The reviewed record of science sign in
Pith

arxiv: 2607.05321 · v1 · pith:HRLRWPMM · submitted 2026-07-06 · cond-mat.mtrl-sci · physics.comp-ph

Data-driven atomistic modelling of hybrid halide perovskite passivation

Laura-Bianca Pa\c{s}ca , Henry J. Snaith , Volker L. Deringer This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-07 17:08 UTCglm-5.2pith:HRLRWPMMrecord.jsonopen to challenge →

classification cond-mat.mtrl-sci physics.comp-ph
keywords machine-learned interatomic potentialshybrid halide perovskitemolecular passivationamino-silanecontinual fine-tuningsurface coverageoctahedral tiltingsolar cells
0
0 comments X

The pith

More passivation molecules can break the perovskite they protect

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

This paper claims that a continual fine-tuning protocol—adapted from large language model training—can improve machine-learned interatomic potentials for simulating amino-silane molecules binding to hybrid halide perovskite surfaces, and uses that protocol to reveal a coverage-dependent structural transition. The central object is the CFT pipeline: a pre-trained foundation model is first adapted to a curated intermediate dataset of diverse perovskite structures (hyP-26), then further fine-tuned on a small specialised dataset of amino-silane passivation configurations. Applied to molecular dynamics simulations of increasing surface coverage, the resulting model shows that as amino-silane density rises, steric crowding forces molecules from multidentate (multi-point) binding into upright, nitrogen-only coordination; the molecules penetrate the surface, octahedral tilting intensifies, Pb–halide bonds elongate, and the perovskite lattice loses corner-sharing connectivity. This offers a microscopic mechanism for the experimentally observed decline in passivation effectiveness and crystallinity at excessive deposition times. The paper also shows that the advantage of the intermediate fine-tuning step depends on the starting model: when the pre-training data already covers organic and surface chemistry well, the additional benefit of the intermediate step is limited.

Core claim

The paper's central finding is twofold. First, a stepwise continual fine-tuning protocol (foundation model → hyP-26 intermediate dataset → specialised passivation dataset) improves force and energy predictions for perovskite–molecule systems when the starting model's pre-training data underrepresents organic and interfacial chemistry, but yields diminishing returns when the starting model is already trained on a diverse chemical space. Second, simulations using this protocol reveal that increasing amino-silane surface coverage drives a transition from multidentate binding to upright nitrogen-mediated coordination, accompanied by molecular penetration into the subsurface, increased octahedral

What carries the argument

Continual fine-tuning (CFT) pipeline: foundation model → hyP-26 intermediate dataset (D1) → AEAPTMS passivation dataset (D2). The hyP-26 dataset contains ~11,000 diverse metal and hybrid halide perovskite structures including distorted crystals, disordered phases, defects, surfaces, and random structures. The D2 dataset captures four coordination modes of the amino-silane AEAPTMS on a Cs0.12FA0.88PbI0.75Br0.25 surface. The coverage-dependent structural transition—from multidentate to upright binding with increasing molecular density—is the key physical mechanism.

Load-bearing premise

The claim that the CFT protocol improves model accuracy depends on the starting foundation model having limited coverage of organic and surface chemistry in its pre-training data. The paper itself shows that when using a model pre-trained on a more diverse dataset, the additional benefits of the intermediate fine-tuning step are limited.

What would settle it

If the coverage-dependent transition from multidentate to upright binding, and the associated lattice disruption, cannot be reproduced by direct DFT molecular dynamics at the same surface coverages, the central physical claim would be undermined. If the CFT protocol shows no accuracy advantage over direct fine-tuning for any starting model, the methodological claim would collapse.

Figures

Figures reproduced from arXiv: 2607.05321 by Henry J. Snaith, Laura-Bianca Pa\c{s}ca, Volker L. Deringer.

Figure 1
Figure 1. Figure 1: Fine-tuning pipelines for ML-driven simulations of hybrid perovskite passivation [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Influence of fitting protocol on model performance for general and specialised tasks. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Error cancellation in MLIP static adsorption energy predictions. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: MLIP-driven relaxation of the amino-silane molecule passivating the hybrid per [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The link between increased passivation coverage and structural disruption of [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
read the original abstract

Molecular passivation of surface defects is key to improving the optoelectronic performance of hybrid halide perovskite materials, but the underlying atomistic mechanisms are incompletely understood. While machine-learned interatomic potentials are now widely used to simulate complex molecular and crystalline systems, their application to experimentally-realistic scenarios - such as molecules coordinating to perovskite surfaces - is still far from trivial. Here, we describe a multistep training pipeline, resembling continuous fine-tuning used for large language models, to underpin atomistic modelling and computational experiments in this domain. Our protocol involves two components: (i) a large, curated, and open dataset of diverse metal and hybrid halide perovskite structures ('hyP-26'); and (ii) a small, specialised dataset for an amino-silane molecule passivating the surface, providing highly specific information for fine-tuning. We apply this approach to explore collective behaviour at a mixed-composition halide perovskite surface passivated with a varying coverage of amino-silane molecules, revealing an evolution of interactions with increasing molecular surface coverage.

