Pith. sign in

REVIEW 3 major objections 4 minor 51 references

Knowledge Distillation Framework for Accelerating High-Accuracy Neural Network-Based Molecular Dynamics Simulations

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A knowledge-distillation recipe that leaves a pre-trained universal neural network potential untuned as the teacher yields student networks that match experimental PEG and LGPS properties with 10x fewer DFT labels and up to 106x faster…

desk verdict A novel KD idea with a plausible mechanism, but the key comparison is missing a control that would isolate the off-the-shelf teacher as the decisive ingredient. read the letter →

arxiv 2506.15337 v2 pith:GBEK6UEA submitted 2025-06-18 cs.LG cond-mat.mtrl-sciphysics.comp-ph

Naoki Matsumura , Yuta Yoshimoto , Yuto Iwasaki , Meguru Yamazaki , Yasufumi Sakai This is my paper · ORCID
classification cs.LGcond-mat.mtrl-sciphysics.comp-ph
keywords knowledgedistillationneuralnetworkpotentialmoleculardynamicspretraineduniversalNNPenergysurfacehigh-energystructuresDFTlabelsDeepPot-SE
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

Knowledge distillation for neural network potentials usually starts by fine-tuning the pre-trained teacher on density functional theory (DFT) data, but this paper argues that step is counterproductive: fine-tuning sharpens energy barriers, so the teacher's own MD runs never visit the high-energy structures a stable student needs. The proposed alternative skips teacher fine-tuning entirely and uses an off-the-shelf universal NNP whose known tendency to underestimate energies softens the landscape, letting MD trajectories explore a wider range of configurations. A student trained on those soft targets and then fine-tuned on only 1,000 DFT-labeled structures reproduces the experimental density and self-diffusion coefficient of polyethylene glycol and the lithium-ion diffusion of LGPS. If the recipe holds, it turns slow general-purpose potentials into fast, material-specific ones with a 10x cut in expensive DFT labels and speedups up to 106x.

What carries the argument

The central object is the effective softness of the untuned teacher's potential energy surface (PES). The framework builds on the documented result [6] that universal NNPs systematically underestimate energies in high-energy regions, which makes their landscapes gentler than the DFT landscape and lowers the energy barriers encountered during MD. The recipe has four steps: (a) NNP-MD runs with the off-the-shelf teacher produce a pool of soft targets; (b) a student DeepPot-SE network is trained on those targets with mean-squared-error loss; (c) a structural feature-based screening step compresses the student's intermediate-layer features with densMAP, adds normalized energy as a third dimension, and selects 1,000 diverse structures; (d) the student's descriptor layers are frozen and only its fitting network is fine-tuned on the DFT-labeled hard targets. The screening supplies the data efficiency, and the frozen-descriptor fine-tuning preserves the broad exploration achieved in the soft-target phase.

What would settle it

Run the same two-stage procedure on a third material where the off-the-shelf teacher overestimates rather than underestimates energies; if the soft-target trajectory does not cover the configurations visited in a reference AIMD run, the 1,000 DFT-labeled points will miss part of the needed high-energy tail and the student will fail in production MD. Concretely, compare the DFT-relabeled energy distribution of the selected soft targets against the energy distribution of a reference AIMD trajectory at the same conditions: the method's premise is that the former spans the same range, so a two-sided gap test with, say, the 5th-to-95th percentile overlap would settle the claim.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the energy underestimation of a pretrained universal NNP—normally treated as a defect—is the feature that makes distillation work. Under the untuned teacher, MD trajectories cross the potential-energy landscape more readily, so the soft-target dataset contains configurations in high-energy regions. The student trained on those soft targets inherits the teacher's sampled landscape; fine-tuning only the student's fitting network, with descriptor layers frozen, on 1000 DFT-labeled points selected for structural and energetic diversity restores accuracy without erasing that explored landscape. The authors demonstrate the contrast explicitly: when the teacher is fine-tuned before distillation, the resulting soft targets are shifted to lower energies and the student, despite a low validation force error, produces a density of $0.015\ \mathrm{g/cm^3}$ and a self-diffusion coefficient of $6156.33 \times 10^{-6}\ \mathrm{cm^2/s}$, far from experiment. This failure of the conventional KD path is presented as the evidence that the teacher's softness, rather than distillation alone, is the load-bearing ingredient.

