Pith. sign in

REVIEW 1 major objections 1 minor 2 cited by

An efficient forgetting-aware fine-tuning framework for pretrained universal machine-learning interatomic potentials

T0 review · 1 major / 1 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 Combining Experience Replay with Elastic Weight Consolidation lets a pretrained universal machine-learning interatomic potential be fine-tuned to a new material while preserving the general knowledge it was pretrained on.

desk verdict A genuinely useful empirical study of replay-plus-EWC fine-tuning for universal MLIPs, but its headline forgetting metric may be partly memorization because the Replay set and the sMPtrj test set are both random 10% samples of MPtrj with no disjointness stated. read the letter →

arxiv 2506.15223 v1 pith:MA3M2NAK submitted 2025-06-18 cond-mat.mtrl-sci

Jisu Kim , Jiho Lee , Sangmin Oh , Yutack Park , Seungwoo Hwang , Seungwu Han , Sungwoo Kang , Youngho Kang This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords machine-learninginteratomicpotentialscatastrophicforgettingfine-tuningelasticweightconsolidationexperiencereplaysolid-stateelectrolytespotentialenergysurfacesofteningcontinuallearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that a universal machine-learning interatomic potential (an MLIP, a neural-network model that predicts atomic energies and forces) can be fine-tuned to a new material without losing the broad chemical knowledge it was pretrained on. The proposed method, reEWC, combines two established forgetting-prevention techniques: Experience Replay, which re-trains on a sample of the original data, and Elastic Weight Consolidation (EWC), which penalizes changes to parameters the pretrained model depends on. The authors test it by fine-tuning the pretrained SevenNet-0 model on the solid electrolyte Li$_6$PS$_5$Cl, and report that reEWC corrects the known 'softening' of the potential-energy surface, gives Li diffusivities close to DFT reference values, and improves accuracy on other sulfide, oxide, nitride, and halide electrolytes. The deeper claim is that Replay and EWC are not redundant: each covers a weakness of the other, and their combination is more robust to replay-set composition than Replay alone. A reader should care because a practical anti-forgetting fine-tuning recipe would let one universal model be continually adapted to new materials while remaining trustworthy for high-throughput screening.

What carries the argument

The load-bearing object is the diagonal Fisher information matrix, computed on a filtered subset of 1,154,095 MPtrj configurations with low Huber loss. It supplies per-parameter importance scores $F_i$; EWC then adds the penalty $\frac{\lambda}{2}\sum_i F_i(\theta_i-\theta_{i,\mathrm{pre}})^2$ to the fine-tuning loss, pinning down parameters whose movement would destroy pretrained knowledge. Replay contributes a second ingredient: in each epoch, after a mini-batch from the LPSC fine-tuning set, a same-size mini-batch drawn from a 10\% random sample of MPtrj is used for an extra parameter update. reEWC is simply the two mechanisms run together, with $\lambda=10^5$ (compared with $10^6$ for EWC alone), and the claim is that the replay signal makes the Fisher penalty more effective while the penalty prevents the large, unconstrained parameter shifts that make Replay-only models unstable on out-of-domain materials. The noise-perturbation experiment (accuracy survives larger perturbations when noise is scaled by the inverse Fisher matrix) is the direct evidence that these importance scores are meaningful.

What would settle it

Recompute the Fisher matrix on the full MPtrj dataset instead of the 1,154,095 high-accuracy subset, retrain reEWC with that matrix, and compare held-out pretraining (sMPtrj) loss and LPSC accuracy; if the results are nearly identical, the subset choice is not load-bearing, and if they differ materially, the core mechanism depends on that filtering.

Watch

Extended reading notes

Core claim

The central claim is that reEWC—running the Replay mini-batch schedule while adding the EWC Fisher-information penalty—achieves a favorable stability–plasticity balance that neither method reaches alone. The paper reports that on a held-out sample of the pretraining set (sMPtrj), reEWC keeps loss close to the pretrained model's own loss, better than EWC even though its parameters move more; on the target LPSC dataset it reduces energy and force errors and removes the systematic softening that makes the pretrained model overestimate Li diffusivity. It also reports that reEWC transfers improved accuracy across chemically distinct solid electrolytes, reproduces DFT quasi-melting behavior, and avoids the unphysical short Li–Li bonds that arise in MD with a Vanilla- or Replay-only fine-tuned model. The conclusion the authors draw is that reEWC should be a default fine-tuning strategy for continual learning in pretrained MLIPs.

