Pith. sign in

REVIEW 4 major objections 7 minor 25 references

Molecular dynamics with a first-principles-validated universal machine-learning potential reveals dynamic elementary processes of growth-related adspecies on GaN(0001)

T0 review · 4 major / 7 minor · reviewed 2026-07-30 · grok-4.5

Pith's one-line read A universal machine-learning potential, checked against first-principles trajectories, shows that GaN growth units migrate by lifting, hop hydrogen, and reversibly split on the hot surface.

desk verdict Solid first MLIP MD of GaN MOVPE surfaces: FPMD finds lifting-assisted GaNH motion and H abstraction; UMA parity is real but the 150 ps rare-event claims rest on energy validation that does not fully pin the Ga–Ga offset or barriers. read the letter →

arxiv 2607.23461 v1 pith:OGLG24H6 submitted 2026-07-26 cond-mat.mtrl-sci physics.comp-ph

classification cond-mat.mtrl-sciphysics.comp-ph
keywords GaNMOVPEmachine-learninginteratomicpotentialmoleculardynamicsNHadmoleculesurfacediffusionhydrogenhoppinguniversalMLIPN(0001)
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

Atomic-scale pictures of how GaN grows by metalorganic vapor-phase epitaxy have long come from zero-temperature energy maps or from first-principles molecular dynamics that last only a few tens of picoseconds. This paper shows that a pretrained universal machine-learning potential can carry those dynamics out to 150 ps after it has been shown to match first-principles relative energies along a short reference trajectory, without any GaN-specific retraining. In the short first-principles runs the GaNH growth unit does not hop as a rigid molecule; its nitrogen atom lifts the attached gallium off the surface while the unit migrates, and the lifted gallium can snatch a neighboring hydrogen atom. The longer machine-learning runs then reveal two processes that never appear in the short window: hydrogen atoms hop from site to site and thereby open or close paths for the growth unit, and the GaNH unit itself repeatedly falls apart into a free Ga adatom and an NH molecule that wander independently before recombining. The result recasts earlier static pictures of concerted “hand-in-hand” transport and hydrogen blocking as intermittent, thermally gated dynamics at growth temperature.

What carries the argument

The GaNH admolecule on hydrogenated GaN(0001), whose lifting-assisted migration, hydrogen-abstraction events, site-to-site H hops, and reversible dissociation GaNH ⇌ Ga_ad + NH_ad are the elementary dynamical objects followed from 1.5 ps first-principles runs into 150 ps machine-learning runs.

What would settle it

A longer first-principles trajectory or an independent ensemble of trajectories that either fails to show reversible GaNH dissociation and hydrogen-gated migration on the same time scale, or shows that the machine-learning potential’s state-dependent energy offset systematically alters the populations or barriers of those events.

Watch

Extended reading notes

Core claim

After single-point validation that a universal machine-learning potential reproduces first-principles relative energies along a GaNH-plus-hydrogen trajectory on GaN(0001) to roughly 8.5 meV per atom without retraining, 150-ps molecular dynamics with that potential shows that the growth unit migrates by a lifting-assisted mode, that surface hydrogen hops gate its paths, and that GaNH reversibly dissociates into independently migrating Ga and NH that later re-form.

Load-bearing premise

That matching relative energies on snapshots from one short first-principles trajectory is enough to trust the machine-learning potential for rare events and long-time surface dynamics, even though it carries a systematic offset on the weak gallium–gallium contacts that define the growing surface.

Editorial extensions

If this is right

  • Mass transport of the growth unit on GaN(0001) at MOVPE temperature proceeds by repeated dissociation and re-association rather than as a permanently bound GaNH complex.
  • Surface hydrogen gates migration paths dynamically by hopping, rather than statically blocking them.
  • Universal machine-learning potentials can be used for GaN MOVPE surface dynamics after lightweight first-principles validation, without system-specific retraining.
  • The length scale of adspecies delivery to step edges must be re-estimated under intermittent Ga–NH partnership.