Load-bearing premise

The recipe rests on the assumption that the untuned teacher's energy underestimation makes it explore the same high-energy structures that matter for real dynamics, and that a 1,000-point DFT fine-tune is enough to correct the student's surface in those regions.

Editorial extensions

If this is right

  • A material-specific NNP can be produced from 1,000 DFT hard targets plus teacher-generated soft targets, a 10x reduction in DFT labels relative to the active-learning baseline for PEG.
  • The student models are fast enough for long production MD: up to 82x speedup for a 3,100-atom PEG system and 20x for a 1,600-atom LGPS system on an H100 GPU, with peak speedups of 106x and 46x respectively.
  • Because the fine-tuned-teacher path demonstrably fails in production MD, the result implies that diversity of sampled configurations matters more than raw teacher accuracy for downstream simulation stability.
  • The two-stage recipe eliminates the iterative retraining loop of active learning, cutting total NNP generation time by 1.9x for PEG and 3.0x for LGPS.

Reading between the lines

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

  • Beyond the paper, the same soft-teacher trick could be repurposed as an enhanced-sampling tool: an untuned universal NNP could cross barriers cheaply, with the bias corrected afterward by DFT relabeling or reweighting; this would be a testable extension on systems where barrier crossing is the bottleneck.
  • The 10x DFT reduction hints at a scaling rule: the number of hard targets needed may scale with the number of distinct structural basins sampled, not with system complexity; a direct test would vary the soft-target pool size and monitor when property accuracy saturates.
  • The authors do not isolate how much of the benefit comes from the teacher's energy underestimation versus the screening step; an experiment replacing the teacher with one of controlled softness would separate the two effects and could guide deliberate barrier tuning.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a two-stage knowledge distillation framework for training lightweight neural network potentials (NNPs) for molecular dynamics: (a) generate soft targets by running MD with a non-fine-tuned, off-the-shelf universal NNP (MatterSim) as the teacher; (b) train a DeepPot-SE student on those soft targets; (c) select 1,000 structures from the soft-target pool using a feature/energy-based screening; (d) label those structures with DFT and fine-tune the student on them. The authors evaluate the approach on PEG (organic) and LGPS (inorganic), reporting that the fine-tuned student reproduces experimental densities and self-diffusion coefficients, reduces DFT training labels by roughly 10x relative to the active-learning baseline GeNNIP4MD, and yields an order-of-magnitude faster inference than the teacher. The central claimed novelty is that using a non-fine-tuned teacher, whose softened PES enables sampling of high-energy configurations, is the key to making the subsequent one-shot DFT fine-tune effective.

Significance. If the causal role of the non-fine-tuned teacher were established, the framework would be a practically valuable contribution: it is simple, applicable to both organic and inorganic systems, and the reported 10x reduction in DFT labels and 10-106x inference speedups are concrete engineering gains. The manuscript is clearly written and the pipeline is reproducible in principle, with specific model versions, DFT settings, and hyperparameters. The central scientific claim, however, rests on a comparison that changes two variables at once, and the accuracy comparisons lack uncertainty quantification. The idea is promising, but the current evidence does not isolate the mechanism the paper advertises.