Load-bearing premise

The method assumes that Fisher-importance scores computed from a filtered subset of the pretraining data correctly identify which parameters must stay fixed, even though fine-tuning follows structured loss gradients rather than random noise.

Editorial extensions

If this is right

  • Fine-tuned reEWC models keep near-pretrained accuracy on the original pretraining domain (sMPtrj loss comparable to SevenNet-0), so they remain safe for general materials screening.
  • The known PES softening and overestimated Li diffusivity of the pretrained model are corrected on LPSC, making simulated ion transport realistic.
  • Knowledge gained from one fine-tuning target transfers to chemically related systems, including PS$_4$-containing sulfides and also oxides, nitrides, and halides.
  • reEWC remains stable when the replay set is changed, e.g., restricted to Li-containing compounds, whereas Replay alone forgets under the same change.
  • The added cost is modest: reEWC takes about twice the vanilla fine-tuning time, which the authors consider acceptable for lab-scale GPU resources.

Reading between the lines

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

  • If this recipe is as robust as reported, the same replay-plus-Fisher combination should apply to other pretrained universal interatomic potentials and other target materials; the paper demonstrates it on one model and one target, so that extension is unverified.
  • The filtered-Fisher trick also suggests a cheap pre-deployment check: compute where the pretrained model is confident, and use those parameters as anchors before adapting it to a niche.
  • The Li$_3$N finding implies that loss-based forgetting metrics can miss catastrophic dynamical failure; reEWC's advantage may be as much about bounding parameter shifts as about protecting loss values. This is an editorial reading, not a claim the paper makes explicitly.
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

1 major / 1 minor

Summary. The paper proposes reEWC, a hybrid continual-learning fine-tuning strategy that combines Experience Replay and Elastic Weight Consolidation (EWC) for pretrained universal machine-learning interatomic potentials (MLIPs). Using SevenNet-0, pretrained on the MPtrj dataset, as the base model and Li6PS5Cl (LPSC) as the fine-tuning target, the authors compare Vanilla fine-tuning, EWC, Replay, and reEWC. They report that reEWC simultaneously achieves low target-domain losses, reduced forgetting on the pretraining domain (as measured by loss on a random 10% MPtrj subset they call sMPtrj), improved energy/force accuracy and softening scales on 126 argyrodite and 9 non-argyrodite solid-electrolyte test sets, Li diffusivities close to DFT references, and quasi-melting ratios consistent with AIMD. The central claim is that reEWC provides a synergistic balance of stability and plasticity, preserving generalizability while learning the target system, at training cost comparable to Replay and with greater robustness to the composition of the Replay set.

Significance. If the central claim holds, the paper offers a practical, computationally efficient recipe for fine-tuning universal MLIPs without catastrophic forgetting, which is directly relevant to materials discovery workflows. The study is unusually thorough in its validation: energy/force MAEs, softening scales, diffusivities, and quasi-melting ratios are benchmarked against independent DFT/AIMD references that were not used in training, and the source data and code are publicly released. The separation of learning, forgetting, and generalizability metrics, the explicit examination of parameter shifts (MAD), the FIM-aware noise-perturbation experiment, and the MD-stability case studies (Li3N, high-entropy argyrodites) are notable strengths. However, the central forgetting metric, sMPtrj, may be contaminated by overlap with the Replay training set, and the FIM-based regularizer relies on hand-tuned-though-robust hyperparameters and on a hand-filtered Fisher subset; these issues must be resolved before the synergistic-advantage claim can be fully credited.

major comments (1)
  1. [§4.2, Fig. S9] The FIM is computed on a hand-filtered subset of the pretrained dataset (1,154,095 configurations with Huber loss below 0.000726), and the regularization strengths are set to different values for EWC (λ=10^6) and reEWC (λ=10^5) after manual exploration. While the paper reports in Fig. S9 that the conclusions are insensitive to reasonable variations in the λ values, no equivalent sensitivity analysis is provided for the FIM-subset loss threshold. Because the EWC and reEWC regularizers rely entirely on the FIM to identify parameters that must stay fixed, an ill-calibrated subset could cause either under-regularization (forgetting) or over-regularization (limited target learning) in a way that is not captured by the current experiments. Please add an ablation over the threshold (or at least report the fraction of parameters that change materially as the threshold is varied) to substantiate that the filter is not load-bearing for the reported comparisons.
minor comments (1)
  1. [§2.2, §4.2] Table 1 reports training cost as ×1 for Vanilla and EWC and ×2 for Replay and reEWC, but the main text (§3) only states that reEWC cost is 'comparable' to Replay; clarifying whether the ×2 factor refers to wall-clock time on the same GPU and whether it includes the FIM computation would improve the efficiency claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central reEWC claims are benchmarked against external DFT references rather than derived from fitted inputs.

