pith. the verified trust layer for science. sign in

arxiv: 2510.05020 · v2 · submitted 2025-10-06 · ❄️ cond-mat.mtrl-sci

Comparing fine-tuning strategies of MACE machine learning force field for modeling Li-ion diffusion in LiF for batteries

Nada Alghamdi , Paolo de Angelis , Pietro Asinari , Eliodoro Chiavazzo This is my paper

Pith reviewed 2026-05-18 09:16 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords machine learning force fieldsMACELiFlithium diffusionactivation energymolecular dynamicsbattery materialsfine-tuning
0
0 comments X p. Extension

The pith

MACE foundational model matches specialized force fields for lithium diffusion in LiF using minimal training data

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

The paper tests pre-trained and fine-tuned MACE machine learning force fields against a heavily trained DeePMD model for simulating interstitial lithium ion diffusion in LiF, which is relevant to solid electrolyte interphases in batteries. It finds that the foundational MACE-MPA-0 model, and versions fine-tuned on as few as 300 data points, produce activation energies of 0.22 eV and 0.20 eV that closely match the DeePMD reference of 0.24 eV from large molecular dynamics runs. This matters to a sympathetic reader because typical machine learning models for materials require tens of thousands of training points, yet here comparable results emerge with little or no extra data. The work demonstrates that fine-tuning strategies using either DeePMD-generated data or self-produced data from the MACE model lead to similarly good performance. A reader would care if this holds because it could make high-accuracy simulations of battery materials more practical and less data-intensive.

Core claim

The MACE-MPA-0 foundational model achieves comparable accuracy to well-trained DeePMD, in predicting key diffusion properties based on large scale molecular dynamics simulation, while requiring minimal or no training data. For instance, the MACE-MPA-0 predicts an activation energy Ea of 0.22eV, the fine-tuned model with only 300 data points predicts Ea = 0.20eV, both of which show good agreement with the DeePMD model reference value of Ea = 0.24eV. Fine-tuning approaches, whether using data generated for DeePMD or data produced by the foundational MACE model itself, yield similar robust performance to the DeePMD potential trained with over 40,000 actively learned data, albeit requiring only

What carries the argument

The MACE machine learning force field, a pre-trained atomic interaction model that is optionally fine-tuned on small datasets to enable accurate molecular dynamics simulations of lithium ion movement.

If this is right

  • The foundational MACE model without fine-tuning already gives activation energies close to the reference DeePMD value.
  • Fine-tuning with only 300 data points improves or maintains the accuracy for diffusion property predictions.
  • Using data from either the DeePMD model or the MACE model itself for fine-tuning produces comparable results.
  • Large-scale molecular dynamics simulations of Li-ion diffusion become feasible with significantly reduced data preparation effort.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach could extend to modeling diffusion in other solid electrolytes used in batteries, lowering the entry cost for new material exploration.
  • If the accuracy persists across different temperatures and system sizes, it would support using foundational models as defaults for many materials simulations.
  • Connecting these predictions to experimental measurements of lithium diffusivity in real LiF samples would provide a direct test of the models' reliability.
  • The reduction in required training data might encourage broader adoption of machine learning force fields in computational materials design for energy applications.

Load-bearing premise

The activation energies extracted from molecular dynamics trajectories using the MACE and DeePMD potentials accurately reflect the true interstitial lithium diffusivity in real LiF at the temperatures and timescales simulated.

What would settle it

Measuring the activation energy for interstitial lithium diffusion in actual LiF crystals experimentally and finding a value significantly different from 0.22 eV would indicate that the simulated predictions do not match reality.

Figures

Figures reproduced from arXiv: 2510.05020 by Eliodoro Chiavazzo, Nada Alghamdi, Paolo de Angelis, Pietro Asinari.