major comments (3)
  1. [§4.1, Table 1 and the 'existing KD' replication] The comparison that motivates the off-the-shelf teacher is confounded. The 'existing KD' baseline trains a student only on soft targets from a fine-tuned teacher, whereas the proposed method trains on soft targets from the off-the-shelf teacher and then fine-tunes with 1,000 DFT hard targets. Table 1 shows that the soft-target-only student from the off-the-shelf teacher also fails to reproduce the experimental values (density 0.994 g/cm3, diffusion 0.775e-6 cm2/s); accuracy appears only after the hard-target fine-tuning. To attribute the success to the non-fine-tuned teacher, the authors need the control: train a student on the 'Soft targets (Fine-tuned)' and then fine-tune it on the same 1,000 hard targets used in the proposed method. If that control also fails, the off-the-shelf teacher is essential; if it succeeds, the advantage may come solely from the hard-target stage, regardless of teacher choice. This control is missing, so the central claim that the non-fine-tuned teacher is the decisive ingredient is not established.
  2. [§4.1, Figure 4 and the 'gentler PES' premise] The energy histograms in Figure 4 show that the off-the-shelf teacher's soft targets are higher in DFT-relabeled energy than the fine-tuned teacher's soft targets, which supports the sampling mechanism. However, this evidence is indirect: it does not show that a student trained on the fine-tuned teacher's soft targets and then fine-tuned on 1,000 hard targets would fail to reproduce the experimental properties. The paper's own argument (that the fine-tuned teacher samples too few high-energy structures to make the student robust) would predict that such a control would fail, but that prediction is not tested. Without this test, the causal connection between the teacher's 'gentler PES' and the final accuracy remains a hypothesis, not a demonstrated result.
  3. [§4.1, Table 1 and §4.2, Figure 5] All reported MD properties are point estimates from a single trajectory: no block averaging, no multiple independent seeds, and no error bars. The sensitivity visible in Table 1 (random seed 0 gives a diffusion coefficient of 0.170e-6 cm2/s versus 0.288e-6 for seed 1, a 40% relative difference) shows that the estimators are not tightly constrained. Claims of 'comparable or superior accuracy' and the specific '10x fewer DFT labels' argument would be more convincing with uncertainty quantification, at minimum for the proposed-method trajectory. This is a load-bearing gap because the accuracy claim is central to the paper's recommendation.
minor comments (4)
  1. [Throughout] There are several typos and formatting issues: 'L10GeP2S12' should be 'Li10GeP2S12' (the element is Li, not L); 'sutdent models' in the Table 1 caption; 'assesstheaccuracyandefficiency' is missing spaces; 'lager GNNs' in §2 should be 'larger GNNs'; and the abstract says '106x speedup' while the Figure 6 caption says 'up to 107 times'.
  2. [§4.1] The description of the 'existing KD' baseline would benefit from clarity: the fine-tuned teacher used for generating 'Soft targets (Fine-tuned)' was fine-tuned on the 'Random, seed 1' hard-target set, while the proposed method's student is fine-tuned on the screening-selected hard targets. This difference in which structures are used for fine-tuning is not discussed and could affect the interpretation.
  3. [§4.1] The paper reports that the proposed fine-tuned student and the existing-KD student have similar force MAEs (0.061 vs 0.063 eV/Å) yet dramatically different MD properties. This is an interesting observation, but the paper does not discuss why the force MAE is not predictive of MD stability in this comparison; a brief comment would help the reader.
  4. [§4.3] The speedup comparison between the teacher (MatterSim, a GNN) and the student (DeePMD) is informative for practitioners, but it conflates model architecture and implementation choices. A sentence acknowledging that the speedup is specific to these implementations and not a universal property of the KD framework would be more precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: predictions are validated against independent experimental properties and DFT relabeled data, with no fitted target being re-issued as a prediction.

full rationale

The paper's central outputs are student NNPs whose predicted PEG density and self-diffusion coefficient and LGPS lithium-ion diffusion coefficients are obtained from long production NNP-MD runs and compared against experimental values. Those experimental values are never used as training labels or as fitted constants, so the headline predictions are not defined by their inputs. The student is trained on soft targets from the off-the-shelf teacher and then fine-tuned on 1,000 DFT hard targets; the DFT validation set used for accuracy assessment is a held-out relabeled portion of the soft-target pool, and the fine-tuning labels are selected separately. The structural-feature screening tool is imported from the authors' prior GeNNIP4MD work, but the paper also demonstrates that random hard-target selection with one seed reproduces the experimental density, so the screening method is not a load-bearing self-citation that forces the result. The claim that high-energy sampling matters is motivated by the external systematic-softening observation [6] and is additionally supported in this paper by Figure 4, which shows DFT-relabeled energy distributions for soft targets from the pre-trained and fine-tuned teachers. The comparison against the existing KD baseline changes two variables at once (teacher fine-tuning and the addition of hard-target fine-tuning), which is a legitimate experimental-design weakness, but it is not circular because no equation or fitted parameter is equivalent to the predicted physical properties. Consequently, no step in the derivation chain reduces to its own input.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the softness of the teacher PES, the representativeness of the feature-based screening, and the sufficiency of 1,000 DFT labels. These are domain assumptions, not fitted constants, but the free parameters listed above affect the outcome. No new physical entities are introduced.