Reading between the lines

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

  • The same validation-plus-long-MD protocol could be applied next to NH3 adsorption and sequential dehydrogenation, testing whether precursor kinetics also change character beyond the picosecond window.
  • If reversible dissociation is general, continuum or kinetic Monte Carlo growth models that treat GaNH as a single rigid carrier will need explicit association–dissociation rates.
  • The missing recurrence of hydrogen abstraction in the 150 ps run suggests its contribution to surface dehydrogenation is rarer than hopping and will require still longer or parallel trajectories to quantify.
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

4 major / 7 minor

Summary. The manuscript combines 1.5 ps CASTEP/PBE FPMD with 150 ps MD driven by the pretrained universal potential UMA (no system-specific retraining) to study a GaNH admolecule plus nine H adatoms on GaN(0001) at 1273 K. FPMD reveals a lifting-assisted migration mode (the N atom lifts Gaadmol off the surface layer) and an H-abstraction event by the lifted Ga. UMA is validated by single-point recalculation of 751 FPMD snapshots (RMSE 8.5 meV/atom, R²=0.946, slope ≈1), with the residuals resolving into two state-dependent branches: lifted configurations near the identity line and surface-engaged configurations (weak Ga-Ga contacts) offset by ≈10 meV/atom. The 150 ps UMA run then shows H-adatom hopping that gates GaNH migration and four reversible GaNH ⇌ Ga_ad + NH_ad dissociation events. The authors claim the first MLIP-based MD study of GaN MOVPE and a dynamical recasting of the static 'hand-in-hand' growth-unit picture.

Significance. If the rare-event claims hold, the paper delivers a qualitatively new picture of MOVPE-relevant surface transport: the GaNH 'growth unit' as an intermittently dissociating complex rather than a rigid carrier, and hydrogen as a dynamic gate rather than a static blocker — a dynamical generalization of the static 'hand-in-hand' migration picture (Ref. 10). Methodologically, the zero-retraining validation of a universal potential against FPMD on a chemically demanding surface (weak metallic Ga-Ga bonds, N-H/Ga-H species, 1273 K) is a useful data point for the community, and the branch-resolved parity analysis (Fig. 4) is a genuinely instructive way to expose state-dependent MLIP error. The FPMD observations themselves (lifting-assisted migration; H abstraction by lifted Ga) are direct, well-supported, and novel relative to static DFT. The main limitation to significance is not novelty but the single-trajectory, single-validation-window evidentiary base for the long-time claims.