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 / 6 minor

Summary. This manuscript presents a continual fine-tuning (CFT) protocol for machine-learned interatomic potentials (MLIPs) applied to amino-silane passivation of hybrid halide perovskite surfaces. The authors introduce a curated dataset ('hyP-26') of diverse perovskite structures as an intermediate fine-tuning step, then further fine-tune on a specialised dataset (D2) of single AEAPTMS molecules on perovskite surfaces. The protocol is benchmarked across four MACE foundation models, with detailed error-cancellation analysis for adsorption energies. The resulting model is then applied to multi-molecule MD simulations (up to 40 molecules) to study coverage-dependent structural disruption of the perovskite lattice, finding a shift from multidentate to upright coordination and increased octahedral tilting with higher coverage.

Significance. The methodological framework is well-constructed and the error analysis is notably thorough. The paper ships an open dataset (hyP-26) with public GitHub repository, performs systematic comparison across four foundation models, provides a transparent error-cancellation waterfall analysis (Fig. 3b) that the authors themselves use to diagnose model limitations, and includes a LASPH sensitivity test (Table S8). The coverage-dependent structural predictions are falsifiable against experimental XRD trends. The CFT protocol and the hyP-26 dataset are useful contributions to the growing toolkit for MLIP fine-tuning in chemically heterogeneous systems.

major comments (2)
  1. §'Data-driven atomistic modelling of molecular passivation' (Fig. 5): The multi-molecule MD simulations (up to 40 AEAPTMS molecules) extrapolate well beyond the D2 training data, which contains only single-molecule passivation configurations (four coordination types of one AEAPTMS molecule on a surface). The steric crowding and intermolecular methoxy-group interactions that the paper invokes to explain the coverage-dependent shift from multidentate to upright coordination (Discussion, final paragraph) are precisely the interactions absent from the training data. All quantitative validation (Tables S7–S12, Figs. 3–4) is performed on single-molecule systems. No DFT cross-check is provided for any multi-molecule snapshot from the Fig. 5 simulations. This is load-bearing for the quantitative claims (91.5% valid octahedra, ~20° tilt angles, E_ads stabilisation at ~−1.5 eV). At minimum, DFT单点单
  2. §'Validation across general and specific tasks' (page 9) and Fig. 4c: The paper acknowledges that 'the additional benefits of CFT over conventional fine-tuning... are limited' when using MACE-MH-1, and Fig. 4c shows the FM→AEAPTMS model achieves a slightly higher average SOAP similarity score (0.975) than the CFT model (0.954). Yet the production simulations in Fig. 5 use the MACE-MH-1-derived CFT model. The justification for selecting the CFT model over the simpler FM→AEAPTMS model for the application simulations should be stated more explicitly, given that the paper's own benchmarks do not clearly favour it for this starting foundation model.
minor comments (6)
  1. Introduction vs. Methods: The molecule is named '(3-aminopropyl)trimethoxysilane (AEAPTMS)' in the Introduction but '[3-(2-aminoethylamino)propyl]trimethoxysilane (AEAPTMS)' in the Methods section. These are different molecules. Please clarify which molecule was actually used and ensure consistency.
  2. Fig. S6a: The panel contains 'Lorem ipsum' placeholder text in the y-axis label area. This should be corrected.
  3. Fig. 5a: The x-axis units for coverage are given as 'molec/nm²' in the figure caption but the in-text values (e.g., '0.03', '0.8', '1.28') lack units in the axis label itself. Please ensure axis labels are self-contained.
  4. §Methods: The Pb–X bond cut-off of 3.8 Å is stated without justification. A brief note on why this value was chosen (e.g., based on the first coordination shell) would help reproducibility.
  5. Table S6: The energy RMSE values are reported as 0.01 eV/atom for all models, which appears to be rounded to two decimal places. Given that the force RMSEs are reported with two decimal places and show meaningful variation, reporting energy RMSEs with higher precision (e.g., meV/atom) would be more informative.
  6. The paper would benefit from a brief discussion of the temperature choice (450 K) for the production MD simulations — whether this reflects experimental processing conditions or was chosen for sampling efficiency.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive assessment. The two major comments are well-taken: (1) the multi-molecule MD simulations extrapolate beyond the single-molecule D2 training data, and DFT cross-checks on multi-molecule snapshots are needed; and (2) the justification for selecting the CFT model over the simpler FM→AEAPTMS model for production simulations should be stated more explicitly. We will address both points in a revised manuscript.