free parameters (3)
  • number_of_hard_targets = 1000
    Set to 1000 for both PEG and LGPS; the paper tests 500 for PEG and reports degraded accuracy (density deviation -8%), showing the central result is sensitive to this choice. It is a hand-picked budget rather than a fitted constant.
  • soft_target_sampling_temperatures = 300-600 K (PEG), 300-1500 K (LGPS)
    Chosen by the authors to cover relevant conditions; not optimized and not shown to be sufficient for other materials.
  • teacher_md_duration_and_extraction_frequency = 150 ps / every 50 fs (PEG), 110 ps / every 50 fs (LGPS)
    Arbitrary choices affecting the pool of soft targets and the subsequent selection; no sensitivity analysis is reported.
assumptions (6)
  • domain assumption Pre-trained universal NNPs underestimate energies in high-energy regions, producing a gentler PES that lowers energy barriers in MD.
    Invoked in Sections 1 and 3(a); based on reference [6]. This is the core premise that the off-the-shelf teacher's MD explores high-energy structures more efficiently than a fine-tuned teacher.
  • domain assumption Fine-tuning a teacher with DFT data increases energy barriers and reduces exploration of high-energy structures.
    Demonstrated for one configuration in Section 4.1 via the Soft targets (Fine-tuned) experiment; assumed to hold generally.
  • domain assumption The structural feature-based screening method (GeNNIP4MD) selects representative structures such that 1000 DFT labels are sufficient to correct the student's PES.
    Borrowed from [26]; the paper's own two-seed random comparison is too small to prove robustness, so this is an unverified premise.
  • domain assumption DFT labels (BLYP-D3 for PEG, PBE for LGPS) are accurate enough to serve as ground truth for training and evaluation.
    Standard practice; quoted DFT settings in Sections 4.1 and 4.2.
  • domain assumption The student DeepPot-SE retains the teacher's knowledge after fine-tuning only the fitting network while fixing descriptors.
    Section 3(d); assumes hard-target structures lie within the soft-target distribution, so descriptor updates are unnecessary.
  • domain assumption Fixed random seeds and a single 21 ns MD trajectory yield converged estimates of density and self-diffusion coefficient.
    Not stated explicitly; the paper does not report block averaging or seed repeats, yet treats these values as definitive.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge Distillation Framework for Accelerating High-Accuracy Neural Network-Based Molecular Dynamics Simulations." pith.science (2026). https://pith.science/paper/GBEK6UEA

@misc{pith2026250615337,
  author       = {Pith},
  title        = {Pith review of: Knowledge Distillation Framework for Accelerating High-Accuracy Neural Network-Based Molecular Dynamics Simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GBEK6UEA}},
  note         = {Machine review of arXiv:2506.15337}
}
abstract

Neural network potentials (NNPs) offer a powerful alternative to traditional force fields for molecular dynamics (MD) simulations. Accurate and stable MD simulations, crucial for evaluating material properties, require training data encompassing both low-energy stable structures and high-energy structures. Conventional knowledge distillation (KD) methods fine-tune a pre-trained NNP as a teacher model to generate training data for a student model. However, in material-specific models, this fine-tuning process increases energy barriers, making it difficult to create training data containing high-energy structures. To address this, we propose a novel KD framework that leverages a non-fine-tuned, off-the-shelf pre-trained NNP as a teacher. Its gentler energy landscape facilitates the exploration of a wider range of structures, including the high-energy structures crucial for stable MD simulations. Our framework employs a two-stage training process: first, the student NNP is trained with a dataset generated by the off-the-shelf teacher; then, it is fine-tuned with a smaller, high-accuracy density functional theory (DFT) dataset. We demonstrate the effectiveness of our framework by applying it to both organic (polyethylene glycol) and inorganic (L$_{10}$GeP$_{2}$S$_{12}$) materials, achieving comparable or superior accuracy in reproducing physical properties compared to existing methods. Importantly, our method reduces the number of expensive DFT calculations by 10x compared to existing NNP generation methods, without sacrificing accuracy. Furthermore, the resulting student NNP achieves up to 106x speedup in inference compared to the teacher NNP, enabling significantly faster and more efficient MD simulations.