major comments (4)
  1. [§Validation of the UMA potential; Fig. 4-5] §Validation / Fig. 4 and §Long-time dynamics / Fig. 5: the central long-time claim (reversible GaNH ⇌ Ga_ad + NH_ad dissociation, with dissociation periods up to ~20 ps) rests on validation performed only along one 1.5 ps FPMD trajectory, using single-point relative energies per atom. The paper's own branch analysis shows this metric hides a systematic state-dependent bias: the surface-engaged branch is offset by ≈10 meV/atom (y = 0.995x + 11.5, Fig. 4a) relative to the lifted branch. For a slab of ~180 atoms, a 10 meV/atom offset corresponds to a total-energy bias on the order of 1-2 eV between the two bonding states — precisely the energetic competition (weak Ga-Ga contacts of Gaadmol with the surface) that controls association, dissociation, and the lifted/engaged population balance. A bias of this size could plausibly destabilize the associated GaNH state or flatten the relevant barr
  2. [§Methods: Machine-learning interatomic potential] §Methods (MLIP MD) and §Long-time dynamics: the mixed integration scheme (adspecies + top bilayer NVE; deeper bilayers Nosé-Hoover at 1273 K, chain length 5, damping time 10 fs) is asserted to 'avoid thermostatting the adspecies dynamics directly', but no diagnostics are reported. The kinetics of rare adspecies events depend directly on the energy content and effective temperature of the NVE subsystem; a 10 fs damping time in the neighboring thermostat layers is strong coupling and could pump or drain energy from the NVE region. The authors should report (i) energy drift of the NVE subsystem over the 150 ps run and (ii) the time-averaged kinetic temperature of the adspecies/top-bilayer group, and ideally demonstrate that event phenomenology is unchanged under a fully thermostatted short run or a different damping time. This is checkable within the existing setup.
  3. [§Results: hydrogen abstraction (Fig. 3) and its non-recurrence] §Results, H-abstraction discrepancy: the FPMD run shows H abstraction by lifted Gaadmol (Fig. 3, t ≈ 1.41-1.50 ps), but this event never recurs in the 100× longer UMA trajectory, and the authors interpret this as rarity contingent on encounter geometry. This interpretation is untested and cuts both ways: if the abstraction is a chemically activated event whose barrier depends on the same Ga-Ga/Ga-H energetics carrying the ≈10 meV/atom state-dependent offset, its absence in UMA could equally indicate that UMA overestimates the abstraction barrier — i.e., the potential may be failing on exactly the class of reactive events one FPMD run happened to sample. The authors should either (a) validate the abstraction pathway directly (e.g., a short UMA run started from the FPMD pre-abstraction configuration, or DFT vs UMA energies along the FPMD abstraction segment, which is already in the 751-sna
  4. [§Methods: event identification; Fig. 5] §Methods / Fig. 5: dissociation and re-formation events are 'identified by visual inspection' and delineated with the Gaadmol-N distance, but no quantitative distance criterion is given (the 3.3 Å Ga-Ga criterion explicitly excludes dissociated periods). Since the count of dissociation events (four) and their durations (0.4-20 ps) are quantitative results quoted in the text, the event-definition threshold and any hysteresis used must be stated; otherwise event counts are not reproducible. Relatedly, with only four events in one trajectory, statements such as 'dynamical equilibrium GaNH ⇌ Ga_ad + NH_ad' and the implied gating kinetics carry no uncertainty estimates; the text should present them as single-trajectory observations, or provide a small ensemble of shorter runs (e.g., 5 × 30 ps from decorrelated initial conditions), which is cheap with UMA.
minor comments (7)
  1. [§Validation of the UMA potential] §Validation, Fig. 4: only energies are benchmarked. UMA-m 1.1 with the OC20 head is used as a force field in LAMMPS; a brief force-error benchmark (e.g., force RMSE on a subset of the 751 snapshots, resolved by branch) would strengthen the validation at modest cost and is standard practice for MLIP-MD papers.
  2. [§Methods] §Methods: the choice of 0.1 fs time step is conservative for FPMD but should be justified for the 1.5 M-step MLIP run; conversely, if a larger step is stable for UMA, the 150 ps length was chosen conservatively — either way, one sentence on energy conservation at 0.1 fs in the pure-NVE limit would suffice.
  3. [Fig. 4(b); §Methods] Fig. 4(b): the 3.3 Å threshold is read off the histogram, but the residual-based branch assignment (6 meV/atom on a 7-point rolling mean, §Methods) and the structural criterion appear to be applied in a slightly circular way (branches defined by residuals, then structurally interpreted, then the structure used as the state classifier). A short clarification that the 3.3 Å cut cleanly separates the residual-defined branches (e.g., a confusion rate) would make the two-state classification airtight.
  4. [§Long-time dynamics revealed by MLIP-based MD] Population comparison (56/44% MLIP vs 50/50% FPMD, §Long-time dynamics): given the ≈10 meV/atom branch offset, this agreement is presented as evidence that the offset does not bias sampling, but with 1.5 ps of FPMD the reference populations have very large uncertainty; the authors' own caveat should be strengthened to note that ~1 eV-scale total-energy biases between branches can leave short-window populations nearly unchanged while strongly affecting rare-event barriers.
  5. [Fig. 5] Fig. 5: state the sampling interval (10 fs) and minimum-image convention in the caption (currently only in Methods), and indicate on the figure or in text how the shaded 'dissociated' intervals were delimited (tie-in with major comment 4).
  6. [§Methods: First-principles molecular dynamics] §Methods: PBE without dispersion correction is used for the reference data; a one-sentence remark on the expected (in)sensitivity of weak Ga-Ga surface interactions to dispersion would preempt an obvious question, given that the state-dependent UMA offset is attributed to that very interaction.
  7. [§Methods] Reproducibility: the workflow uses public tools (fairchem-core 2.19.0, fairchem-lammps 0.4.0, LAMMPS 2 Aug 2023 update 3) and a public model — commendable. Depositing the initial slab geometry, LAMMPS input, and the 751-snapshot parity dataset (e.g., on Zenodo) would make the validation fully reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: UMA is externally pretrained; dynamics are validated against independent CASTEP/PBE energies, not fitted to the claimed rare events.