full rationale

The paper's derivation chain does not reduce to its own inputs. reEWC is defined as a combination of standard Replay and EWC losses (Eq. 2 plus the replay mini-batch update), and its claimed benefits are evaluated on LPSC and transfer materials through energy/force MAEs, softening scales, Li diffusivities, and quasi-melting ratios, all compared with DFT data generated independently of the fine-tuning procedure. The Fisher information matrix is a standard sensitivity measure computed from pretrained gradients; selecting its subset by low Huber loss is a stated methodological choice, not a target-derived fit. The main potential benchmark-hygiene issues—that the Replay set and sMPtrj are both random 10% samples of MPtrj without a stated disjointness guarantee, and that lambda values are chosen with sMPtrj losses in view—are validity concerns about the forgetting metric, but they are not circular reductions: the external DFT benchmarks and the reported insensitivity of conclusions to lambda (Fig. S9) provide independent support for the central claim. No load-bearing self-citation, imported uniqueness theorem, or ansatz-by-citation appears; citations to SevenNet-0 and atomic-energy mapping are tools, not premises that assume the paper's conclusions.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The load-bearing inputs are the EWC Fisher approximation, the choice of a low-loss subset for computing the FIM, hand-tuned regularizer strengths, the Replay sampling ratio, and the MD-derived reference data. These are engineering choices rather than new physics, but they affect the strength of all empirical claims.

free parameters (5)
  • EWC regularization strength lambda_EWC = 10^6
    Chosen by exploring values and balancing losses on the fine-tuning validation set and sMPtrj (Methods 4.2, Fig. S9). The comparison uses a different lambda for EWC than for reEWC.
  • reEWC regularization strength lambda_reEWC = 10^5
    Chosen by the same balancing procedure as lambda_EWC (Methods 4.2, Fig. S9). The differing values complicate a like-for-like method comparison.
  • FIM subset loss threshold = 0.000726
    Hand-selected to yield 1,154,095 pretrained configurations with low Huber loss for computing the Fisher information (Methods 4.2). Affects which parameters EWC and reEWC protect.
  • Replay set size = 10% of MPtrj
    Random 10% of the pretrained dataset is used as the Replay set (Section 2.2). The fraction is not swept, yet the method's cost and forgetting prevention depend on it.
  • Quasi-melting MSD threshold = 0.36 Angstrom^2/ps
    Defines when a structure is considered melted (Section 2.5). The choice of this cutoff affects the WMAE comparison between MLIP and DFT.
assumptions (6)
  • standard math The pretraining posterior p(theta|D_pre) is approximated as Gaussian with curvature given by the diagonal Fisher information matrix (Eqs. 4-7).
    This is the standard EWC approximation from Kirkpatrick et al. (Ref. 33). It is an assumption about the loss landscape, and the paper does not verify it for SevenNet-0 beyond the noise-perturbation experiment.
  • domain assumption DFT (PBE) energies and forces are ground truth for all test metrics.
    All benchmark references and the fine-tuning labels come from PBE DFT calculations. The comparisons inherit any DFT functional errors.
  • domain assumption SevenNet-0 pretrained on MPtrj is representative of universal MLIPs for the purpose of transfer conclusions.
    Only one base model is tested. The claimed generalizability of reEWC across pretrained MLIP architectures is not established.
  • domain assumption The fine-tuning target LPSC is representative of argyrodite solid electrolytes and its PES features (PS4 backbone) transfer to related chemistries.
    Knowledge transfer is attributed to shared PS4 units (Section 2.4), but this is inferred from correlations, not proven causally.
  • domain assumption Atomic energy decomposition used in Fig. 7d is physically meaningful and the reference MLIP provides valid atomic energies.
    The paper acknowledges DFT cannot provide atomic energies directly, and uses an MLIP fine-tuned on the same material as reference. This assumes local energy decompositions are comparable across MLIPs.
  • domain assumption 100 ps MD trajectories are sufficient to estimate Li diffusivity and quasi-melting behavior.
    The authors use five independent 100 ps runs and standard analysis windows, but longer trajectories would improve statistical reliability, especially at lower temperatures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An efficient forgetting-aware fine-tuning framework for pretrained universal machine-learning interatomic potentials." pith.science (2026). https://pith.science/paper/MA3M2NAK