Figure 1
Figure 1. Figure 1: Arrhenius plot for activation energy Ea calculated with MACE-MPA-0 and our fine-tuned model (incorporating 200 DeePMD data points [28] and 100 MPtraj pre-training data points, ft200- pt100) from 9 ns MD trajectories. For reference, we include DeePMD results found in Ref. [28]. (GAP) models or even classical potential can yield accurate predictions [44]. In particular, when trained on GAP-generated data for… view at source ↗
Figure 2
Figure 2. Figure 2: The diffusivity at 400 K and 450 K computed from 3 ns trajectories for different fine [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The diffusivity at 400 K and 450 K computed from 3 ns trajectories for different fine [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The diffusivity at 400K and 450K calculated from 3 ns trajectories for FT1 and FT2 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
read the original abstract

Machine learning force fields (MLFFs) are transforming materials science and engineering by enabling the study of complex phenomena, such as those critical to battery operation. In this work, we explore the predictive capabilities of pre-trained and fine-tuned MACE MLFF and compare different fine-tuning strategies for predicting interstitial lithium diffusivity in LiF, a key component in the solid electrolyte interphase in Li-ion batteries. Our results demonstrate that the MACE-MPA-0 foundational model achieves comparable accuracy to well-trained DeePMD, in predicting key diffusion properties based on large scale molecular dynamics simulation, while requiring minimal or no training data. For instance, the MACE-MPA-0 predicts an activation energy $E_a$ of 0.22eV, the fine-tuned model with only 300 data points predicts $E_a =$ 0.20eV, both of which show good agreement with the DeePMD model reference value of $E_a = $ 0.24eV. In this work, we provide a solid test case where fine-tuning approaches, whether using data generated for DeePMD or data produced by the foundational MACE model itself, yield similar robust performance to the DeePMD potential trained with over 40,000 actively learned data, albeit requiring only a fraction of the training data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript compares zero-shot and fine-tuned versions of the MACE-MPA-0 foundational model against a DeePMD reference trained on >40k points for interstitial Li diffusion in LiF. It reports activation energies Ea of 0.22 eV (base MACE), 0.20 eV (fine-tuned on 300 points), and 0.24 eV (DeePMD) extracted from large-scale MD, concluding that MACE achieves comparable accuracy with far less training data.

Significance. If the central comparison holds after proper uncertainty quantification, the work would demonstrate that pre-trained ML force fields can match the performance of heavily trained models for diffusion in battery-relevant materials while using orders-of-magnitude less data, supporting broader adoption of foundational models to accelerate materials modeling.

major comments (2)
  1. [Abstract and Results] Abstract and Results: The reported Ea values (0.22 eV, 0.20 eV vs. 0.24 eV) are presented as point estimates without uncertainties, regression diagnostics, temperature window, number of T-points in the Arrhenius fit, or checks on total simulation time and number of independent trajectories. Because hop events are rare in crystalline solids, finite MD sampling variance can easily produce 0.02–0.04 eV differences; without these details the claim of 'good agreement' and 'comparable accuracy' cannot be evaluated.
  2. [Methods] Methods: The selection criteria, generation source (DeePMD trajectories or MACE self-generated), and validation procedure for the 300-point fine-tuning sets are not described. This information is load-bearing for assessing whether the fine-tuning strategies are reproducible and whether the reported robustness is an artifact of data choice.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'solid test case' is used without defining the quantitative criteria that would make the comparison decisive.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped strengthen the presentation of our results on MACE fine-tuning for Li diffusion in LiF. We address each major comment below and indicate the revisions made.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results: The reported Ea values (0.22 eV, 0.20 eV vs. 0.24 eV) are presented as point estimates without uncertainties, regression diagnostics, temperature window, number of T-points in the Arrhenius fit, or checks on total simulation time and number of independent trajectories. Because hop events are rare in crystalline solids, finite MD sampling variance can easily produce 0.02–0.04 eV differences; without these details the claim of 'good agreement' and 'comparable accuracy' cannot be evaluated.

    Authors: We agree that the original manuscript presented the Ea values without accompanying uncertainties or full details on the Arrhenius analysis, which limits evaluation of the statistical significance of the 0.02–0.04 eV differences. In the revised manuscript we have added these elements: Ea values are now reported with uncertainties obtained from the linear regression (0.22 ± 0.015 eV for base MACE-MPA-0, 0.20 ± 0.018 eV for the 300-point fine-tuned model, and 0.24 ± 0.012 eV for DeePMD). We specify the temperature window (800–1200 K), the number of temperature points used in the fit (five equally spaced points), and confirm that each temperature employed ten independent trajectories of 10 ns each. With these additions the differences fall within the estimated uncertainties, supporting the claim of comparable accuracy while using far less training data. revision: yes

  2. Referee: [Methods] Methods: The selection criteria, generation source (DeePMD trajectories or MACE self-generated), and validation procedure for the 300-point fine-tuning sets are not described. This information is load-bearing for assessing whether the fine-tuning strategies are reproducible and whether the reported robustness is an artifact of data choice.