full rationale

The paper’s load-bearing chain is (1) CASTEP FPMD observations of lifting-assisted GaNH migration and H abstraction, (2) single-point UMA energy parity on those FPMD snapshots (RMSE ≈ 8.5 meV/atom, no retraining), and (3) longer UMA MD that reports H hops and reversible GaNH dissociation. UMA is a universal potential pretrained on ~500M external DFT calculations; it is not fit to the target trajectories or to the rare-event outcomes. Relative energies are compared to an independent first-principles code (CASTEP/PBE) on the same geometries. Prior author citations supply surface reconstructions and static pathway context but do not algebraically force the new dynamical claims. Concerns that energy parity may not guarantee barrier/force fidelity for rare events are correctness/extrapolation risks, not circular reductions of outputs to inputs. No self-definitional loop, fitted-input-as-prediction, or load-bearing uniqueness import is present. Score 0; steps empty.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

Claims rest on standard DFT-MD practice, a pretrained universal potential used without fine-tuning, a specific slab/composition chosen to mimic H2-rich MOVPE, and analysis thresholds (Ga–Ga distance, residual smoothing) that classify bonding states. No new physical entities are postulated; ‘lifting-assisted migration’ and ‘reversible dissociation’ are observed modes. Trust in long MD equals trust that UMA errors (especially on weak Ga–Ga) do not invent those modes.

free parameters (3)
  • Ga–Ga distance threshold 3.3 Å for lifted vs surface-engaged = 3.3 Å
    Chosen from the bimodal distance distribution in Fig. 4(b) to classify bonding state and to exclude dissociated periods in population analysis; directly affects reported 56%/44% populations.
  • Residual branch-assignment threshold 6 meV/atom (7-point rolling mean) = 6 meV/atom
    Hand-chosen cutoff on smoothed ΔE_MLIP−ΔE_DFT used to color the two parity branches; analysis convenience parameter.
  • Nosé–Hoover damping time and chain length in mixed thermostat = 10 fs, chain length 5
    Damping 10 fs, chain length 5 on subsurface layers; standard but tunable choices that can affect heat flow into the NVE adspecies region.
assumptions (4)
  • domain assumption PBE-DFT without dispersion, 280 eV cutoff, Γ-only sampling on the 4×4 five-bilayer slab is an adequate reference for relative adspecies energetics at 1273 K.
    Methods: entire FPMD setup and the validation target for UMA.
  • domain assumption The chosen coverage (nine H adatoms + one GaNH on 4×4, modeling 3Ga–H plus growth unit) represents the essential MOVPE H2-rich growth front for the reported mechanisms.
    Methods and Introduction linking to prior reconstruction thermodynamics.
  • domain assumption Pretrained UMA-m 1.1 (OC20 head) transfers to metallic Ga–Ga, polar Ga–N, and hydride adspecies on GaN without system-specific retraining for dynamical conclusions.
    Validation section and all 150 ps results; supported only by single-trajectory energy parity.
  • ad hoc to paper Mixed NVE (adspecies + top bilayer) / NVT (deeper layers) integration does not distort rare-event statistics relative to a fully thermostatted run.
    Methods: scheme adopted ‘to avoid thermostatting the adspecies dynamics directly,’ without comparative tests.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Molecular dynamics with a first-principles-validated universal machine-learning potential reveals dynamic elementary processes of growth-related adspecies on GaN(0001)." pith.science (2026). https://pith.science/paper/OGLG24H6

