

    Initialize the selected point setS={p 1}, wherep 1 is an arbitrary starting point

    work page

  5. [5]

    For each unselected pointpi, calculate its shortest distance to the selected point setS: d(pi, S) = min pj ∈S ∥pi −p j∥

    work page

  6. [6]

    Select the pointpk that maximizesd(p i, S): pk = arg max pi /∈S d(pi, S) 5

    work page

  7. [7]

    Addp k to the selected point setS

    work page

  8. [8]

    Meanwhile, the-for–-filterparameter allows users to enable or disable filtering based on the minimum bond length (disabled by default)

    Repeat steps 2-4 until the size ofSreaches the de- sired number of points. Meanwhile, the-for–-filterparameter allows users to enable or disable filtering based on the minimum bond length (disabled by default). This functionality helps to exclude non-physical structures from the MD trajectory. The underlying principle of this filtering method is de- taile...

    work page

Show all 68 references
  1. [9]

    NepTrainKit Prior to this work, researchers primarily used OVITO

  2. [10]

    Zhang, J

    L. Zhang, J. Han, H. Wang, R. Car, and W. E, Deep potential molecular dynamics: A scalable model with the accuracy of quantum mechanics, Physical Review Letters 120, 143001 (2018)

    2018

  3. [11]

    Examples In this section, we present the details of a fully au- tomated workflow for generating a NEP model, using CsPbI3 as a representative example and implementing the process through NepTrain. We demonstrate that this tool enables the construction of highly accurate NEP 8 ...

  4. [12]

    NepTrain inte- 10 FIG

    Summary and conclusions In this work, we introduced the software tools Nep- Train and NepTrainKit for the convenient preparation and manipulation of training datasets to construct high- quality training sets for NEP models. NepTrain inte- 10 FIG. 6. Schematic illustration of t...

  5. [13]

    O. T. Unke, S. Chmiela, H. E. Sauceda, M. Gastegger, I. Poltavsky, K. T. Schütt, A. Tkatchenko, and K.-R. Müller, Machine learning force fields, Chemical Reviews 11 121, 10142 (2021)

    2021

  6. [14]

    Behler and G

    J. Behler and G. Csányi, Machine learning potentials for extended systems: a perspective, European Physical Journal B94, 142 (2021)

    2021

  7. [15]

    Friederich, F

    P. Friederich, F. Häse, J. Proppe, and A. Aspuru-Guzik, Machine-learned potentials for next-generation matter simulations, Nature Materials20, 750 (2021)

    2021

  8. [16]

    Behler, Perspective: Machine learning potentials for atomistic simulations, Journal of Chemical Physics145, 170901 (2016)

    J. Behler, Perspective: Machine learning potentials for atomistic simulations, Journal of Chemical Physics145, 170901 (2016)

    2016

  9. [17]

    Behler and M

    J. Behler and M. Parrinello, Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces, Physical Review Letters98, 146401 (2007)

    2007

  10. [18]

    A. P. Bartók, M. C. Payne, R. Kondor, and G. Csányi, Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons, Physical Re- view Letters104, 136403 (2010)

    2010

  11. [19]

    Thompson, L

    A. Thompson, L. Swiler, C. Trott, S. Foiles, and G. Tucker, Spectral neighbor analysis method for auto- mated generation of quantum-accurate interatomic po- tentials, Journal of Computational Physics285, 316 (2015)

    2015

  12. [20]

    A. V. Shapeev, Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials, Mul- tiscale Modeling & Simulation14, 1153 (2016)

    2016

  13. [21]

    H. Wang, L. Zhang, J. Han, and W. E, Deepmd-kit: A deep learning package for many-body potential en- ergy representation and molecular dynamics, Computer Physics Communications228, 178 (2018)

    2018

  14. [22]

    C. Qian, D. Hedman, P. Li, S. Y. Kim, and F. Ding, The reconstruction of Pt(001) surface and the shell-like reconstruction of the vicinal Pt(001) surfaces revealed by neural network potential, Small20, 2404274 (2024)

    2024

  15. [23]

    Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Physical Review B 99, 014104 (2019)

    R. Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Physical Review B 99, 014104 (2019)

    2019

  16. [24]