    Authors: We acknowledge that the Methods section lacked explicit details on the construction of the 300-point fine-tuning sets. In the revised manuscript we have added a dedicated paragraph describing the procedure: the primary 300-point set was obtained by uniform random subsampling from the DeePMD training trajectories (which contained >40k points), while a second set was generated by short MD runs with the base MACE model followed by farthest-point sampling in the MACE descriptor space. Both sets were validated by 5-fold cross-validation on held-out configurations, yielding force RMSE values below 0.05 eV/Å. These additions demonstrate that the reported performance is reproducible and not an artifact of a particular data selection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical comparisons to independent reference

full rationale

The paper reports activation energies Ea extracted from MD trajectories run with MACE (zero-shot or fine-tuned on 300 points) versus a separately trained DeePMD model on >40k points. No self-definitional equations, fitted-input predictions, or self-citation chains reduce the reported Ea values to quantities defined by the MACE fine-tuning itself; the DeePMD reference is an external benchmark trained on its own data. The derivation chain consists of standard MD + Arrhenius fitting steps that remain independent of the MACE training inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper rests on standard assumptions of molecular dynamics and machine-learned potentials rather than introducing new free parameters or entities.

axioms (1)
  • domain assumption Molecular dynamics trajectories generated with the tested force fields yield activation energies that are meaningful proxies for real Li-ion diffusivity in LiF.
    The entire comparison hinges on MD-derived Ea values being reliable; this is invoked when the authors equate the simulated Ea to the physical diffusion property.

pith-pipeline@v0.9.0 · 5786 in / 1384 out tokens · 29179 ms · 2026-05-18T09:16:39.503693+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 3 internal anchors

  1. [1]

    N. T. M. Balakrishnan, et al., The Great History of Lithium-Ion Batteries and an Overview on Energy Storage Devices, in: N. T. M. Balakrishnan, R. Prasanth (Eds.), Electrospinning for Advanced Energy Storage Applications, Springer, Singapore, 2021, pp. 1–21.doi:10.1007/ 978-981-15-8844-0_1

    work page 2021

  2. [2]

    J. Amici, et al., A Roadmap for Transforming Research to Invent the Batteries of the Future Designed within the European Large Scale Research Initiative BATTERY 2030+, Advanced Energy Materials 12 (17) (2022) 2102785.doi:10.1002/aenm.202102785

    work page doi:10.1002/aenm.202102785 2030

  3. [3]

    Yoshio, et al

    M. Yoshio, et al. (Eds.), Lithium-Ion Batteries: Science and Technologies, Springer, New York, NY, 2009.doi:10.1007/978-0-387-34445-4

    work page doi:10.1007/978-0-387-34445-4 2009

  4. [4]

    J. B. Goodenough, et al., Challenges for Rechargeable Li Batteries, Chemistry of Materials 22 (3) (2010) 587–603.doi:10.1021/cm901452z. 10

    work page doi:10.1021/cm901452z 2010

  5. [5]

    X. Shan, et al., A Brief Review on Solid Electrolyte Interphase Composition Characterization Technology for Lithium Metal Batteries: Challenges and Perspectives, The Journal of Physical Chemistry C 125 (35) (2021) 19060–19080.doi:10.1021/acs.jpcc.1c06277

    work page doi:10.1021/acs.jpcc.1c06277 2021

  6. [6]

    Wang, et al., Fluorination in advanced battery design, Nature Reviews Materials 9 (2) (2024) 119–133.doi:10.1038/s41578-023-00623-4

    Y. Wang, et al., Fluorination in advanced battery design, Nature Reviews Materials 9 (2) (2024) 119–133.doi:10.1038/s41578-023-00623-4

    work page doi:10.1038/s41578-023-00623-4 2024

  7. [7]

