{"id":"2d5be66c-1b06-4675-a818-99d9b068a215","arxiv_id":"2607.23461","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FPMD plus the universal MLIP UMA reveal lifting-assisted GaNH migration, H-gated paths, and reversible GaNH dissociation on GaN(0001) at MOVPE temperature.","lead":"Simulations of GaN growth surfaces show a GaNH growth unit that lifts its Ga atom while migrating, hops hydrogen that opens and closes paths, and reversibly splits into separate Ga and NH pieces. The work shows a general machine-learned potential can track these events for 150 ps after checking against short first-principles runs.","discovery_kind":"new_application","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The rare-event claims rest on validation along the FPMD trajectory, not on the local forces/barriers that create dissociation and H-gated hops; the state-dependent Ga–Ga offset could invert those rare-event energetics.","rationale":"This is the same soft spot identified by the reader, sharpened from “relative-energy RMSE may be insufficient” to the specific mechanism by which it can fail: global per-atom energy parity along a short FPMD trajectory does not constrain the local force/barrier errors that control rare dissociation and H-gated migration, and the detected Ga–Ga branch offset sits exactly on the controlling bond type. The paper deserves credit for reporting the branch split rather than hiding it, for using a public UMA/LAMMPS stack with methods-level detail, for dynamically cross-checking that the lifting-assisted mode persists under UMA with similar lifted/engaged populations, and for candidly noting that H abstraction did not recur in 150 ps. Those points make REJECT too strong: the FPMD results stand, and the UMA results are plausible rather than asserted without evidence. But ACCEPT is not warranted until a force/barrier-level check on UMA-sampled rare-event configurations is done. The proposed CASTEP force/NEB audit is decisive because it tests the exact quantity MD uses—forces—and the exact kinetic quantities—barriers and basin stability—for the claimed new processes, rather than relying on trajectory-wide energy correlation. Until then the reader’s CONDITIONAL verdict and medium correctness risk remain appropriate.","tokens_in":10705,"tokens_out":3301,"duration_ms":95830,"concrete_test":"Take UMA-only frames around the first dissociation/re-formation (≈23.4/43.9 ps), the H-hop/gating sequence (≈105–113 ps), and several lifted↔engaged transitions; build CI-NEB paths in UMA, then recompute CASTEP single-point energies/forces for endpoints and saddles and DFT-relax/NEB the same configurations. Compare max/RMS force errors on Gaadmol, N, nearby H and surface Ga, force signs along the reaction coordinate, associated-vs-dissociated minima, and barrier heights. If DFT confirms the same basins and UMA forces/barriers agree within ~0.1 eV with unchanged signs, the concern does not land; if DFT collapses dissociated states, reverses forces, or shifts barriers enough to change 1273 K rates by orders, the long-time claims are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The FPMD observations themselves are direct and reasonably supported. The load-bearing gap is in the extrapolation from them to the 150 ps UMA events. The validation uses 751 snapshots every 2 fs from the same 1.5 ps CASTEP trajectory and compares only single-point relative energies per atom. That metric can look excellent—RMSE 8.5 meV/atom, R2 0.946, slope ≈1—because it is dominated by large whole-slab thermal fluctuations, while MD is driven by forces and rare events by local curvatures and transition states that the 1.5 ps run may never visit. The paper’s own branch analysis strengthens rather than settles this concern: residuals split into lifted vs surface-engaged configurations, with the weak Ga–Ga-contact branch offset by ≈10 meV/atom (y=0.995x+11.5). That is precisely the bonding motif controlling GaNH association/dissociation, Gaad migration, and H abstraction. A bias that over/under-stabilizes Gaad+NHad relative to GaNH, or flattens the relevant local barrier, could generate repeated “reversible dissociation” or H-path gating in UMA even if global parity remains good. The dissociation/re-formation events are also called visually from one trajectory, and the FPMD H-abstraction event does not recur under UMA, so recurrence statistics are absent.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","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.","tokens_in":11033,"tokens_out":3224,"duration_ms":123081,"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":[{"comment":"§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","section":"§Validation of the UMA potential; Fig. 4-5"},{"comment":"§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.","section":"§Methods: Machine-learning interatomic potential"},{"comment":"§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","section":"§Results: hydrogen abstraction (Fig. 3) and its non-recurrence"},{"comment":"§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.","section":"§Methods: event identification; Fig. 5"}],"minor_comments":[{"comment":"§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.","section":"§Validation of the UMA potential"},{"comment":"§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.","section":"§Methods"},{"comment":"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.","section":"Fig. 