{"id":"226103a4-2729-4e11-909a-1acfe33dff62","arxiv_id":"2607.24871","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A kinetic Monte Carlo framework simulates GaN molecular-beam epitaxy using catalog and on-the-fly machine-learned barriers, reproducing triangular islands and proposing a desorption-driven island-walking regime.","lead":"This paper describes a computer-simulation framework for growing gallium nitride (GaN) crystals atom by atom, tracking how atoms land, move, attach, and evaporate during molecular beam epitaxy. It pairs conventional energy-barrier tables with machine-learning-based on-the-fly barrier calculations and predicts a striking 'walking island' behavior at high temperatures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated BEP+CHGNet OTF barriers: the +0.14 eV Ga2N correction that drives the claimed morphology is within the α-sensitivity and CHGNet's unquantified surface error, so the predictive claim is unsupported without DFT/NEB checks.","rationale":"The paper explicitly acknowledges the ambiguity in α in SI S1 but does not validate CHGNet's ΔE against DFT or NEB for the specific step-edge configurations. The consistency statement in §3.2 ('cluster within ±0.01 eV') is circular: the BEP formula with ΔE≈0 returns E0 by construction, so it cannot validate the method. The three configurations that do show nonzero ΔE are precisely those that test the framework, and they are unverified. The claimed morphological impact (fewer, larger islands; extended AdGa patches) is attributed to barrier changes of 0.09–0.14 eV, which are the same magnitude as the spread from the α sensitivity and plausibly the same magnitude as CHGNet's error on open surfaces. A single set of DFT-NEB calculations for these three hops would settle whether the OTF predictions are accurate. This concern is load-bearing because the paper's title and abstract present on-the-fly barrier evaluation as the principal advance; if the barriers are wrong, the framework's predictive claim reduces to that of the predefined catalog, which is already known. The reader's verdict of CONDITIONAL correctly captures this; our analysis agrees. Therefore, the verdict should remain unchanged.","tokens_in":13656,"tokens_out":8814,"duration_ms":82266,"concrete_test":"Compute DFT-NEB barriers, using the same slab model and functional as Chugh et al. (2017), for the three OTF-corrected configurations: Ga2N with nn2=2 (predicted 1.59 eV), Ga2AdGa (predicted 0.514 eV), and Ga2N,1AdGa (predicted 1.657 eV). Compare each to the BEP-predicted barrier. If the mean absolute deviation exceeds ~0.1 eV, the OTF corrections—and the morphological differences attributed to them—are unsupported. Additionally, compare CHGNet ΔE values to DFT reaction energies for the same hops.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that OTF-KMC with BEP+CHGNet captures environment-dependent step-edge kinetics rests on E_a = E0 + αΔE (§3.2), with E0 from the DFT catalog and ΔE from CHGNet relaxations. The only reported consistency check—novel barriers clustering within ±0.01 eV of E0 when ΔE≈0—is tautological because BEP maps ΔE=0 back to E0. The three asymmetric configurations that do differ (Ga2N +0.14 eV, Ga2AdGa −0.09 eV, Ga2N,1AdGa +0.11 eV) are exactly where the unvalidated ΔE and assumed α=0.5 matter. Table S1 shows varying α∈[0.25,0.75] shifts these barriers by 0.086–0.14 eV—the same order as the corrections. CHGNet is bulk-trained with no demonstrated accuracy for GaN(0001) adatoms, step edges, or N vacancies; its surface reaction energies likely carry errors ~0.1 eV. If so, the 'fewer, larger islands' and 'extended AdGa patches' in Fig. 9 are not robust predictions but artifacts of BEP linearity and CHGNet's unknown surface error. No DFT/NEB validation for the specific novel configurations is provided, so the OTF results are not yet predictive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a lattice-based kinetic Monte Carlo framework for simulating GaN(0001) MBE growth. The framework operates in two modes: a predefined-barrier mode using a DFT-derived coordination catalog, and an on-the-fly (OTF) mode that computes activation barriers for local environments absent from the catalog using a Brønsted–Evans–Polanyi (BEP) relation with CHGNet reaction energies. Simulations with predefined barriers reproduce compact triangular islands, Ostwald ripening, Ehrlich–Schwoebel barrier-induced multilayer nucleation, and, at high temperature with desorption, an 'island walking' regime attributed to N–Ga exchange and preferential AdGa desorption. OTF simulations produce slightly fewer, larger islands and more extended AdGa patches than the predefined-barrier runs. The paper claims the framework is a flexible platform for predictive atomic-scale simulations of GaN epitaxy.","tokens_in":14103,"tokens_out":4295,"duration_ms":40774,"significance":"If the OTF barrier evaluation were properly validated, the framework would be a valuable contribution: it extends