read point-by-point responses

  1. Referee: The multi-molecule MD simulations (up to 40 AEAPTMS molecules) extrapolate well beyond the D2 training data, which contains only single-molecule passivation configurations. The steric crowding and intermolecular methoxy-group interactions invoked to explain the coverage-dependent shift are absent from the training data. No DFT cross-check is provided for any multi-molecule snapshot. This is load-bearing for quantitative claims (91.5% valid octahedra, ~20° tilt angles, E_ads stabilisation at ~−1.5 eV). At minimum, DFT single-point calculations on selected multi-molecule snapshots are needed.

    Authors: The referee is correct that the multi-molecule simulations in Fig. 5 extrapolate beyond the single-molecule configurations in D2, and that DFT cross-checks on multi-molecule snapshots are currently absent. This is a fair and important point. We will address it by computing DFT single-point energies on a representative selection of multi-molecule snapshots from the Fig. 5 simulations (at low, intermediate, and high coverage) and reporting the MLIP-vs-DFT energy and force deviations in a revised manuscript. We agree this is necessary to support the quantitative claims made in the application section. We will also add explicit discussion of the extrapolation risk and temper the quantitative claims accordingly where the DFT validation does not fully constrain them. Specifically, we will reframe the 91.5% valid octahedra and ~20° tilt values as MLIP predictions that are qualitatively consistent with experimental XRD trends, rather than as fully DFT-validated quantitative results, pending the DFT cross-checks we will now add. We note that the qualitative trends—increasing octahedral tilting with coverage, shift from multidentate to upright coordination, and progressive lattice disruption—are consistent with experimental observations (Refs. 2 and 4), which provides some external validation of the physical reasonableness of the simulations. However, we agree that this does not substitute for direct DFT validation on the multi-molecule configurations. revision: yes

  2. Referee: The paper acknowledges that the additional benefits of CFT over conventional fine-tuning are limited when using MACE-MH-1, and Fig. 4c shows the FM→AEAPTMS model achieves a slightly higher average SOAP similarity score (0.975) than the CFT model (0.954). Yet the production simulations in Fig. 5 use the MACE-MH-1-derived CFT model. The justification for selecting the CFT model over the simpler FM→AEAPTMS model should be stated more explicitly.

    Authors: The referee raises a valid point. The SOAP similarity scores in Fig. 4c are indeed very close (0.975 vs. 0.954), and the manuscript does not currently provide an explicit justification for choosing the CFT model for the production simulations. We will add a clearer justification in the revised manuscript. The rationale is as follows: while the average SOAP similarity scores are comparable, the CFT model shows improved force RMSE on the AEAPTMS test set (0.05 eV/Å vs. 0.06 eV/Å, Table S6) and, more importantly, improved accuracy on the surface slab component of the adsorption energy decomposition (Table S7: surface ΔE = 0.6 meV/atom for CFT vs. 1.0 meV/atom for FM→AEAPTMS). Since the production simulations involve extensive surface restructuring at high coverage, we considered the improved surface accuracy to be the deciding factor. We acknowledge, however, that the difference is modest and that the choice is not strongly favoured by the benchmarks. We will state this explicitly and note that the FM→AEAPTMS model would likely have been an equally reasonable choice, with the qualitative conclusions of the application simulations expected to be robust to this selection. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; the MLIP is trained on DFT data and produces structural predictions that are compared against independent experimental observations.

full rationale

The paper's central claim — that increasing AEAPTMS coverage shifts molecules from multidentate to upright coordination and disrupts the perovskite lattice — comes from MLIP-driven MD simulations (Fig. 5). The MLIP is trained on DFT-labelled data (D1='hyP-26', D2=single-molecule AEAPTMS passivation configurations) and then used to run multi-molecule MD simulations whose structural predictions (octahedral tilting, valid octahedra percentages, coordination shifts) are emergent properties of the simulation, not fitted quantities. The E_ads metric (Eq. 1) is a standard thermodynamic definition, not a circular construct. The qualitative agreement with experimental XRD trends (Ref. 4) is used as independent external validation. The paper transparently acknowledges that CFT benefits depend on the starting model (page 9: 'the additional benefits of CFT over conventional fine-tuning... are limited' for MACE-MH-1), which is an honest limitation, not a circular argument. Self-citations (e.g., to MACE architecture, graph-pes software, autoplex) are to tools and datasets, not to load-bearing 'uniqueness theorems' that would force the conclusion. The skeptic's concern about extrapolation beyond single-molecule training data is a correctness/extrapolation risk, not a circularity issue: the multi-molecule predictions are not equivalent to the training inputs by construction. No step in the derivation chain reduces to its own inputs by definition or by self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, particles, or forces. The free parameters are standard MLIP hyperparameters and simulation settings, chosen empirically. The axioms are standard domain assumptions for MLIP-based materials modelling.