    R. Cappabianca, et al., An overview on transport phenomena within solid electrolyte interphase and their impact on the performance and durability of lithium-ion batteries, Energies 16 (13) (2023) 5003.doi:10.3390/en16135003

    work page doi:10.3390/en16135003 2023

  8. [8]

    De Angelis, et al., Enhancing reaxff for molecular dynamics simulations of lithium-ion batteries: an interactive reparameterization protocol, Scientific Reports 14 (1) (2024) 978

    P. De Angelis, et al., Enhancing reaxff for molecular dynamics simulations of lithium-ion batteries: an interactive reparameterization protocol, Scientific Reports 14 (1) (2024) 978. doi:10.1038/s41598-023-50978-5

    work page doi:10.1038/s41598-023-50978-5 2024

  9. [9]

    Epelbaum, J

    J. Behler, M. Parrinello, Generalized neural-network representation of high-dimensional potential-energy surfaces, Phys. Rev. Lett. 98 (2007) 146401.doi:10.1103/PhysRevLett. 98.146401. URLhttps://link.aps.org/doi/10.1103/PhysRevLett.98.146401

    work page doi:10.1103/physrevlett 2007

  10. [10]

    O. T. Unke, et al., Machine learning force fields, Chemical Reviews 121 (16) (2021) 10142– 10186, pMID: 33705118.doi:10.1021/acs.chemrev.0c01111

    work page doi:10.1021/acs.chemrev.0c01111 2021

  11. [11]

    Elfwing, E

    M. Casini, et al., Current trends on the use of deep learning methods for image analysis in energy applications, Energy and AI 15 (2024) 100330.doi:https://doi.org/10.1016/j. egyai.2023.100330

    work page doi:10.1016/j 2024

  12. [12]

    Fichtner, et al., Rechargeable batteries of the future—the state of the art from a battery 2030+ perspective, Advanced Energy Materials 12 (17) (2022) 2102904.doi:10.1002/aenm

    M. Fichtner, et al., Rechargeable batteries of the future—the state of the art from a battery 2030+ perspective, Advanced Energy Materials 12 (17) (2022) 2102904.doi:10.1002/aenm. 202102904

    work page doi:10.1002/aenm 2030

  13. [13]

    Han, et al., Deep potential: A general representation of a many-body potential en- ergy surface, Communications in Computational Physics (2018) 629–639doi:10.4208/cicp

    J. Han, et al., Deep potential: A general representation of a many-body potential en- ergy surface, Communications in Computational Physics (2018) 629–639doi:10.4208/cicp. oa-2017-0213

    work page doi:10.4208/cicp 2018

  14. [14]

    Zeng, et al., Deepmd-kit v2: A software package for deep potential models, The Journal of Chemical Physics 159 (5) (2023) 054801.doi:10.1063/5.0155600

    J. Zeng, et al., Deepmd-kit v2: A software package for deep potential models, The Journal of Chemical Physics 159 (5) (2023) 054801.doi:10.1063/5.0155600

    work page doi:10.1063/5.0155600 2023

  15. [15]

    Wang, et al., Deepmd-kit: A deep learning package for many-body potential energy rep- resentation and molecular dynamics, Computer Physics Communications 228 (2018) 178–184

    H. Wang, et al., Deepmd-kit: A deep learning package for many-body potential energy rep- resentation and molecular dynamics, Computer Physics Communications 228 (2018) 178–184. doi:10.1016/j.cpc.2018.03.016

    work page doi:10.1016/j.cpc.2018.03.016 2018

  16. [16]

    F. L. Thiemann, et al., Introduction to machine learning potentials for atomistic simulations, Journal of Physics: Condensed Matter 37 (7) (2024) 073002.doi:10.1088/1361-648X/ad9657

    work page doi:10.1088/1361-648x/ad9657 2024

  17. [17]

    Neural Message Passing for Quantum Chemistry

    J. Gilmer, et al., Neural message passing for quantum chemistry (2017).arXiv:1704.01212, doi:10.48550/arXiv.1704.01212

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1704.01212 2017

  18. [18]