Figures

Figures reproduced from arXiv: 2506.15337 by the authors.

Figure 1
Figure 1. Illustration of the potential energy surface (PES). (a) PES calculated from DFT calculations and (b) PES of pre-trained universal NNP (blue) and a DFT fine-tuned universal NNP (orange). While fine-tuning with DFT data enhances the reproducibil￾ity of the DFT PES, especially in the high-energy region, it can lead to a reduced exploration of structures in the high-energy region. empirical force fields. By leveraging t… view at source ↗
Figure 2
Figure 2. Overview of the proposed knowledge distillation framework. (a) NNP-MD sim￾ulations are performed on initial structures using a non-fine-tuned, off-the-shelf teacher model to generate soft targets. (b) A student model is then trained on the soft tar￾gets. (c) Data points are selected from the soft targets using a structural feature-based screening method [26] and converted into hard targets using DFT calculations. (d… view at source ↗
Figure 3
Figure 3. Atomic structures used in this work. (a) PEG system with five 4-mer molecules (155 atoms) for training data creation. (b) PEG system with one hundred 4-mer molecules (3100 atoms) for production MD simulations. (c) LGPS system containing two formula units (50 atoms) for training data creation. (d) LGPS system containing 64 formula units (1600 atoms) for production MD simulations. datasets with an 8:2 ratio, yielding … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Energy histograms at 300 K for structures in the Soft targets (Pre-trained) (blue) and the Soft targets (Fine-tuned) (orange). These energies were relabeled by DFT as the ground truth. The distribution of the Soft targets (Fine-tuned) shifts towards lower values compar…
Figure 5
Figure 5. Figure 5: Arrhenius plot of lithium-ion self-diffusion coefficients for LGPS, obtained from experimental results [23], AIMD simulations [27], NNP-MD simulations using PaiNN model [42], GeNNIP4MD [26], the teacher model (MatterSim-v1.0.0-5M) [44] and our proposed DP model (traine…
Figure 6
Figure 6. Figure 6: Computation time of teacher model (MatterSim-v1.0.0-5M [44]) and student model (DP [50]) for (a) PEG and (b) LGPS on an NVIDIA H100 80 GB GPU. For the PEG, the teacher model’s high memory usage restricts simulations of PEG to a maximum of 4096 atoms. The student model …

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages

  1. [1]

    Physical Review A38(6), 3098–3100 (1988)

    Becke, A.D.: Density-functional exchange-energy approximation with correct asymptotic behavior. Physical Review A38(6), 3098–3100 (1988)

    work page 1988

  2. [2]

    Physical Review Letters98(14), 146401 (2007)

    Behler, J., Parrinello, M.: Generalized Neural-Network Representation of High- Dimensional Potential-Energy Surfaces. Physical Review Letters98(14), 146401 (2007)

    work page 2007

  3. [3]

    Physical Review B 50(24), 17953–17979 (1994)

    Blöchl, P.E.: Projector augmented-wave method. Physical Review B 50(24), 17953–17979 (1994)

    work page 1994

  4. [4]

    ACS Applied Materials & Interfaces16(28), 36878–36891 (2024)

    Chahal, R., Toomey, M.D., Kearney, L.T., et al.: Deep-Learning Interatomic Po- tential Connects Molecular Structural Ordering to the Macroscale Properties of Polyacrylonitrile. ACS Applied Materials & Interfaces16(28), 36878–36891 (2024)

    work page 2024

  5. [5]

    Nature Computational Science2(11), 718–728 (2022)

    Chen, C., Ong, S.P.: A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science2(11), 718–728 (2022)

    2022

  6. [6]

    npj Computational Materials11(1), 9 (2025)