@misc{pith2026260723461,
  author       = {Pith},
  title        = {Pith review of: Molecular dynamics with a first-principles-validated universal machine-learning potential reveals dynamic elementary processes of growth-related adspecies on GaN(0001)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OGLG24H6}},
  note         = {Machine review of arXiv:2607.23461}
}
read the original abstract

Atomic-scale understanding of the surface elementary processes in metalorganic vapor phase epitaxy (MOVPE) of GaN has so far relied on static density-functional-theory (DFT) energetics and on first-principles molecular dynamics (FPMD) limited to a few tens of picoseconds. Here we combine FPMD with a universal machine-learning interatomic potential (MLIP), UMA, to follow the dynamics of growth-related adspecies on GaN(0001) over time scales inaccessible to purely first-principles approaches. FPMD simulations of a GaNH admolecule coexisting with H adatoms reveal a hitherto unrecognized diffusion mode, in which the N atom lifts the Ga atom of the GaNH unit off the surface layer during migration, and show that the lifted Ga abstracts an H adatom from the surface, events invisible to static DFT. Single-point UMA calculations on FPMD snapshots reproduce the first-principles relative energies along the trajectory (RMSE of about 8.5 meV/atom) without any retraining. Long-time MLIP-based MD (150 ps) then reveals dynamics never observed within the FPMD window: site-to-site H-adatom hopping, which gates the migration paths of the growth unit, and reversible dissociation of the GaNH unit into independently migrating Ga and NH adspecies. This work constitutes, to our knowledge, the first application of an MLIP to the molecular dynamics of GaN MOVPE.

Figures

Figures reproduced from arXiv: 2607.23461 by the authors.

Figure 1
Figure 1. Top views of the GaN(0001) slab model (Ga: brown, N: blue, H: white); the black frame denotes the 4×4 supercell. Top: initial configuration with the GaNH admolecule and nine H adatoms. Bottom: migration trajectories of Gaadmol (green) and the N atom of the GaNH unit (magenta) during the 1.5 ps FPMD run, superposed on the final configuration. algorithm23 (force convergence 0.05 eV/Å) in ASE24, and the relaxed structu… view at source ↗
Figure 2
Figure 2. Side-view snapshots from the FPMD trajectory showing the lifting-assisted migration of the GaNH admolecule on GaN(0001) (Ga: brown, N: blue, H: white). The N atom of the GaNH unit lifts Gaadmol off the topmost surface layer (top panel) and the unit migrates while holding it in the elevated configuration (middle panel), after which the unit settles at a new site (bottom panel). Snapshots at t = 0.2, 0.4, and 0.8 ps (… view at source ↗
Figure 3
Figure 3. Side-view snapshots from the FPMD trajectory showing hydrogen abstraction by the lifted Gaadmol (Ga: brown, N: blue, H: white). The lifted Gaadmol captures a neighboring H adatom (top panel) and abstracts it from the surface (bottom panel). Snapshots at t ≈ 1.41 and 1.50 ps (top and bottom) from the FPMD run at 1273 K. essentially the identity line (slope 1.001) with only a small positive offset of ∼7 meV/atom. It i… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Reversible dissociation and re-formation of the GaNH unit observed in the 150-ps MLIP-based MD run. The distance between Gaadmol and the N atom of the unit is shown as a function of time (sampled every 10 fs, minimum-image convention). Shaded intervals denote the disso…
Figure 4
Figure 4. Figure 4: Validation of the UMA potential against first-principles data, resolved by the bonding state of Gaadmol. (a) Parity plot of the relative potential energies (in meV/atom, referenced to the initial structure) of 751 snapshots sampled every 2 fs along the 1.5 ps FPMD traj…
Figure 6
Figure 6. Figure 6: Top views of migration trajectories from the MLIP-based MD run (Ga: brown, N: blue, H: white); the black frame denotes the 4×4 supercell. Top: dissociated period following the first dissociation event (23.4–30 ps); the Ga adatom (green) and the NH admolecule (magenta; …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references