4(b); §Methods"},{"comment":"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.","section":"§Long-time dynamics revealed by MLIP-based MD"},{"comment":"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).","section":"Fig. 5"},{"comment":"§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.","section":"§Methods: First-principles molecular dynamics"},{"comment":"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.","section":"§Methods"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a good fit for the journal's scope. The reference list leans on the authors' own research program and its close collaborators (Refs. 3-12), but these papers genuinely supply the surface models and static-pathway context, so this is not a concern for acceptance. My main hesitation is that the two-headline long-time phenomena each rest on a single 150 ps trajectory whose potential was validated only along one 1.5 ps FPMD run; the requested DFT spot-checks on MLIP-generated geometries seem essential and inexpensive."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new content is dynamical, not conceptual. Short CASTEP FPMD shows GaNH migrating by N lifting Gaadmol off the top layer, plus one clear H-abstraction event by the lifted Ga—neither visible to static pathway searches. UMA single-points on 751 snapshots match relative energies at RMSE ~8.5 meV/atom, R² 0.946, near-unit slope, no retraining. The 150 ps UMA run then shows repeated H hops that open/close paths and reversible GaNH ⇌ Gaad + NHad with independent migration and re-formation. That is the first claimed MLIP MD of GaN MOVPE and a useful reframing of the old “hand-in-hand” picture as intermittent partnership plus H gating.\n\nWhat they do well: explicit trajectories, a parity plot resolved by bonding state, honest reporting that H abstraction did not recur under UMA, and low circularity (UMA is not fitted to these trajectories; prior author DFT supplies context, not forced identities). Methods are clear enough to attempt reproduction.\n\nSoft spots are real but scoped. Validation is single-point energies along one 1.5 ps FPMD trajectory, not forces or barriers on the dissociation/H-hop coordinates. Their own branch analysis shows a ~10 meV/atom offset on the weak surface Ga–Ga (surface-engaged) contacts—the motif that controls association and migration. That offset could bias rare-event sampling even while global parity looks good. Events are called from one long trajectory by visual inspection; no rates, no ensemble, no error bars. Free parameters (3.3 Å Ga–Ga cut, residual threshold) are stated. None of this sinks the FPMD observations or the workflow demonstration; it does mean the long-time kinetic picture is provisional.\n\nThis is for people who already care about GaN/III-nitride growth atomistics or who want a concrete test of universal MLIPs on a chemically messy high-T surface. Worth a serious referee. I would engage, cite the dynamical modes and the validation protocol, and push for ensembles plus a tighter handle on the Ga–Ga offset before treating the dissociation frequencies as settled.","headline":"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.","tokens_in":12130,"tokens_out":557,"would_cite":true,"duration_ms":11316,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"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.","keywords":["GaN MOVPE","machine-learning interatomic potential","molecular dynamics","GaNH admolecule","surface diffusion","hydrogen hopping","universal MLIP","GaN(0001)"],"falsifier":"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.","tokens_in":11932,"feed_emoji":"⚛️","tokens_out":996,"duration_ms":17245,"temperature":0.7,"pith_summary":"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.","feed_headline":"GaN growth units lift, split, and follow hopping hydrogen","feed_subtitle":"A checked universal potential stretches surface dynamics to 150 ps and rewrites the static growth picture","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["MLIP-MD shows GaNH lifting, H-gated paths, and reversible split","GaN growth unit lifts Ga, tracks hopping H, then dissociates","Validated universal potential captures 150-ps GaN adspecies dynamics","H-adatom hops gate GaNH migration and enable reversible breakup","First MLIP dynamics of GaN MOVPE reveal lifting diffusion mode"],"cache_read_input_tokens":128,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["MLIP-MD shows GaNH lifting, H-gated paths, and reversible split","GaN growth unit lifts Ga, tracks hopping H, then dissociates","Validated universal potential captures 150-ps GaN adspecies dynamics","H-adatom hops gate GaNH migration and enable reversible breakup","First MLIP dynamics of GaN MOVPE reveal lifting diffusion mode"]},"model":"grok-4.5","effort":"low","cost_usd":0.004025,"raw_usage":{"total_tokens":1322,"prompt_tokens":868,"num_sources_used":0,"completion_tokens":83,"cost_in_usd_ticks":40248000,"prompt_tokens_details":{"text_tokens":868,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":371,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":868,"tokens_out":83,"duration_ms":8179,"temperature":1.0,"reasoning_tokens":371,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T21:36:59.464819+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"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.","supporting_citations":[],"review_version":1}