    Deng, B., Choi, Y., Zhong, P., et al.: Systematic softening in universal machine learning interatomic potentials. npj Computational Materials11(1), 9 (2025)

    work page 2025

  7. [7]

    Nature Machine Intel- ligence 5(9), 1031–1041 (2023)

    Deng, B., Zhong, P., Jun, K., et al.: CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intel- ligence 5(9), 1031–1041 (2023)

    work page 2023

  8. [8]

    Nature Materials20(6), 750–761 (2021)

    Friederich, P., Häse, F., Proppe, J., Aspuru-Guzik, A.: Machine-learned potentials for next-generation matter simulations. Nature Materials20(6), 750–761 (2021)

    2021

Show all 51 references
  1. [9]

    Journal of Physics

    Giannozzi, P., Baroni, S., Bonini, N., et al.: QUANTUM ESPRESSO: a modular and open-source software project for quantum simulations of materials. Journal of Physics. Condensed Matter: An Institute of Physics Journal21(39), 395502 (2009)

    2009

  2. [10]

    The Journal of Chemical Physics132(15), 154104 (2010)

    Grimme, S., Antony, J., Ehrlich, S., Krieg, H.: A consistent and accurateab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. The Journal of Chemical Physics132(15), 154104 (2010)

    2010

  3. [11]

    ACM Trans

    Gupta, M., Agrawal, P.: Compression of Deep Learning Models for Text: A Survey. ACM Trans. Knowl. Discov. Data16(4), 61:1–61:55 (2022) Title Suppressed Due to Excessive Length 15

    2022

  4. [12]

    Physical Review B58(7), 3641–3662 (1998)

    Hartwigsen, C., Goedecker, S., Hutter, J.: Relativistic separable dual-space Gaus- sian pseudopotentials from H to Rn. Physical Review B58(7), 3641–3662 (1998)

    1998

  5. [13]

    npj Computational Materials8(1), 1–15 (2022)

    Hayashi, Y., Shiomi, J., Morikawa, J., Yoshida, R.: RadonPy: automated physical property calculation using all-atom classical molecular dynamics simulations for polymer informatics. npj Computational Materials8(1), 1–15 (2022)

    2022

  6. [14]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778. IEEE (2016)

    2016

  7. [15]

    Hinton, G., Vinyals, O., Dean, J.: Distilling the Knowledge in a Neural Network (2015), arXiv:1503.02531

    2015 arXiv

  8. [16]

    Journal of Chemical & Engineering Data66(6), 2480–2500 (2021)

    Hoffmann, M.M., Horowitz, R.H., Gutmann, T., Buntkowsky, G.: Densities, Vis- cosities, and Self-Diffusion Coefficients of Ethylene Glycol Oligomers. Journal of Chemical & Engineering Data66(6), 2480–2500 (2021)

    2021

  9. [17]

    Current Opinion in Solid State and Materials Science35, 101214 (2025)

    Jacobs, R., Morgan, D., Attarian, S., et al.: A practical guide to machine learning interatomic potentials – Status and future. Current Opinion in Solid State and Materials Science35, 101214 (2025)

    2025

  10. [18]

    APL Materials 1(1), 011002 (2013)

    Jain, A., Ong, S.P., Hautier, G., et al.: Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Materials 1(1), 011002 (2013)

    2013

  11. [19]

    Jiao, X., Yin, Y., Shang, L., et al.: TinyBERT: Distilling BERT for Natural Lan- guage Understanding (2019), arXiv:1909.10351

    2019 arXiv

  12. [20]

    Kaplan, A.D., Liu, R., Qi, J., et al.: A Foundational Potential Energy Surface Dataset for Materials (2025)

    2025

  13. [21]

    In: Proceedings of the 37th Inter- national Conference on Neural Information Processing Systems

    Kelvinius, F.E., Georgiev, D., Toshev, A.P., Gasteiger, J.: Accelerating molecular graph neural networks via knowledge distillation. In: Proceedings of the 37th Inter- national Conference on Neural Information Processing Systems. Curran Associates Inc. (2023)

    2023

  14. [22]