@misc{pith2026250615223,
  author       = {Pith},
  title        = {Pith review of: An efficient forgetting-aware fine-tuning framework for pretrained universal machine-learning interatomic potentials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MA3M2NAK}},
  note         = {Machine review of arXiv:2506.15223}
}
abstract

Pretrained universal machine-learning interatomic potentials (MLIPs) have revolutionized computational materials science by enabling rapid atomistic simulations as efficient alternatives to ab initio methods. Fine-tuning pretrained MLIPs offers a practical approach to improving accuracy for materials and properties where predictive performance is insufficient. However, this approach often induces catastrophic forgetting, undermining the generalizability that is a key advantage of pretrained MLIPs. Herein, we propose reEWC, an advanced fine-tuning strategy that integrates Experience Replay and Elastic Weight Consolidation (EWC) to effectively balance forgetting prevention with fine-tuning efficiency. Using Li$_6$PS$_5$Cl (LPSC), a sulfide-based Li solid-state electrolyte, as a fine-tuning target, we show that reEWC significantly improves the accuracy of a pretrained MLIP, resolving well-known issues of potential energy surface softening and overestimated Li diffusivities. Moreover, reEWC preserves the generalizability of the pretrained MLIP and enables knowledge transfer to chemically distinct systems, including other sulfide, oxide, nitride, and halide electrolytes. Compared to Experience Replay and EWC used individually, reEWC delivers clear synergistic benefits, mitigating their respective limitations while maintaining computational efficiency. These results establish reEWC as a robust and effective solution for continual learning in MLIPs, enabling universal models that can advance materials research through large-scale, high-throughput simulations across diverse chemistries.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Active learning and explicit electrostatics enable accurate modeling of electrolytes

    physics.chem-ph 2025-10 conditional novelty 6.0 of 10

    Automated active learning produces transferable MTPs for EC/EMC/LiPF6 electrolytes, while explicit electrostatics (QRd) matches accuracy with fewer parameters only where its MD remains stable.

  2. Heterogeneous Ensemble Enables a Universal Uncertainty Metric for Atomistic Foundation Models

    cond-mat.mtrl-sci 2025-07 conditional novelty 5.0 of 10

    Weights from OMat24 force errors turn an eleven-model heterogeneous uMLIP ensemble into an uncertainty metric U that correlates with true force errors across material families and drives low-DFT distillation.

Reference graph

Works this paper leans on

69 extracted references · 61 canonical work pages · cited by 2 Pith papers

  1. [1]

    Deringer, V.L., Caro, M.A., Cs´ anyi, G.: Machine learning interatomic potentials as emerging tools for materials science. Adv. Mater. 31(46), 1902765 (2019)

    work page 2019

  2. [2]

    Behler, J., Parrinello, M.: Generalized neural- network representation of high-dimensional potential-energy surfaces. Phys. Rev. Lett. 98(14), 146401 (2007)

    work page 2007

  3. [3]

    Bart´ ok, A.P., Payne, M.C., Kondor, R., Cs´ anyi, G.: Gaussian approximation poten- tials: The accuracy of quantum mechan- ics, without the electrons. Phys. Rev. Lett. 104(13), 136403 (2010)

    work page 2010

  4. [4]

    Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J.P., Kornbluth, M., Molinari, N., Smidt, T.E., Kozinsky, B.: E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nat. Com- mun. 13(1), 2453 (2022)

    work page 2022

  5. [5]

    Preprint at arXiv:2206.07697 (2022)

    Batatia, I., Kovacs, D.P., Simm, G., Ortner, C., Csanyi, G.: Mace: Higher order equiv- ariant message passing neural networks for fast and accurate force fields. Preprint at arXiv:2206.07697 (2022)

    arXiv 2022

  6. [6]

    Griesemer, S.D., Xia, Y., Wolverton, C.: Accelerating the prediction of stable materi- als with machine learning. Nat. Comput. Sci. 3(11), 934–945 (2023)

    work page 2023

  7. [7]

    Pilania, G., Wang, C., Jiang, X., Rajasekaran, S., Ramprasad, R.: Acceler- ating materials property predictions using machine learning. Sci. Rep. 3(1), 2810 (2013)

    work page 2013

  8. [8]

    Hwang, S., Jung, J., Hong, C., Jeong, W., Kang, S., Han, S.: Stability and equilib- rium structures of unknown ternary metal oxides explored by machine-learned poten- tials. J. Am. Chem. Soc. 145(35), 19378– 19386 (2023)

    work page 2023

Show all 69 references
  1. [9]