  1. [1]

    Hiroshi Amano. Nobel lecture: Growth of GaN on sap- phire via low-temperature deposited buffer layer and 6/8 realization of p-type GaN by Mg doping followed by low-energy electron beam irradiation.Rev. Mod. Phys., 87:1133–1138, 2015

  2. [2]

    Amano, Y

    H. Amano, Y . Baines, E. Beam, et al. The 2018 GaN power electronics roadmap.J. Phys. D: Appl. Phys., 51:163001, 2018

  3. [3]

    A new theoretical ap- proach to adsorption–desorption behavior of Ga on GaAs surfaces.Surf

    Yoshihiro Kangawa, Tomonori Ito, Akira Taguchi, Kenji Shiraishi, and Tadashi Ohachi. A new theoretical ap- proach to adsorption–desorption behavior of Ga on GaAs surfaces.Surf. Sci., 493:178–181, 2001

  4. [4]

    Thermodynamic analy- sis of (0001) and (000¯1) GaN metalorganic vapor phase epitaxy.Jpn

    Akira Kusaba, Yoshihiro Kangawa, Pawel Kempisty, Hec- tor Valencia, Kenji Shiraishi, Yoshinao Kumagai, Koichi Kakimoto, and Akinori Koukitu. Thermodynamic analy- sis of (0001) and (000¯1) GaN metalorganic vapor phase epitaxy.Jpn. J. Appl. Phys., 56:070304, 2017

  5. [5]

    Re- action pathway of surface-catalyzed ammonia decompo- sition and nitrogen incorporation in epitaxial growth of gallium nitride.J

    Kieu My Bui, Jun-Ichi Iwata, Yoshihiro Kangawa, Kenji Shiraishi, Yasuteru Shigeta, and Atsushi Oshiyama. Re- action pathway of surface-catalyzed ammonia decompo- sition and nitrogen incorporation in epitaxial growth of gallium nitride.J. Phys. Chem. C, 122:24665–24671, 2018

  6. [6]

    First- principle study of ammonia decomposition and nitrogen incorporation on the GaN surface in metal organic vapor phase epitaxy.J

    Kieu My Bui, Jun-Ichi Iwata, Yoshihiro Kangawa, Kenji Shiraishi, Yasuteru Shigeta, and Atsushi Oshiyama. First- principle study of ammonia decomposition and nitrogen incorporation on the GaN surface in metal organic vapor phase epitaxy.J. Cryst. Growth, 507:421–424, 2019

  7. [7]

    Gallium–gallium weak bond that incorporates nitrogen at atomic steps during GaN epitaxial growth.Appl

    Kieu My Bui, Kenji Shiraishi, and Atsushi Oshiyama. Gallium–gallium weak bond that incorporates nitrogen at atomic steps during GaN epitaxial growth.Appl. Surf. Sci., 557:149542, 2021

  8. [8]

    Exploration of a large-scale re- constructed structure on GaN(0001) surface by Bayesian optimization.Appl

    Akira Kusaba, Yoshihiro Kangawa, Tetsuji Kuboyama, and Atsushi Oshiyama. Exploration of a large-scale re- constructed structure on GaN(0001) surface by Bayesian optimization.Appl. Phys. Lett., 120:021602, 2022

Show all 25 references
  1. [9]

    Bowler, and Akira Kusaba

    Karol Kawka, Paweł Kempisty, Konrad Sakowski, Stanisław Krukowski, Michał Bo ´ckowski, David R. Bowler, and Akira Kusaba. Augmentation of the elec- tron counting rule with Ising model.J. Appl. Phys., 135:225302, 2024

  2. [10]

    Insight into the step flow growth of gallium nitride based on density functional theory.Appl

    Kieu My Bui, Kenji Shiraishi, and Atsushi Oshiyama. Insight into the step flow growth of gallium nitride based on density functional theory.Appl. Surf. Sci., 613:155840, 2023

  3. [11]

    A two-dimensional liquid-like phase on Ga- rich GaN(0001) surfaces evidenced by first principles molecular dynamics.Jpn