    In: Interna- tional Conference on Learning Representations (ICLR) (2015)

    Kingma, D.P., Lei, J.: Adam: A Method for Stochastic Optimization. In: Interna- tional Conference on Learning Representations (ICLR) (2015)

    2015

  15. [23]

    Energy & Environmental Science 6(12), 3548–3552 (2013)

    Kuhn, A., Duppel, V., Lotsch, B.V.: Tetragonal Li10GeP2S12 and Li7GePS8 – exploring the Li ion dynamics in LGPS Li electrolytes. Energy & Environmental Science 6(12), 3548–3552 (2013)

    2013

  16. [24]

    Chemistry of Materials36(6), 2898–2906 (2024)

    Liu, D., Wu, Y., Samatov, M.R., et al.: Compression Eliminates Charge Traps by Stabilizing Perovskite Grain Boundary Structures: An Ab Initio Analysis with Machine Learning Force Field. Chemistry of Materials36(6), 2898–2906 (2024)

    2024

  17. [25]

    Journal of Chemical Theory and Com- putation 17(7), 4291–4300 (2021)

    Lu, C., Wu, C., Ghoreishi, D., et al.: OPLS4: Improving Force Field Accuracy on Challenging Regimes of Chemical Space. Journal of Chemical Theory and Com- putation 17(7), 4291–4300 (2021)

    2021

  18. [26]

    Journal of Chemical The- ory and Computation21(8), 3832–3846 (2025)

    Matsumura, N., Yoshimoto, Y., Yamazaki, T., et al.: Generator of Neural Network Potential for Molecular Dynamics: Constructing Robust and Accurate Potentials with Active Learning for Nanosecond-Scale Simulations. Journal of Chemical The- ory and Computation21(8), 3832–3846 (2025)

    2025

  19. [27]

    Chemistry of Materials 24(1), 15–17 (2012)

    Mo, Y., Ong, S.P., Ceder, G.: First Principles Study of the Li10 GeP 2 S 12 Lithium Super Ionic Conductor Material. Chemistry of Materials 24(1), 15–17 (2012)

    2012

  20. [28]

    Scientific Reports13(1), 17251 (2023)

    Mohanty, S., Stevenson, J., Browning, A.R., et al.: Development of scalable and generalizable machine learned force field for polymers. Scientific Reports13(1), 17251 (2023)

    2023

  21. [29]

    Nature Biotechnology39(6), 765– 774 (2021) 16 N

    Narayan, A., Berger, B., Cho, H.: Assessing single-cell transcriptomic variability through density-preserving data visualization. Nature Biotechnology39(6), 765– 774 (2021) 16 N. Matsumura et al

    2021

  22. [30]

    Physical Review Letters77(18), 3865–3868 (1996)

    Perdew,J.P.,Burke,K.,Ernzerhof,M.:GeneralizedGradientApproximationMade Simple. Physical Review Letters77(18), 3865–3868 (1996)

    1996

  23. [31]

    npj Computational Materials10(1), 43 (2024)

    Qi, J., Ko, T.W., Wood, B.C., et al.: Robust training of machine learning in- teratomic potentials with dimensionality reduction and stratified sampling. npj Computational Materials10(1), 43 (2024)

    2024

  24. [32]

    Romero, A., Ballas, N., Kahou, S.E., et al.: FitNets: Hints for Thin Deep Nets (2014), arXiv:1412.6550

    2014 arXiv

  25. [33]

    Sanh, V., Debut, L., Chaumond, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter (2019), arXiv:1910.01108

    2019 arXiv

  26. [34]

    In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV)

    Saputra, M.R.U., Gusmao, P., Almalioglu, Y., et al.: Distilling Knowledge From a Deep Pose Regressor Network. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 263–272. IEEE (2019)

    2019

  27. [35]

    In: Proceedings of the 38th International Conference on Machine Learning

    Schütt, K., Unke, O., Gastegger, M.: Equivariant message passing for the pre- diction of tensorial properties and molecular spectra. In: Proceedings of the 38th International Conference on Machine Learning. pp. 9377–9388. PMLR (2021)

    2021

  28. [36]

    Advanced Materials 35(22), 2210788 (2023)