    ACS Appl

    Lee, J., Ju, S., Hwang, S., You, J., Jung, J., Kang, Y., Han, S.: Disorder-dependent Li dif- fusion in Li 6PS5Cl investigated by machine- learning potential. ACS Appl. Mater. Inter- faces 16(35), 46442–46453 (2024)

    2024

  2. [10]

    ACS Appl

    Hong, C., Oh, S., An, H., Kim, P.-h., Kim, Y., Ko, J.-h., Sue, J., Oh, D., Park, S., Han, S.: Atomistic simulation of hf etching process of amorphous Si 3N4 using machine learn- ing potential. ACS Appl. Mater. Interfaces 16(36), 48457–48469 (2024)

    2024

  3. [11]

    Choi, J.M., Lee, K., Kim, S., Moon, M., Jeong, W., Han, S.: Accelerated computation 21 of lattice thermal conductivity using neu- ral network interatomic potentials. Comput. Mater. Sci. 211, 111472 (2022)

    2022

  4. [12]

    ACS Mater

    Kang, S., Han, S., Kang, Y.: First-principles calculations of luminescence spectra of real- scale quantum dots. ACS Mater. Au 2(2), 103–109 (2021)

    2021

  5. [13]

    ACS Catal

    Jung, J., Ju, S., Kim, P.-h., Hong, D., Jeong, W., Lee, J., Han, S., Kang, S.: Electro- chemical degradation of Pt 3Co nanoparticles investigated by off-lattice kinetic monte carlo simulations with machine-learned potentials. ACS Catal. 13(24), 16078–16087 (2023)

    2023

  6. [14]

    APL Mater

    Jain, A., Ong, S.P., Hautier, G., Chen, W., Richards, W.D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., Persson, K.A.: Commentary: The Materials Project: A materials genome approach to accelerat- ing materials innovation. APL Mater. 1(1), 011002 (2013)

    2013

  7. [15]

    Schmidt, J., Cerqueira, T.F., Romero, A.H., Loew, A., J¨ ager, F., Wang, H.-C., Botti, S., Marques, M.A.: Improving machine-learning models in materials science through large datasets. Mater. Today Phys. 48, 101560 (2024)

    2024

  8. [16]

    Preprint at arXiv:2410.12771 (2024)

    Barroso-Luque, L., Shuaibi, M., Fu, X., Wood, B.M., Dzamba, M., Gao, M., Rizvi, A., Zitnick, C.L., Ulissi, Z.W.: Open materials 2024 (OMat24) inorganic materials dataset and models. Preprint at arXiv:2410.12771 (2024)

    2024 arXiv

  9. [17]

    Park, Y., Kim, J., Hwang, S., Han, S.: Scalable parallel algorithm for graph neural network interatomic potentials in molecu- lar dynamics simulations. J. Chem. Theory Comput. 20(11), 4857–4868 (2024)

    2024

  10. [18]

    Deng, B., Zhong, P., Jun, K., Riebesell, J., Han, K., Bartel, C.J., Ceder, G.: CHGNet as a pretrained universal neural network poten- tial for charge-informed atomistic modelling. Nat. Mach. Intell. 5(9), 1031–1041 (2023)

    2023

  11. [19]

    Preprint at arXiv:2401.00096 (2024)

    Batatia, I., Benner, P., Chiang, Y., Elena, A.M., Kov´ acs, D.P., Riebesell, J., Advin- cula, X.R., Asta, M., Avaylon, M., Baldwin, W.J., Berger, F., Bernstein, N., Bhowmik, A., Blau, S.M., C˘ arare, V., Darby, J.P., De, S., Pia, F.D., Deringer, V.L., Elijoˇ sius, R., El-Macha...

    2024 arXiv

  12. [20]

    Chen, C., Ong, S.P.: A universal graph deep learning interatomic potential for the peri- odic table. Nat. Comput. Sci. 2(11), 718–728 (2022)

    2022

  13. [21]

    Preprint at arxiv:2502.12147 (2025)

    Fu, X., Wood, B.M., Barroso-Luque, L., Levine, D.S., Gao, M., Dzamba, M., Zitnick, C.L.: Learning smooth and expressive inter- atomic potentials for physical property pre- diction. Preprint at arxiv:2502.12147 (2025)

    2025 arXiv

  14. [22]

    Bochkarev, A., Lysogorskiy, Y., Drautz, R.: Graph atomic cluster expansion for semilo- cal interactions beyond equivariant message passing. Phys. Rev. X 14(2), 021036 (2024)