    Z. Fan, Z. Zeng, C. Zhang, Y. Wang, K. Song, H. Dong, Y. Chen, and T. Ala-Nissila, Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport, Physical Review B104, 104309 (2021)

    2021

  17. [25]

    Z. Fan, Y. Wang, P. Ying, K. Song, J. Wang, Y. Wang, Z. Zeng, K. Xu, E. Lindgren, J. M. Rahm, A. J. Gabourie, J. Liu, H. Dong, J. Wu, Y. Chen, Z. Zhong, J. Sun, P. Erhart, Y. Su, and T. Ala-Nissila, GPUMD: A package for constructing accurate machine-learned po- tentials and pe...

    2022

  18. [26]

    Z. Fan, W. Chen, V. Vierimaa, and A. Harju, Efficient molecular dynamics simulations with many-body poten- tials on graphics processing units, Computer Physics Communications218, 10 (2017)

    2017

  19. [27]

    H. Dong, Y. Shi, P. Ying, K. Xu, T. Liang, Y. Wang, Z. Zeng, X. Wu, W. Zhou, S. Xiong, S. Chen, and Z. Fan, Molecular dynamics simulations of heat transport using machine-learned potentials: A mini-review and tutorial on gpumd with neuroevolution potentials, Journal of Ap- pli...

    2024

  20. [28]

    P. Ying, C. Qian, R. Zhao, Y. Wang, K. Xu, F. Ding, S. Chen, and Z. Fan, Advances in modeling complex ma- terials: The rise of neuroevolution potentials, Chemical Physics Reviews6, 011310 (2025)

    2025

  21. [29]

    Y. Wang, Z. Fan, P. Qian, M. A. Caro, and T. Ala- Nissila, Density dependence of thermal conductivity in nanoporous and amorphous carbon with machine- learned molecular dynamics (2024), arXiv:2408.12390 [cond-mat]

    2024 arXiv

  22. [30]

    K. Xu, T. Liang, N. Xu, P. Ying, S. Chen, N. Wei, J. Xu, and Z. Fan, NEP-MB-pol: A unified machine- learned framework for fast and accurate prediction of water’s thermodynamic and transport properties (2024), arXiv:2411.09631 [physics]

    2024 arXiv

  23. [31]

    S. Chen, X. Jin, W. Zhao, and T. Li, Intricate short- range order in GeSn alloys revealed by atomistic simula- tions with highly accurate and efficient machine-learning potentials, Physical Review Materials8, 043805 (2024)

    2024

  24. [32]

    Fransson, J

    E. Fransson, J. Wiktor, and P. Erhart, Phase transitions in inorganic halide perovskites from machine-learned po- tentials, The Journal of Physical Chemistry C127, 13773 (2023)

    2023

  25. [33]

    Wiktor, E

    J. Wiktor, E. Fransson, D. Kubicki, and P. Erhart, Quan- tifying dynamic tilting in halide perovskites: Chemical trends and local correlations, Chemistry of Materials35, 6737 (2023)

    2023

  26. [34]

    Batatia, D

    I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Csanyi, MACE: Higher order equivariant message passing neural networks for fast and accurate force fields, inAdvances in Neural Information Processing Systems, Vol. 35, edited by S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrav...

  27. [35]

    J. Liu, J. Byggmästar, Z. Fan, P. Qian, and Y. Su, Large- scale machine-learning molecular dynamics simulation of primary radiation damage in tungsten, Physical Review B108, 054312 (2023)

    2023

  28. [36]

    P. Ying, H. Dong, T. Liang, Z. Fan, Z. Zhong, and J. Zhang, Atomistic insights into the mechanical anisotropy and fragility of monolayer fullerene networks using quantum mechanical calculations and machine- learning molecular dynamics simulations, Extreme Me- chanics Letters58...

    2023

  29. [37]

    M. Yu, Z. Zhao, W. Guo, and Z. Zhang, Fracture tough- ness of two-dimensional materials dominated by edge en- ergy anisotropy, Journal of the Mechanics and Physics of Solids186, 105579 (2024)

    2024

  30. [38]

    K. Song, R. Zhao, J. Liu, Y. Wang, E. Lindgren, Y. Wang, S. Chen, K. Xu, T. Liang, P. Ying, N. Xu, Z. Zhao, J. Shi, J. Wang, S. Lyu, Z. Zeng, S. Liang, H. Dong, L. Sun, Y. Chen, Z. Zhang, W. Guo, P. Qian, J. Sun, P. Erhart, T. Ala-Nissila, Y. Su, and Z. Fan, General-purpose ma...