environment-aware KMC to a compound semiconductor, introduces a configurable descriptor that resolves step-edge vs. terrace environments, and demonstrates multiphenomenon morphology in a single code. The caching database and MPI-parallel architecture are practical strengths. However, the current manuscript is essentially a method demonstration with qualitative morphological outputs. The central novelty—adaptive OTF barrier evaluation—rests on a BEP correction that is not tested against DFT/NEB for the specific environments that drive the reported morphology changes. The established qualitative behaviors (triangular islands, ripening, ES mounds) are consistent with prior work and are not in themselves evidence of predictive power. The framework has potential, but the quantitative validation needed to support the 'predictive' claim is missing.","major_comments":[{"comment":"The OTF barrier is computed as E_a = E0 + αΔE with E0 taken from the same nn1 catalog and α=0.5 assumed. The only consistency check reported—that novel barriers with ΔE≈0 coincide with E0—is tautological because the BEP form forces this limit. The three configurations with nonzero ΔE (Ga2N +0.14 eV, Ga2AdGa −0.09 eV, Ga2N,1AdGa +0.11 eV) are exactly where the assumed α and CHGNet's unquantified surface error matter. SI Table S1 shows that varying α∈[0.25,0.75] changes these barriers by 0.086–0.14 eV, the same order as the claimed corrections. CHGNet is bulk-trained with no demonstrated accuracy for GaN(0001) adatoms, step edges, or N vacancies. Without DFT/NEB reference calculations for at least these configurations, the OTF corrections and the resulting morphological differences in Fig. 9 are unsupported. This undermines the central claim that OTF barriers capture environment-dependent","section":"§3.2, SI S1, Eq. (BEP relation)"},{"comment":"The paper repeatedly describes the framework as 'predictive' and concludes that predicted morphologies are 'directly testable by in situ STM or post-growth AFM,' but no quantitative validation against experiments or against an independent benchmark is provided. All comparisons are qualitative (e.g., triangular islands 'consistent with' STM observations; ripening 'characteristic of' Ostwald ripening). There are no comparisons of island densities, island size distributions, or temperature-dependent scaling laws. Given the large number of hand-set parameters (ES barriers, desorption bases and increments, N2 desorption barrier, flux ratio, α), the absence of any quantitative benchmark makes it impossible to assess whether the model's predictions are accurate or merely plausible. At minimum, one quantitative comparison—e.g., island density vs. temperature or a known scaling law—is needed to s","section":"§3.1–3.2, Conclusions"},{"comment":"The N2 associative desorption rule ('whenever two adjacent surface N atoms are both free of Ga coverage above them, fixed 1.5 eV barrier') and the coordination-dependent desorption barrier hierarchy (N>Ga>AdGa with fixed per-neighbor increments) are introduced as model assumptions. If the 'island walking' regime and the claimed stoichiometric drain are offered as physical predictions, these assumptions need at least a sensitivity analysis or a comparison to dedicated experimental/DFT data. Without this, the walking regime should be framed as a model-dependent scenario rather than a validated prediction. This is especially relevant because the paper's abstract highlights the walking regime as one of its phenomena.","section":"§3.1.4, §2 (desorption and N2 model)"}],"minor_comments":[{"comment":"Typo in 'Brønsted–Evansf–Polanyi'; should read 'Brønsted–Evans–Polanyi'.","section":"§3.2"},{"comment":"The phrase 'their coincidence with predefined barriers for most event types is a physical consequence of near-zero BEP reaction energies (ΔE≈0)' is misleading. This is a mathematical consequence of the BEP form, not a physical validation of the barrier catalog; please rephrase to avoid overclaiming.","section":"§3.2, Fig. 8 caption"},{"comment":"The abstract and methods state that barriers are computed via NEB, PES, or BEP, but the simulations reported here use only the BEP route. Please clarify that the other methods are available but not exercised in this work.","section":"Abstract, §2"},{"comment":"The ES barrier values (0.15/0.25 eV) are described as 'conservative lower bounds relative to previously reported values', but the references (7,30,31) are not specific about the ranges. Please give the literature values explicitly so the reader can see the factor being used.","section":"§3.1.3"},{"comment":"The configuration-space analysis is useful, but the tables would benefit from a short explanation of why the 'nn1 only' counts differ between Tables S2 (20) and S3 (29) and how the predefined catalog of 21 barriers arises from 20 fingerprints.","section":"SI S2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible method presentation, but the central OTF claim is not yet validated. A