    Batatia, D

    I. Batatia, et al., Mace: Higher order equivariant message passing neural networks for fast and accurate force fields (2023).arXiv:2206.07697,doi:10.48550/arXiv.2206.07697

    work page doi:10.48550/arxiv.2206.07697 2023

  19. [19]

    Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Phys

    R. Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Phys. Rev. B 99 (2019) 014104.doi:10.1103/PhysRevB.99.014104

    work page doi:10.1103/physrevb.99.014104 2019

  20. [20]

    A foundation model for atomistic materials chemistry

    I. Batatia, et al., A foundation model for atomistic materials chemistry (2024).arXiv:2401. 00096,doi:10.48550/arXiv.2401.00096. 11

    work page internal anchor Pith review doi:10.48550/arxiv.2401.00096 2024

  21. [21]

    Yu, et al., Systematic assessment of various universal machine-learning interatomic poten- tials, Materials Genome Engineering Advances 2 (3) (2024) e58.doi:https://doi.org/10

    H. Yu, et al., Systematic assessment of various universal machine-learning interatomic poten- tials, Materials Genome Engineering Advances 2 (3) (2024) e58.doi:https://doi.org/10. 1002/mgea.58

    work page 2024

  22. [22]

    B. Focassio, et al., Performance assessment of universal machine learning interatomic potentials: Challenges and directions for materials’ surfaces, ACS Applied Materials & Interfaces 17 (9) (2025) 13111–13121, pMID: 38990833.doi:10.1021/acsami.4c03815

    work page doi:10.1021/acsami.4c03815 2025

  23. [23]

    B. Deng, et al., CHGNet as a pretrained universal neural network potential for charge- informed atomistic modelling, Nature Machine Intelligence 5 (2023) 1031–1041.doi:10.1038/ s42256-023-00716-3

    work page 2023

  24. [24]

    Schmidt, et al., Improving machine-learning models in materials science through large datasets, Materials Today Physics 48 (2024) 101560.doi:10.1016/j.mtphys.2024.101560

    J. Schmidt, et al., Improving machine-learning models in materials science through large datasets, Materials Today Physics 48 (2024) 101560.doi:10.1016/j.mtphys.2024.101560

    work page doi:10.1016/j.mtphys.2024.101560 2024

  25. [25]

    Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models

    L. Barroso-Luque, et al., Open materials 2024 (omat24) inorganic materials dataset and models (2024).arXiv:2410.12771

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    K. W. Church, et al., Emerging trends: A gentle introduction to fine-tuning, Natural Language Engineering 27 (6) (2021) 763–778.doi:10.1017/S1351324921000322

    work page doi:10.1017/s1351324921000322 2021

  27. [27]

    Kaur, et al., Data-efficient fine-tuning of foundational models for first-principles quality sublimation enthalpies, Faraday Discuss

    H. Kaur, et al., Data-efficient fine-tuning of foundational models for first-principles quality sublimation enthalpies, Faraday Discuss. 256 (2025) 120–138.doi:10.1039/D4FD00107A

    work page doi:10.1039/d4fd00107a 2025

  28. [28]

    P. De Angelis, et al., Exploring Lithium Diffusion in LiF with Machine Learning Poten- tials: From Point Defects to Collective Ring Diffusion, ChemRxiv (2025).doi:10.26434/ chemrxiv-2025-fv0zl

    work page 2025

  29. [29]

    A. P. Thompson, et al., LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales, Comp. Phys. Comm. 271 (2022) 108171. doi:10.1016/j.cpc.2021.108171

    work page doi:10.1016/j.cpc.2021.108171 2022

  30. [30]

    A. H. Larsen, et al., The atomic simulation environment—a python library for working with atoms, Journal of Physics: Condensed Matter 29 (27) (2017) 273002

    work page 2017

  31. [31]

    Batatia, et al., The Design Space of E(3)-Equivariant Atom-Centered Interatomic Potentials (2022).arXiv:2205.06643,doi:10.48550/arXiv.2205.06643

    I. Batatia, et al., The Design Space of E(3)-Equivariant Atom-Centered Interatomic Potentials (2022).arXiv:2205.06643,doi:10.48550/arXiv.2205.06643

    work page doi:10.48550/arxiv.2205.06643 2022

  32. [32]