    2024

  31. [39]

    metatensor, metatrain,https://github.com/ metatensor/metatrain(2025)

    2025

  32. [40]

    M. R. Schäfer, N. Segreto, F. Zills, C. Holm, and J. Käst- ner, Apax: A flexible and performant framework for the development of machine-learned interatomic potentials (2025), arXiv:2505.22168 [physics]

    2025

  33. [41]

    Benner, equitrain,https://github.com/ BAMeScience/equitrain(2025)

    P. Benner, equitrain,https://github.com/ BAMeScience/equitrain(2025)

    2025

  34. [42]

    Brunken, O

    C. Brunken, O. Peltre, H. Chomet, L. Walewski, M. McAuliffe, V. Heyraud, S. Attias, M. Maarand, Y. Khanfir, E. Toledo, F. Falcioni, M. Bluntzer, S. Acosta-Gutiérrez, and J. Tilly, Machine learning in- teratomic potentials: Library for efficient training, model development and ...

    2025 arXiv

  35. [43]

    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, An automated framework for ex- 12 ploring and learning potential-energy surfaces (2024), arXiv:2412.16736 [physics]

    2024 arXiv

  36. [44]

    Zhang, H

    Y. Zhang, H. Wang, W. Chen, J. Zeng, L. Zhang, H. Wang, and W. E, DP-GEN: A concurrent learning platformforthegenerationofreliabledeeplearningbased potential energy models, Computer Physics Communica- tions253, 107206 (2020)

    2020

  37. [45]

    Galib, M

    M. Galib, M. Isiet, and M. Ponga, AtomProNet: Data flow to and from machine learning interatomic potentials in materials science (2025), arXiv:2501.14039 [cond-mat]

    2025 arXiv

  38. [46]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville, UMAP: Uniform manifold approximation and projection for dimension re- duction (2020), arXiv:1802.03426 [stat]

    2020 arXiv

  39. [47]

    How- ever, these two tools operate independently, lacking the capability for integrated analysis

    or VMD [48] to inspect dataset structures and em- ployed Python scripts to visualize NEP output. How- ever, these two tools operate independently, lacking the capability for integrated analysis. In contrast, Chemis- cope enables users to intuitively explore structural fea- tur...

  40. [48]

    Lindgren, M

    E. Lindgren, M. Rahm, E. Fransson, F. Eriksson, N. Österbacka, Z. Fan, and P. Erhart, calorine: A Python package for constructing and sampling neuroevolution potential models, Journal of Open Source Software9, 6264 (2024)

    2024

  41. [49]

    Kresse and J

    G. Kresse and J. Hafner,Ab Initiomolecular-dynamics simulation of the liquid-metal–amorphous-semiconductor transition in germanium, Physical Review B49, 14251 (1994)

    1994

  42. [50]

    Kresse and J

    G. Kresse and J. Furthmüller, Efficient iterative schemes forab initiototal-energy calculations using a plane-wave basis set, Physical Review B54, 11169 (1996)

    1996

  43. [51]

    T.Schaul, T.Glasmachers,andJ.Schmidhuber,HighDi- mensions and Heavy Tails for Natural Evolution Strate- gies,inProceedings of the 13th Annual Conference on Ge- netic and Evolutionary Computation, GECCO ’11 (Asso- ciation for Computing Machinery, New York, NY, USA,

  44. [52]

    Liang, K

    T. Liang, K. Xu, E. Lindgren, Z. Chen, R. Zhao, J. Liu, B. Tang, B. Zhang, Y. Wang, K. Song, P. Ying, H. Dong, S. Chen, P. Erhart, Z. Fan, T. Ala-Nissila, and J. Xu, Nep89: Universal neuroevolution potential for in- organic and organic materials across 89 elements (2025), arXi...