    Kieu My Bui, Mauro Boero, Kenji Shiraishi, and Atsushi Oshiyama. A two-dimensional liquid-like phase on Ga- rich GaN(0001) surfaces evidenced by first principles molecular dynamics.Jpn. J. Appl. Phys., 59:SGGK04, 2020

  4. [12]

    An atomistic insight into reactions and free-energy pro- files of NH3 and Ga on GaN surfaces during the epitaxial growth.Appl

    Mauro Boero, Kieu My Bui, Kenji Shiraishi, Kana Ishisone, Yoshihiro Kangawa, and Atsushi Oshiyama. An atomistic insight into reactions and free-energy pro- files of NH3 and Ga on GaN surfaces during the epitaxial growth.Appl. Surf. Sci., 599:153935, 2022

  5. [13]

    Generalized neural- network representation of high-dimensional potential- energy surfaces.Phys

    Jörg Behler and Michele Parrinello. Generalized neural- network representation of high-dimensional potential- energy surfaces.Phys. Rev. Lett., 98:146401, 2007

  6. [14]

    Deringer, Miguel A

    V olker L. Deringer, Miguel A. Caro, and Gábor Csányi. Machine learning interatomic potentials as emerging tools for materials science.Adv. Mater., 31:1902765, 2019

  7. [15]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)- equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nat. Commun., 13:2453, 2022

  8. [16]

    A universal graph deep learning interatomic potential for the periodic table.Nat

    Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table.Nat. Comput. Sci., 2:718–728, 2022

  9. [17]

    Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. InAdvances in Neural Information Processing Systems, volume 35, 2022

  10. [18]

    Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R

    Brandon M. Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R. Kitchin, Daniel S. Levine, Kyle Michel, Anuroop Sriram, Taco Co- hen, Abhishek Das, Ammar Rizvi, Sushree Jagriti Sahoo, Zachary W. Ulissi, a...

  11. [19]

    Le, and Paulette Clancy

    Xiangyu Chen, Nam Q. Le, and Paulette Clancy. Diffusion-limited crystal growth of gallium nitride using active machine learning.Cryst. Growth Des., 24:2855– 2863, 2024

  12. [20]

    Clark, Matthew D

    Stewart J. Clark, Matthew D. Segall, Chris J. Pickard, Phil J. Hasnip, Matt I. J. Probert, Keith Refson, and Mike C. Payne. First principles methods using CASTEP. Z. Kristallogr., 220:567–570, 2005

  13. [21]

    Perdew, Kieron Burke, and Matthias Ernzerhof

    John P. Perdew, Kieron Burke, and Matthias Ernzerhof. Generalized gradient approximation made simple.Phys. Rev. Lett., 77:3865–3868, 1996

  14. [22]

    Thompson, H

    Aidan P. Thompson, H. Metin Aktulga, Richard Berger, Dan S. Bolintineanu, W. Michael Brown, Paul S. Crozier, Pieter J. in ’t Veld, Axel Kohlmeyer, Stan G. Moore, Trung Dac Nguyen, Ray Shan, Mark J. Stevens, Julien Tranchida, Christian Trott, and Steven J. Plimpton. LAMMPS – a ...

  15. [23]

    Structural relaxation made simple.Phys

    Erik Bitzek, Pekka Koskinen, Franz Gähler, Michael Moseler, and Peter Gumbsch. Structural relaxation made simple.Phys. Rev. Lett., 97:170201, 2006. 7/8

  16. [24]

    The atomic simulation environment—a Python library for working with atoms.J

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, et al. The atomic simulation environment—a Python library for working with atoms.J. Phys.: Condens. Matter, 29:273002, 2017

  17. [25]

    Visualization and analysis of atom- istic simulation data with OVITO–the Open Visualiza- tion Tool.Modelling Simul

    Alexander Stukowski. Visualization and analysis of atom- istic simulation data with OVITO–the Open Visualiza- tion Tool.Modelling Simul. Mater. Sci. Eng., 18:015012, 2010. 8/8

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.