    Schmidt, J., Hoffmann, N., Wang, H.C., et al.: Machine-Learning-Assisted Deter- mination of the Global Zero-Temperature Phase Diagram of Materials. Advanced Materials 35(22), 2210788 (2023)

    2023

  29. [37]

    Chemical Reviews121(16), 10142–10186 (2021)

    Unke, O.T., Chmiela, S., Sauceda, H.E., et al.: Machine Learning Force Fields. Chemical Reviews121(16), 10142–10186 (2021)

    2021

  30. [38]

    Computer Physics Communications228, 178–184 (2018)

    Wang, H., Zhang, L., Han, J., E, W.: DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics. Computer Physics Communications228, 178–184 (2018)

    2018

  31. [39]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(6), 3048–3068 (2022)

    Wang, L., Yoon, K.J.: Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks. IEEE Transactions on Pattern Analysis and Machine Intelligence44(6), 3048–3068 (2022)

    2022

  32. [40]

    Wang, R., Gao, Y., Wu, H., Zhong, Z.: PFD: Automatically Generating Machine Learning Force Fields from Universal Models (2025), arXiv:2502.20809

    2025

  33. [41]

    ACS Materials Letters7(6), 2105–2114 (2025)

    Wines, D., Choudhary, K.: CHIPS-FF: Evaluating Universal Machine Learning Force Fields for Material Properties. ACS Materials Letters7(6), 2105–2114 (2025)

    2025

  34. [42]

    Journal of Physics: Energy5(2), 024004 (2023)

    Winter, G., Gómez-Bombarelli, R.: Simulations with machine learning poten- tials identify the ion conduction mechanism mediating non-Arrhenius behavior in LGPS. Journal of Physics: Energy5(2), 024004 (2023)

    2023

  35. [43]

    ACM Trans

    Yang, C., Zhu, Y., Lu, W., et al.: Survey on Knowledge Distillation for Large Language Models: Methods, Evaluation, and Application. ACM Trans. Intell. Syst. Technol. (2024)

    2024

  36. [44]

    Yang, H., Hu, C., Zhou, Y., et al.: MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures (2024), arXiv:2405.04967

    2024 arXiv

  37. [45]

    In: 2017 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR)

    Yim, J., Joo, D., Bae, J., Kim, J.: A Gift from Knowledge Distillation: Fast Op- timization, Network Minimization and Transfer Learning. In: 2017 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR). pp. 7130–7138. IEEE (2017)

    2017

  38. [46]

    Yoshimoto, Y., Matsumura, N., Iwasaki, Y., et al.: Large-Scale, Long-Time Atom- istic Simulations of Proton Transport in Polymer Electrolyte Membranes Using a Neural Network Interatomic Potential (2025), arXiv:2503.20412

    2025 arXiv

  39. [47]

    The Journal of Chemical Physics159(5), 054801 (2023)

    Zeng, J., Zhang, D., Lu, D., et al.: DeePMD-kit v2: A software package for deep potential models. The Journal of Chemical Physics159(5), 054801 (2023)

    2023

  40. [48]

    Nature639(8055), 624–632 (2025)

    Zeni, C., Pinsler, R., Zügner, D., et al.: A generative model for inorganic materials design. Nature639(8055), 624–632 (2025)

    2025

  41. [49]

    npj Computational Materials10(1), 293 (2024) Title Suppressed Due to Excessive Length 17

    Zhang, D., Liu, X., Zhang, X., et al.: DPA-2: a large atomic model as a multi-task learner. npj Computational Materials10(1), 293 (2024) Title Suppressed Due to Excessive Length 17

    2024

  42. [50]

    In: Advances in Neural Information Processing Systems

    Zhang, L., Han, J., Wang, H., et al.: End-to-end Symmetry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems. In: Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018)

    2018

  43. [51]

    Computer Physics Communications253, 107206 (2020)

    Zhang,Y.,Wang,H.,Chen,W.,etal.:DP-GEN:Aconcurrentlearningplatformfor the generation of reliable deep learning based potential energy models. Computer Physics Communications253, 107206 (2020)

    2020

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.