    2025 arXiv

  45. [53]

    Fan, NEP_CPU,https://github.com/ brucefan1983/NEP_CPU(2025)

    Z. Fan, NEP_CPU,https://github.com/ brucefan1983/NEP_CPU(2025)

    2025

  46. [54]

    Hjorth Larsen, J

    A. Hjorth Larsen, J. Jørgen Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P. C. Jennings, P. Bjerre Jensen, J. Kermode, J. R. Kitchin, E. Leonhard Kolsbjerg, J. Kubal, K. Kaasb- jerg, S. Lysgaard,...

    2017

  47. [55]

    M. O. J. Jäger, E. V. Morooka, F. Federici Canova, L. Hi- manen, and A. S. Foster, Machine learning hydrogen ad- sorption on nanoclusters through structural descriptors, npj Computational Materials4, 37 (2018)

    2018

  48. [56]

    A. P. Bartók, S. De, C. Poelking, N. Bernstein, J. R. Kermode, G. Csányi, and M. Ceriotti, Machine learning unifies the modeling of materials and molecules, Science Advances3, e1701816 (2017)

    2017

  49. [57]

    Himanen, M

    L. Himanen, M. O. Jäger, E. V. Morooka, F. Fed- erici Canova, Y. S. Ranawat, D. Z. Gao, P. Rinke, and A. S. Foster, DScribe: Library of descriptors for machine learning in materials science, Computer Physics Commu- nications247, 106949 (2020)

    2020

  50. [58]

    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.Cour- napeau, M. Brucher, M. Perrot, and É. Duchesnay, Scikit-learn: Machine Learning in Python, Journal of Machine Learning Re...

    2011

  51. [59]

    A. Stukowski, Visualization and analysis of atomistic simulationdatawithOVITO–theopenvisualizationtool, Modelling and Simulation in Materials Science and En- gineering18, 015012 (2010)

    2010

  52. [60]

    Humphrey, A

    W. Humphrey, A. Dalke, and K. Schulten, Vmd: Visual molecular dynamics, Journal of Molecular Graphics14, 33 (1996)

    1996

  53. [61]

    Fraux, R

    G. Fraux, R. K. Cersonsky, and M. Ceriotti, Chemis- cope: Interactive structure-property explorer for materi- als and molecules, Journal of Open Source Software5, 2117 (2020)

    2020

  54. [62]

    Campagnola, E

    L. Campagnola, E. Larson, A. Klein, D. Hoese, Sid- dharth, C. Rossant, A. Griffiths, N. P. Rougier, asnt, L. Gaifas, K. Mühlbauer, A. Taylor, MSS, T. Lambert, sylm21, A. Anderson, A. J. Champandard, M. Hunter, T. Robitaille, M. F. Kaptan, E. S. de Andrade, G. Bokota, G. Faveli...

    2025

  55. [63]

    D. Xu, M. Wu, Y. Bai, B. Wang, H. Zhou, Z. Fan, N. Zhang, J. Tan, H. Li, H. Bian, and Z. Liu, Record- efficiency inverted CsPbI3 perovskite solar cells enabled by rearrangement and hydrophilic modification of SAMs, Advanced Functional Materials35, 2412946 (2025)

    2025

  56. [64]

    Marronnier, G

    A. Marronnier, G. Roma, S. Boyer-Richard, L. Pedesseau, J.-M. Jancu, Y. Bonnassieux, C. Katan, C. C. Stoumpos, M. G. Kanatzidis, and J. Even, Anhar- monicity and disorder in the black phases of cesium lead iodide used for stable inorganic perovskite solar cells, ACS Nano12, 34...

    2018

  57. [65]

    Momma and F

    K. Momma and F. Izumi,VESTA 3for three- dimensional visualization of crystal, volumetric and mor- phology data, Journal of Applied Crystallography44, 1272 (2011)

    2011

  58. [66]

    P. E. Blöchl, Projector augmented-wave method, Physi- cal Review B50, 17953 (1994)

    1994

  59. [67]

    J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized Gradient Approximation Made Simple, Physical Review Letters77, 3865 (1996)

    1996

  60. [68]

    Bernetti and G

    M. Bernetti and G. Bussi, Pressure control using stochas- tic cell rescaling, Journal of Chemical Physics153, 114107 (2020)

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.