    2024

  15. [23]

    Kang, S.: How graph neural network inter- atomic potentials extrapolate: Role of the message-passing algorithm. J. Chem. Phys. 161(24) (2024)

    2024

  16. [24]

    Mod- ell

    Zhang, Y.-W., Sorkin, V., Aitken, Z.H., Poli- tano, A., Behler, J., Thompson, A.P., Ko, T.W., Ong, S.P., Chalykh, O., Korogod, D., Podryabinkin, E., Shapeev, A., Li, J., Mishin, 22 Y., Pei, Z., Liu, X., Kim, J., Park, Y., Hwang, S., Han, S., Sheriff, K., Cao, Y., Freitas, R.: ...

    2025

  17. [25]

    Preprint at arXiv:2308.14920 (2023)

    Riebesell, J., Goodall, R.E., Benner, P., Chi- ang, Y., Deng, B., Lee, A.A., Jain, A., Pers- son, K.A.: Matbench discovery–a framework to evaluate machine learning crystal stabil- ity predictions. Preprint at arXiv:2308.14920 (2023)

    2023 arXiv

  18. [26]

    Ju, S., You, J., Kim, G., Park, Y., An, H., Han, S.: Application of pretrained uni- versal machine-learning interatomic potential for physicochemical simulation of liquid elec- trolytes in Li-ion batteries. Digit. Discov. 4(6), 1544–1559 (2025)

    2025

  19. [27]

    npj Comput

    Deng, B., Choi, Y., Zhong, P., Riebesell, J., Anand, S., Li, Z., Jun, K., Persson, K.A., Ceder, G.: Systematic softening in universal machine learning interatomic potentials. npj Comput. Mater. 11(1), 1–9 (2025)

    2025

  20. [28]

    Freitas, L.P., Schleder, G.R.: Performance assessment of universal machine learning interatomic potentials: Challenges and directions for materials’ surfaces

    Focassio, B., M. Freitas, L.P., Schleder, G.R.: Performance assessment of universal machine learning interatomic potentials: Challenges and directions for materials’ surfaces. ACS Appl. Mater. Interfaces (2024)

    2024

  21. [29]

    Pitfield, J., Brix, F., Tang, Z., Slaven- sky, A.M., Rønne, N., Christiansen, M.- P.V., Hammer, B.: Augmentation of universal potentials for broad applications. Phys. Rev. Lett. 134(5), 056201 (2025)

    2025

  22. [30]

    Preprint at arXiv:2502.15582 (2025)

    Radova, M., Stark, W.G., Allen, C.S., Mau- rer, R.J., Bart´ ok, A.P.: Fine-tuning founda- tion models of materials interatomic poten- tials with frozen transfer learning. Preprint at arXiv:2502.15582 (2025)

    2025 arXiv

  23. [31]

    Faraday Discuss

    Kaur, H., Della Pia, F., Batatia, I., Advin- cula, X.R., Shi, B.X., Lan, J., Cs´ anyi, G., Michaelides, A., Kapil, V.: Data-efficient fine-tuning of foundational models for first- principles quality sublimation enthalpies. Faraday Discuss. 256, 120–138 (2025)

    2025

  24. [32]

    Clausen, C.M., Rossmeisl, J., Ulissi, Z.W.: Adapting OC20-trained EquiformerV2 mod- els for high-entropy materials. J. Phys. Chem. C 128(27), 11190–11195 (2024)

    2024

  25. [33]

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwi´ nska, A., Hassabis, D., Clopath, C., Kumaran, D., Hadsell, R.: Overcoming catastrophic forgetting in neu- ral networks. Proc. Natl. Acad. Sci....

    2017

  26. [34]

    Zenke, F., Poole, B., Ganguli, S.: Contin- ual learning through synaptic intelligence. In Proc. 34th International Conference on Machine Learning 3987–3995 (PMLR, 2017)

    2017

  27. [35]

    Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for contin- ual learning. Adv. Neural Inf. Process. Syst. 32 (2019)

    2019

  28. [36]

    Nature 518(7540), 529–533 (2015)

    Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostro- vski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wier- stra, D., Legg, S., Hassabis, D.: Human-level contro...