free parameters (4)
  • Loss function weights (λE, λF, λv) = (1, 10, 100) for hyP-26; (1, 50, 1) for D2
    Chosen to improve force predictions; the D2 weights were found to substantially improve force RMSE.
  • Learning rate = 10^-4 (hyP-26 for MACE-MP-0b3); 10^-5 (hyP-26 for MACE-MH-1)
    Selected empirically based on performance in the CFT pipeline.
  • MD simulation temperature = 450 K
    Chosen for the MD simulations of surface passivation.
  • Pb-X bond cut-off = 3.8 Å
    Used for bond-length calculations.
axioms (4)
  • domain assumption The MACE architecture is a valid model for interatomic potentials.
    The entire approach relies on MACE-based MLIPs being suitable for this chemical space.
  • domain assumption SCAN DFT is accurate enough for this system.
    All reference data is generated at the SCAN level of theory.
  • domain assumption The Pb-terminated (001) surface is representative.
    The MD simulations focus on this specific surface without additional vacancies or interstitial defects.
  • domain assumption Naive fine-tuning is sufficient for the target application.
    The paper assumes catastrophic forgetting does not need to be actively prevented for the target task.

pith-pipeline@v1.1.0-glm · 30821 in / 2217 out tokens · 228893 ms · 2026-07-07T17:08:48.726397+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

86 extracted references · 86 canonical work pages · 7 internal anchors

  1. [1]

    https://github.com/vldgroup/graph-pes

    work page

  2. [2]

    , author Kovacs, D

    author Batatia, I. , author Kovacs, D. P. , author Simm, G. , author Ortner, C. & author Csanyi, G. title MACE : H igher O rder E quivariant M essage P assing N eural N etworks for F ast and A ccurate F orce F ields . In editor Koyejo, S. et al. (eds.) booktitle Advances in Neural Information Processing Systems , vol. volume 35 , pages 11423--11436 ( year 2022 )

    work page 2022

  3. [3]

    author Batatia, I. et al. title Cross Learning between Electronic Structure Theories for Unifying Molecular , Surface , and Inorganic Crystal Foundation Force Fields ( year 2025 ). note arXiv:2510.25380 [physics.chem-ph]

    work page arXiv 2025

  4. [4]

    author Batatia, I. et al. title A foundation model for atomistic materials chemistry ( year 2023 ). note arXiv:2401.00096v1 [physics.chem-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    author Kingma, D. P. & author Ba, J. title Adam: A method for stochastic optimization ( year 2014 ). note arXiv:1412.6980 [cs.LG]

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  6. [6]

    , author Fransson, E

    author Eriksson, F. , author Fransson, E. & author Erhart, P. title The hiphive package for the extraction of high-order force constants by machine learning . journal Advanced Theory and Simulations volume 2 , pages 1800184 ( year 2019 )

    work page 2019

  7. [7]

    author Islam, M. A. , author Rondinelli, J. M. & author Spanier, J. E. title Normal mode determination of perovskite crystal structures with octahedral rotations: theory and applications . journal Journal of Physics: Condensed Matter volume 25 , pages 175902 ( year 2013 )

    work page 2013

  8. [8]

    title Crystal structures of perovskite halide compounds used for solar cells

    author Oku, T. title Crystal structures of perovskite halide compounds used for solar cells . journal Reviews on Advanced Materials Science volume 59 , pages 264--305 ( year 2020 )

    work page 2020

  9. [9]

    , author Shi, T

    author Yin, W.-J. , author Shi, T. & author Yan, Y. title Unusual defect physics in CH3NH3PbI3 perovskite solar cell absorber . journal Applied Physics Letters volume 104 , pages 063903 ( year 2014 )

    work page 2014

  10. [10]

    , author Fujisawa, J.-I

    author Giorgi, G. , author Fujisawa, J.-I. , author Segawa, H. & author Yamashita, K. title Cation Role in Structural and Electronic Properties of 3D Organic – Inorganic Halide Perovskites : A DFT Analysis . journal The Journal of Physical Chemistry C volume 118 , pages 12176--12183 ( year 2014 )

    work page 2014

  11. [11]