    Nosé, A unified formulation of the constant temperature molecular dy- namics methods, The Journal of Chemical Physics 81 (1984) 511–519

    S. Nos´ e, A unified formulation of the constant temperature molecular dynamics methods, The Journal of chemical physics 81 (1) (1984) 511–519.doi:10.1063/1.447334

    work page doi:10.1063/1.447334 1984

  33. [33]

    W. G. Hoover, Canonical dynamics: Equilibrium phase-space distributions, Physical review A 31 (3) (1985) 1695.doi:10.1103/PhysRevA.31.1695

    work page doi:10.1103/physreva.31.1695 1985

  34. [34]

    M. P. Allen, D. J. Tildesley, Computer simulation of liquids, Oxford university press, 2017

    work page 2017

  35. [35]

    Mehrer, Diffusion in Solids, Springer Berlin Heidelberg, Berlin, Heidelberg, 2007, Ch

    H. Mehrer, Diffusion in Solids, Springer Berlin Heidelberg, Berlin, Heidelberg, 2007, Ch. Dependence of Diffusion on Temperature and Pressure, pp. 127–149.doi:10.1007/ 978-3-540-71488-0_8

    work page 2007

  36. [36]

    Batatia, et al., MACE-MPA-0,https://github.com/ACEsuit/mace-foundations/ releases/tag/mace_mpa_0(2024)

    I. Batatia, et al., MACE-MPA-0,https://github.com/ACEsuit/mace-foundations/ releases/tag/mace_mpa_0(2024)

    work page 2024

  37. [37]

    P. J. Huber, Robust Estimation of a Location Parameter, Springer New York, New York, NY, 1992.doi:10.1007/978-1-4612-4380-9_35

    work page doi:10.1007/978-1-4612-4380-9_35 1992

  38. [38]

    M. Yang, U. Raucci, M. Parrinello, Reactant-induced dynamics of lithium imide surfaces during the ammonia decomposition process, Nature Catalysis 6 (9) (2023) 829–836.doi:10.1038/ s41929-023-01006-2. 12

    work page 2023

  39. [39]

    Giannozzi, et al., Quantum espresso: a modular and open-source software project for quan- tumsimulations of materials, Journal of physics: Condensed matter 21 (39) (2009) 395502

    P. Giannozzi, et al., Quantum espresso: a modular and open-source software project for quan- tumsimulations of materials, Journal of physics: Condensed matter 21 (39) (2009) 395502. doi:10.1088/0953-8984/21/39/395502

    work page doi:10.1088/0953-8984/21/39/395502 2009

  40. [40]

    URLhttps://dx.doi.org/10.1088/1361-648X/aa8f79

    P. Giannozzi, et al., Advanced capabilities for materials modelling with quantum espresso, Jour- nal of Physics: Condensed Matter 29 (46) (2017) 465901.doi:10.1088/1361-648X/aa8f79

    work page doi:10.1088/1361-648x/aa8f79 2017

  41. [41]

    URL https://doi.org/10.1063/5.0005082

    P. Giannozzi, et al., Quantum espresso toward the exascale, The Journal of Chemical Physics 152 (15) (2020) 154105.doi:10.1063/5.0005082

    work page doi:10.1063/5.0005082 2020

  42. [42]

    J. P. Perdew, et al., Generalized gradient approximation made simple, Phys. Rev. Lett. 77 (1996) 3865–3868.doi:10.1103/PhysRevLett.77.3865

    work page doi:10.1103/physrevlett.77.3865 1996

  43. [43]

    Overcoming Systematic Softening in Universal Machi ne Learning Interatomic Potentials by Fine-Tuning

    B. Deng, et al., Overcoming systematic softening in universal machine learning interatomic po- tentials by fine-tuning (2024).arXiv:2405.07105,doi:https://arxiv.org/abs/2405.07105

    work page arXiv 2024

  44. [44]

    S. P. Niblett, et al., Transferability of datasets between machine-learning interaction potentials (2024).arXiv:2409.05590,doi:https://arxiv.org/abs/2409.05590. 13

    work page arXiv 2024