    2015

  29. [37]

    Preprint at arXiv:1606.04671 (2016)

    Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neu- ral networks. Preprint at arXiv:1606.04671 (2016)

    2016 arXiv

  30. [38]

    Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Life- long learning with dynamically expandable networks. In Proc. 6th International Con- ference on Learning Representations (ICLR, 2018)

    2018

  31. [39]

    Preprint at arXiv:2406.18263 (2024) 23

    Wang, R., Guo, M., Gao, Y., Wang, X., Zhang, Y., Deng, B., Chen, X., Shi, M., Zhang, L., Zhong, Z.: A pre-trained deep potential model for sulfide solid electrolytes with broad coverage and high accuracy. Preprint at arXiv:2406.18263 (2024) 23

    2024 arXiv

  32. [40]

    Zhao, Q., Stalin, S., Zhao, C.-Z., Archer, L.A.: Designing solid-state electrolytes for safe, energy-dense batteries. Nat. Rev. Mater. 5(3), 229–252 (2020)

    2020

  33. [41]

    Bachman, J.C., Muy, S., Grimaud, A., Chang, H.-H., Pour, N., Lux, S.F., Paschos, O., Maglia, F., Lupart, S., Lamp, P., Gior- dano, L., Shao-Horn, Y.: Inorganic solid-state electrolytes for lithium batteries: mechanisms and properties governing ion conduction. Chem. Rev. 116(1)...

    2016

  34. [42]

    Manthiram, A., Yu, X., Wang, S.: Lithium battery chemistries enabled by solid-state electrolytes. Nat. Rev. Mater. 2(4), 1–16 (2017)

    2017

  35. [43]

    Nature 632(8026), 768–774 (2024)

    Dohare, S., Hernandez-Garcia, J.F., Lan, Q., Rahman, P., Mahmood, A.R., Sutton, R.S.: Loss of plasticity in deep continual learning. Nature 632(8026), 768–774 (2024)

    2024

  36. [44]

    https://github.com/MDIL-SNU/ SevenNet

    MDIL-SNU: SevenNet GitHub Repository (2024). https://github.com/MDIL-SNU/ SevenNet

    2024

  37. [45]

    Zhou, L., Minafra, N., Zeier, W.G., Nazar, L.F.: Innovative approaches to Li-argyrodite solid electrolytes for all-solid-state lithium batteries. Acc. Chem. Res. 54(12), 2717–2728 (2021)

    2021

  38. [46]

    ACS Appl

    Yu, C., Ganapathy, S., Hageman, J., Van Eijck, L., Van Eck, E.R., Zhang, L., Schwietert, T., Basak, S., Kelder, E.M., Wagemaker, M.: Facile synthesis toward the optimal structure-conductivity characteris- tics of the argyrodite Li 6PS5Cl solid-state electrolyte. ACS Appl. Mate...

    2018

  39. [47]

    Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C., Chen, W., Jing, Y., Zhao, W., Wang, X., Liu, Z., Zheng, H., Chen, J., Liu, Y., Tang, J., Li, J., Sun, M.: Parameter-efficient fine-tuning of large-scale pre-trained language models. Nat. Mach. ...

    2023

  40. [48]

    An, H., Kim, J., Kim, G., Han, S.: Atom- istic simulation of HF diffusion on ammonium fluorosilicate surface using neural network potential. Modell. Simul. Mater. Sci. Eng. (2025)

    2025

  41. [49]

    Li, H., Chaudhari, P., Yang, H., Lam, M., Ravichandran, A., Bhotika, R., Soatto, S.: Rethinking the hyperparameters for fine- tuning. In Proc. 8th International Conference on Learning Representations (ICLR, 2020)

    2020

  42. [50]

    In European Conference on Computer Vision , 491–507 (2020)

    Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., Houlsby, N.: Big transfer (BiT): General visual repre- sentation learning. In European Conference on Computer Vision , 491–507 (2020)

    2020

  43. [51]

    Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Gontijo-Lopes, R., Hajishirzi, H., Farhadi, A., Namkoong, H., Schmidt, L.: Robust fine-tuning of zero- shot models. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition 7959–7971 (IEEE, 2022)

    2022

  44. [52]

    Murugan, R., Thangadurai, V., Weppner, W.: Fast lithium ion conduction in garnet- type Li 7La3Zr2O12. Angew. Chem. Int. Ed. 46(41), 7778 (2007)

    2007

  45. [53]

    npj Comput

    He, X., Zhu, Y., Epstein, A., Mo, Y.: Statis- tical variances of diffusional properties from ab initio molecular dynamics simulations. npj Comput. Mater. 4(1), 18 (2018)

    2018

  46. [54]

    Kamaya, N., Homma, K., Yamakawa, Y., Hirayama, M., Kanno, R., Yonemura, M., Kamiyama, T., Kato, Y., Hama, S., Kawamoto, K., Mitsui, A.: A lithium superi- onic conductor. Nat. Mater. 10(9), 682–686 (2011)