    , author Hamad, S

    author Grau-Crespo, R. , author Hamad, S. , author Catlow, C. R. A. & author de Leeuw, N. H. title Symmetry-adapted configurational modelling of fractional site occupancy in solids . journal Journal of Physics: Condensed Matter volume 19 , pages 256201 ( year 2007 )

    work page 2007

  12. [12]

    author McMeekin, D. P. et al. title Crystallization Kinetics and Morphology Control of Formamidinium – Cesium Mixed - Cation Lead Mixed - Halide Perovskite via Tunability of the Colloidal Precursor Solution . journal Advanced Materials volume 29 , pages 1607039 ( year 2017 )

    work page 2017

  13. [13]

    author Fabini, D. H. et al. title Reentrant structural and optical properties and large positive thermal expansion in perovskite formamidinium lead iodide . journal Angewandte Chemie International Edition volume 55 , pages 15392--15396 ( year 2016 )

    work page 2016

  14. [14]

    author Unger, E. L. et al. title Chloride in lead chloride-derived organo-metal halides for perovskite-absorber solar cells . journal Chemistry of Materials volume 26 , pages 7158--7165 ( year 2014 )

    work page 2014

  15. [15]

    author Jeon, N. J. et al. title Compositional engineering of perovskite materials for high-performance solar cells . journal Nature volume 517 , pages 476--480 ( year 2015 )

    work page 2015

  16. [16]

    author Chen, Q. et al. title Under the spotlight: The organic–inorganic hybrid halide perovskite for optoelectronic applications . journal Nano Today volume 10 , pages 355--396 ( year 2015 )

    work page 2015

  17. [17]

    author Lyahovitskaya, V. et al. title Formation and thermal stability of quasi-amorphous thin films . journal Physical Review B volume 71 , pages 094205 ( year 2005 )

    work page 2005

  18. [18]

    , author Csányi, G

    author Bernstein, N. , author Csányi, G. & author Deringer, V. L. title De novo exploration and self-guided learning of potential-energy surfaces . journal npj Computational Materials volume 5 , pages 99 ( year 2019 )

    work page 2019

  19. [19]

    , author Liu, Y

    author Paşca, L.-B. , author Liu, Y. , author Anker, A. S. , author Steier, L. & author Deringer, V. L. title Machine-learning-driven modelling of amorphous and polycrystalline BaZrS3 . journal Journal of Materials Chemistry A volume 13 , pages 35447--35454 ( year 2025 )

    work page 2025

  20. [20]

    author Deringer, V. L. , author Pickard, C. J. & author Cs\'anyi, G. title Data-driven learning of total and local energies in elemental boron . journal Physical Review Letters volume 120 , pages 156001 ( year 2018 )

    work page 2018

  21. [21]

    author Liu, Y. et al. title An automated framework for exploring and learning potential-energy surfaces . journal Nature Communications volume 16 , pages 7666 ( year 2025 )

    work page 2025

  22. [22]

    author Bart\'ok, A. P. , author Payne, M. C. , author Kondor, R. & author Cs\'anyi, G. title Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons . journal Physical Review Letters volume 104 , pages 136403 ( year 2010 )

    work page 2010

  23. [23]

    author Bart\'ok, A. P. , author Kondor, R. & author Cs\'anyi, G. title On representing chemical environments . journal Physical Review B volume 87 , pages 184115 ( year 2013 )

    work page 2013

  24. [24]

    , author Wei, S.-H

    author Zunger, A. , author Wei, S.-H. , author Ferreira, L. G. & author Bernard, J. E. title Special quasirandom structures . journal Physical Review Letters volume 65 , pages 353--356 ( year 1990 )

    work page 1990

  25. [25]

    , author Fri \' a k, M

    author Gehringer, D. , author Fri \' a k, M. & author Holec, D. title Models of configurationally-complex alloys made simple . journal Computer Physics Communications volume 286 , pages 108664 ( year 2023 )

    work page 2023

  26. [26]

    author Kavanagh, S. R. et al. title doped: Python toolkit for robust and repeatable charged defect supercell calculations . journal Journal of Open Source Software volume 9 , pages 6433 ( year 2024 )

    work page 2024

  27. [27]

    author Ong, S. P. et al. title Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis . journal Computational Materials Science volume 68 , pages 314--319 ( year 2013 )

    work page 2013

  28. [28]

    , author Ruzsinszky, A

    author Sun, J. , author Ruzsinszky, A. & author Perdew, J. title Strongly Constrained and Appropriately Normed Semilocal Density Functional . journal Physical Review Letters volume 115 , pages 036402 ( year 2015 )

    work page 2015

  29. [29]

    & author Hafner, J

    author Kresse, G. & author Hafner, J. title Ab initio molecular dynamics for liquid metals . journal Physical Review B volume 47 , pages 558--561 ( year 1993 )

    work page 1993

  30. [30]

    & author Furthmüller, J

    author Kresse, G. & author Furthmüller, J. title Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set . journal Computational Materials Science volume 6 , pages 15--50 ( year 1996 )

    work page 1996

  31. [31]

    & author Furthmüller, J

    author Kresse, G. & author Furthmüller, J. title Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set . journal Physical Review B volume 54 , pages 11169--11186 ( year 1996 )

    work page 1996

  32. [32]

    author Blöchl, P. E. title Projector augmented-wave method . journal Physical Review B volume 50 , pages 17953--17979 ( year 1994 )

    work page 1994

  33. [33]

    & author Joubert, D

    author Kresse, G. & author Joubert, D. title From ultrasoft pseudopotentials to the projector augmented-wave method . journal Physical Review B volume 59 , pages 1758--1775 ( year 1999 )

    work page 1999

  34. [34]

    https://vasp.at/wiki/LASPH (last accessed 10 June 2026)

    work page 2026

  35. [35]

    & author Kaczkowski, J

    author Płowaś-Korus, I. & author Kaczkowski, J. title Comparative density functional studies of BiMO _3 polymorphs (M = Al, Ga, In) based on LDA, GGA, and meta-GGA functionals . journal New Journal of Chemistry volume 46 , pages 15381--15391 ( year 2022 )

    work page 2022

  36. [36]

    , author Walker, A

    author Brivio, F. , author Walker, A. B. & author Walsh, A. title Structural and electronic properties of hybrid perovskites for high-efficiency thin-film photovoltaics from first-principles . journal APL Materials volume 1 , pages 042111 ( year 2013 )

    work page 2013

  37. [37]

    author Weller, M. T. , author Weber, O. J. , author Frost, J. M. & author Walsh, A. title Cubic perovskite structure of black formamidinium lead iodide, - [HC(NH2)2]PbI3 , at 298 K . journal The Journal of Physical Chemistry Letters volume 6 , pages 3209--3212 ( year 2015 )

    work page 2015

  38. [38]

    , author Healy, J

    author McInnes, L. , author Healy, J. , author Saul, N. & author Großberger, L. title UMAP : Uniform Manifold Approximation and Projection . journal Journal of Open Source Software volume 3 , pages 861 ( year 2018 )

    work page 2018

  39. [39]

    title Visualization and analysis of atomistic simulation data with OVITO-the Open Visualization Tool

    author Stukowski, A. title Visualization and analysis of atomistic simulation data with OVITO-the Open Visualization Tool . journal Modelling and Simulation in Materials Science and Engineering volume 18 , pages 015012 ( year 2010 )

    work page 2010

  40. [40]

    author Helal Miah, M. et al. title Key degradation mechanisms of perovskite solar cells and strategies for enhanced stability: issues and prospects . journal RSC Advances volume 15 , pages 628--654 ( year 2025 )

    work page 2025

  41. [41]

    author Jariwala, S. et al. title Reducing surface recombination velocity of methylammonium-free mixed-cation mixed-halide perovskites via surface passivation . journal Chemistry of Materials volume 33 , pages 5035--5044 ( year 2021 )

    work page 2021

  42. [42]

    author Lin, Y.-H. et al. title Bandgap-universal passivation enables stable perovskite solar cells with low photovoltage loss . journal Science volume 384 , pages 767--775 ( year 2024 )

    work page 2024

  43. [43]

    author Huang, Z. et al. title Surface passivation for halide optoelectronics: Comparing optimization and reactivity of amino-silanes with formamidinium . journal Journal of the American Chemical Society volume 147 , pages 42918--42925 ( year 2025 )

    work page 2025

  44. [44]

    author Lu, D. et al. title 86 pflops deep potential molecular dynamics simulation of 100 million atoms with ab initio accuracy . journal Computer Physics Communications volume 259 , pages 107624 ( year 2021 )

    work page 2021

  45. [45]

    author Morrow, J. D. et al. title Understanding defects in amorphous silicon with million-atom simulations and machine learning . journal Angewandte Chemie International Edition volume 63 , pages e202403842 ( year 2024 )

    work page 2024

  46. [46]

    , author Zhu, W

    author Bian, T. , author Zhu, W. , author Lei, Q. & author Yin, J. title Machine learning potentials for inorganic and hybrid lead halide perovskites: From phase stability to defects and interfaces . journal ACS Applied Materials & Interfaces volume 18 , pages 17219--17238 ( year 2026 )

    work page 2026

  47. [47]

    author Deng, B. et al. title Overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning ( year 2024 ). note arXiv:2405.07105 [cond-mat.mtrl-sci]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [48]

    author Batatia, I. et al. title A foundation model for atomistic materials chemistry . journal The Journal of Chemical Physics volume 163 , pages 184110 ( year 2025 )

    work page 2025

  49. [49]

    author El Aisnada, A. N. et al. title Cost-effective strategy of enhancing machine learning potentials by transfer learning from a multicomponent data set on ænet-pytorch . journal The Journal of Physical Chemistry C volume 129 , pages 658--669 ( year 2025 )

    work page 2025

  50. [50]

    author Tompa, T. et al. title Fine-tuning MLIP foundation models: strategies for accuracy and transferability ( year 2026 ). note arXiv:2606.12704 [physics.chem-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  51. [51]