    2011

  47. [55]

    Energy Environ

    Seino, Y., Ota, T., Takada, K., Hayashi, A., Tatsumisago, M.: A sulphide lithium super ion conductor is superior to liquid ion conduc- tors for use in rechargeable batteries. Energy Environ. Sci. 7(2), 627–631 (2014)

    2014

  48. [56]

    Asano, T., Sakai, A., Ouchi, S., Sakaida, M., Miyazaki, A., Hasegawa, S.: Solid halide electrolytes with high lithium-ion conductiv- ity for application in 4 V class bulk-type all-solid-state batteries. Adv. Mater. 30(44), 24 1803075 (2018)

    2018

  49. [57]

    Energy Environ

    Li, W., Wu, G., Ara´ ujo, C.M., Scheicher, R.H., Blomqvist, A., Ahuja, R., Xiong, Z., Feng, Y., Chen, P.: Li+ ion conductivity and diffusion mechanism in α-Li3N and β- Li3N. Energy Environ. Sci. 3(10), 1524–1530 (2010)

    2010

  50. [58]

    Miara, L.J., Suzuki, N., Richards, W.D., Wang, Y., Kim, J.C., Ceder, G.: Li-ion con- ductivity in Li9S3N. J. Mater. Chem. A3(40), 20338–20344 (2015)

    2015

  51. [59]

    Matsuo, M., Orimo, S.-i.: Lithium fast-ionic conduction in complex hydrides: review and prospects. Adv. Energy Mater. 1(2), 161–172 (2011)

    2011

  52. [60]

    Yoo, D., Lee, K., Jeong, W., Lee, D., Watan- abe, S., Han, S.: Atomic energy mapping of neural network potential. Phys. Rev. Mater. 3(9), 093802 (2019)

    2019

  53. [61]

    Kraft, M.A., Culver, S.P., Calderon, M., B”ocher, F., Krauskopf, T., Senyshyn, A., Dietrich, C., Zevalkink, A., Janek, J., Zeier, W.G.: Influence of lattice polarizability on the ionic conductivity in the lithium supe- rionic argyrodites Li 6PS5X (X= Cl, Br, I). J. Am. Chem. S...

    2017

  54. [62]

    Evans, D.J., Holian, B.L.: The nos` e–hoover thermostat. J. Chem. Phys. 83(8), 4069–4074 (1985)

    1985

  55. [63]

    Kresse, G., Joubert, D.: From ultrasoft pseu- dopotentials to the projector augmented- wave method. Phys. Rev. B 59, 1758–1775 (1999)

    1999

  56. [64]

    Bl¨ ochl, P.E.: Projector augmented-wave method. Phys. Rev. B 50(24), 17953–17979 (1994)

    1994

  57. [65]

    Perdew, J.P., Burke, K., Ernzerhof, M.: Gen- eralized gradient approximation made simple. Phys. Rev. Lett. 77(18), 3865 (1996)

    1996

  58. [66]

    Larsen, A.H., Mortensen, J.J., Blomqvist, J., Castelli, I.E., Christensen, R., Du lak, M., Friis, J., Groves, M.N., Hammer, B., Har- gus, C., Hermes, E.D., Jennings, P.C., Jensen, P.B., Kermode, J., Kitchin, J.R., Kolsbjerg, E.L., Kubal, J., Kaasbjerg, K., Lysgaard, S., Marons...

    2017

  59. [67]

    Com- put

    Thompson, A.P., Aktulga, H.M., Berger, R., Bolintineanu, D.S., Brown, W.M., Crozier, P.S., Veld, P.J.i.t., Kohlmeyer, A., Moore, S.G., Nguyen, T.D., Shan, R., Stevens, M.J., Tranchida, J., Trott, C., Plimpton, S.J.: Lammps–a flexible simulation tool for particle-based material...

    2022

  60. [68]

    Hogrefe, K., Minafra, N., Hanghofer, I., Banik, A., Zeier, W.G., Wilkening, H.M.R.: Opening diffusion pathways through site disorder: The interplay of local struc- ture and ion dynamics in the solid elec- trolyte Li 6+xP1–xGexS5I as probed by neu- tron diffraction and NMR. J. ...

    2022

  61. [69]

    Friauf, R.J.: Correlation effects for diffusion in ionic crystals. J. Appl. Phys. 33(1), 494– 505 (1962) 25

    1962

Pith tools

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