    , author Mohandas, N

    author Echeverri Restrepo, S. , author Mohandas, N. K. , author Sluiter, M. H. F. & author Paxton, A. T. title Applicability of universal machine learning interatomic potentials to the simulation of steels . journal Modelling and Simulation in Materials Science and Engineering volume 33 , pages 035003 ( year 2025 )

    work page 2025

  52. [52]

    author Kempen, L. H. E. , author Cheula, R. & author Andersen, M. title How accurate are foundational machine learning interatomic potentials for heterogeneous catalysis? ( year 2025 ). note arXiv:2512.16702 [cond-mat.mtrl-sci]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  53. [53]

    author Wu, Z. et al. title A machine learning interatomic potential data set and model for catalysis with local fine-tuning to chemical accuracy . journal JACS Au volume 5 , pages 6151--6161 ( year 2025 )

    work page 2025

  54. [54]

    , author Desai, R

    author Biswas, M. , author Desai, R. , author Bidna, G. & author Mannodi-Kanakkithodi, A. title Unified graph-based interatomic potential for perovskite structure optimization . journal Journal of Chemical Information and Modeling volume 66 , pages 1353--1370 ( year 2026 )

    work page 2026

  55. [55]

    CITB: A Benchmark for Continual Instruction Tuning

    author Zhang, Z. , author Fang, M. , author Chen, L. & author Namazi-Rad, M.-R. title CITB : A Benchmark for Continual Instruction Tuning ( year 2023 ). note arXiv:2310.14510 [cs.CL]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  56. [56]

    Exploring Continual Fine-Tuning for Enhancing Language Ability in Large Language Model

    author Aggarwal, D. , author Damle, S. , author Goyal, N. , author Lokam, S. & author Sitaram, S. title Exploring Continual Fine - Tuning for Enhancing Language Ability in Large Language Model ( year 2024 ). note arXiv.2410.16006 [cs.CL]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  57. [57]

    author Guan, C. et al. title Multi-stage LLM fine-tuning with a continual learning setting . In editor Chiruzzo, L. , editor Ritter, A. & editor Wang, L. (eds.) booktitle Findings of the Association for Computational Linguistics: NAACL 2025 , pages 5499--5513 ( year 2025 )

    work page 2025

  58. [58]

    author Ahlawat, D. et al. title A family of large language models for materials research with insights into model adaptability in continued pretraining . journal Nature Machine Intelligence volume 8 , pages 435--448 ( year 2026 )

    work page 2026

  59. [59]

    & author Reiher, M

    author Eckhoff, M. & author Reiher, M. title Lifelong machine learning potentials for chemical reaction network explorations . journal Journal of Chemical Theory and Computation volume 21 , pages 9641--9656 ( year 2025 )

    work page 2025

  60. [60]

    author Kim, J. et al. title An efficient forgetting-aware fine-tuning framework for pretrained universal machine-learning interatomic potentials . journal npj Computational Materials volume 12 , pages 26 ( year 2026 )

    work page 2026

  61. [61]

    author Liu, X. et al. title An AI -ready fine-tuning framework for accurate machine-learning interatomic potentials in solid-solid battery interfaces ( year 2026 ). note arXiv:2601.17847 [cond-mat.mtrl-sci]

    work page arXiv 2026

  62. [62]

    author Wong, N. H. & author Yang, J. H. title Bias in Universal Machine-Learned Interatomic Potentials and Its Effects on Fine-Tuning . journal Journal of Chemical Theory and Computation , DOI: 10.1021/acs.jctc.6c00425 ( year 2026 )

    work page doi:10.1021/acs.jctc.6c00425 2026

  63. [63]

    author Chanussot, L. et al. title Open catalyst 2020 (OC20) dataset and community challenges . journal ACS Catalysis volume 11 , pages 6059--6072 ( year 2021 )

    work page 2020

  64. [64]

    author Deng, B. et al. title CHGN et as a pretrained universal neural network potential for charge-informed atomistic modelling . journal Nature Machine Intelligence volume 5 , pages 1031--1041 ( year 2023 )

    work page 2023

  65. [65]

    author Barroso-Luque, L. et al. title The Open Materials 2024 ( OMat24 ) inorganic materials dataset and models . journal Nature Computational Science , DOI: 10.1038/s43588-026-00996-w ( year 2026 )

    work page doi:10.1038/s43588-026-00996-w 2024

  66. [66]

    author Levine, D. et al. title The O pen M olecules 2025 ( OMol25 ) D ataset, E valuations, and M odels ( year 2025 ). note arXiv:2505.08762 [physics.chem-ph]

    work page arXiv 2025

  67. [67]

    author Cavignac, T. et al. title AI - Driven Expansion and Application of the Alexandria Database . journal Journal of Physics: Materials volume 9 , pages 025014 ( year 2026 )

    work page 2026

  68. [68]

    title Knowledge distillation of neural network potential for molecular crystals

    author Taniguchi, T. title Knowledge distillation of neural network potential for molecular crystals . journal Faraday Discussions volume 256 , pages 139--155 ( year 2025 )

    work page 2025

  69. [69]

    author Johnson, M. S. et al. title Fusing Delta Learning and Machine Learned Interatomic Potentials for Efficient High Throughput Calculation of Surface Kinetic Parameters ( year 2026 ). note DOI: 10.26434/chemrxiv.15000328/v1

    work page doi:10.26434/chemrxiv.15000328/v1 2026

  70. [70]

    author Shi, Y. et al. title (3-aminopropyl)trimethoxysilane surface passivation improves perovskite solar cell performance by reducing surface recombination velocity . journal ACS Energy Letters volume 7 , pages 4081--4088 ( year 2022 )

    work page 2022

  71. [71]

    author Ng, E. Z. X. , author Ang, S. J. , author Yang, H. , author Chen, Y. & author Wong, M. W. title Accelerated and efficient modeling of low- organosilicate glass with the M3GNet machine learning interatomic potentials . journal Computational Materials Today volume 8 , pages 100042 ( year 2025 )

    work page 2025

  72. [72]

    , author Walsh Atkins, P

    author Ghosh, D. , author Walsh Atkins, P. , author Islam, M. S. , author Walker, A. B. & author Eames, C. title Good Vibrations : Locking of Octahedral Tilting in Mixed - Cation Iodide Perovskites for Solar Cells . journal ACS Energy Letters volume 2 , pages 2424--2429 ( year 2017 )

    work page 2017

  73. [73]

    D evice-scale atomistic modelling of phase-change memory materials

    author Zhou, Y. , author Zhang, W. , author Ma, E. & author Deringer, V. L. title Research data for " D evice-scale atomistic modelling of phase-change memory materials" ( year 2023 ). note DOI: 10.5281/zenodo.8208202

    work page doi:10.5281/zenodo.8208202 2023

  74. [74]

    , author Kuklin, M

    author Ibragimova, R. , author Kuklin, M. S. & author Caro, M. A. title Dataset for CH GAP interatomic potential ( year 2024 ). note DOI: 10.5281/zenodo.10925480

    work page doi:10.5281/zenodo.10925480 2024

  75. [75]

    title Unified graph-based interatomic potential for perovskite structure optimization ( year 2025 )

    author Biswas, M. title Unified graph-based interatomic potential for perovskite structure optimization ( year 2025 ). note DOI: 10.5281/zenodo.17363611

    work page doi:10.5281/zenodo.17363611 2025

  76. [76]

    author Lee, K. et al. title A comprehensive dataset of chemical reactions covering second and third row elements with million-scale quantum calculations ( year 2025 ). note DOI: 10.26434/chemrxiv-2025-bh42x

    work page doi:10.26434/chemrxiv-2025-bh42x 2025

  77. [77]

    author Chenebuah, E. T. , author Nganbe, M. & author Tchagang, A. B. title A deep generative modeling architecture for designing lattice-constrained perovskite materials . journal npj Computational Materials volume 10 , pages 198 ( year 2024 )

    work page 2024

  78. [78]

    , author Jacobs, R

    author Li, W. , author Jacobs, R. & author Morgan, D. title Predicting the thermodynamic stability of perovskite oxides using machine learning models . journal Computational Materials Science volume 150 , pages 454--463 ( year 2018 )

    work page 2018

  79. [79]

    author Zhang, Y. et al. title Machine Learning for Designing Perovskites and Perovskite-Inspired Solar Materials: Emerging Opportunities and Challenges . journal Advanced Science volume 13 , pages e74952 ( year 2026 )

    work page 2026

  80. [80]

    , author Madanchi, A

    author Liu, Y. , author Madanchi, A. , author Anker, A. S. , author Simine, L. & author Deringer, V. L. title The amorphous state as a frontier in computational materials design . journal Nature Reviews Materials volume 10 , pages 228--241 ( year 2025 )

    work page 2025

Showing first 80 references.