focused DFT/NEB benchmark of a handful of step-edge configurations (especially Ga2N, Ga2AdGa, and Ga2N,1AdGa) would largely address the main criticism and would make this a strong contribution. The qualitative morphology results are nice but do not by themselves justify 'predictive.' I recommend major revision rather than rejection because the framework is clearly useful and the gaps are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does one genuinely new thing—on-the-fly barrier evaluation with MLIP/BEP within a lattice KMC for GaN—and builds a framework that should be useful to the community. But the central quantitative claims ride on unvalidated BEP+CHGNet barriers, and the agreement between OTF and predefined barriers is partly built into the BEP form.\n\nThe predefined-barrier simulations reproduce known results: triangular islands, Ostwald ripening, ES-induced mounds. Those are necessary sanity checks, not new science. The island-walking regime driven by N–Ga exchange and AdGa desorption is a nice qualitative prediction, though it rests on hand-set desorption parameters.\n\nThe OTF extension is the real contribution. Extending the fingerprint from nn1 to nn1×nn2 (140/497 configurations) and computing barriers on demand with BEP and a cached database is a legitimate extension of Trushin et al.'s self-learning idea to a compound semiconductor. The MPI architecture is sensible.\n\nNow the soft spots. The BEP relation Ea = E0 + αΔE uses E0 from the DFT catalog and ΔE from CHGNet. The three significant corrections (Ga2N +0.14 eV, Ga2AdGa −0.09 eV, Ga2N,1AdGa +0.11 eV) are precisely where ΔE matters, and they drive the morphology changes in Fig. 9. But there is no DFT/NEB validation for these configurations. Table S1 shows α=0.25–0.75 changes barriers by 0.086–0.14 eV, the same size as the corrections. CHGNet's accuracy on GaN(0001) step edges is not quantified. So the 'fewer, larger islands' prediction is not yet robust.\n\nThe paper's own consistency check—novel barriers clustering within ±0.01 eV of E0—is tautological: BEP with ΔE≈0 maps back to E0. It does not validate the nn1 catalog or the OTF procedure.\n\nAlso, the desorption parameters, ES barriers, and flux ratio are set by hand, and no quantitative comparison to experiment (island density, size distribution, temperature trends) is provided. That's acceptable for a framework paper, but the authors should stop short of claiming predictive status.\n\nRecommendation: send to peer review. The framework is worth presenting to the community, and the OTF concept needs referee scrutiny. Require (1) DFT/NEB checks of the three key barriers, (2) an error estimate for CHGNet on surfaces, and (3) code release as a condition of acceptance. The paper is not a breakthough, but it's a solid contribution with one big unvalidated assumption.","headline":"A solid framework paper whose main new feature—MLIP/BEP on-the-fly barriers—is plausible but unvalidated; the OTF morphology changes fall within the uncertainty of α and CHGNet.","tokens_in":14544,"tokens_out":3505,"would_cite":false,"duration_ms":34155,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A lattice kinetic Monte Carlo framework for GaN growth now computes step-edge barriers on the fly, revealing environment-dependent corrections that change predicted island morphology.","keywords":["kinetic Monte Carlo","GaN(0001) growth","molecular beam epitaxy","on-the-fly barrier evaluation","Brønsted–Evans–Polanyi","machine-learned interatomic potential","Ehrlich–Schwoebel barrier","Ostwald ripening"],"falsifier":"Compute full nudged elastic band barriers directly for the Ga2N step-edge configuration (nn2Ga = 2) using a reliable DFT functional and compare with the predicted 1.59 eV; a discrepancy larger than roughly 0.1 eV would indicate the BEP-CHGNet estimate is not capturing the step-edge stabilization. Alternatively, measure island size distributions in GaN MBE growth under slightly Ga-rich conditions and compare against the predicted shift to fewer, larger islands.","tokens_in":13574,"feed_emoji":"⚛️","tokens_out":2919,"duration_ms":30913,"temperature":0.7,"pith_summary":"The paper tries to establish that a lattice kinetic Monte Carlo (KMC) framework can simulate GaN(0001) molecular beam epitaxy across multiple growth regimes, and that adding adaptive on-the-fly barrier evaluation—where activation energies are computed from the local atomic environment rather than pulled from a fixed catalog—captures step-edge physics that nearest-neighbor-only catalogs miss. Using predefined barriers, the framework reproduces compact triangular islands, Ostwald ripening during growth interruptions, Ehrlich–Schwoebel-barrier-induced multilayer nucleation, and a desorption-driven 'island walking' regime at high temperature. With on-the-fly barriers, it finds that specific step-edge configurations carry distinctly different activation energies (for example, a Ga2N step-edge hop is 0.14 eV higher than the flat-terrace value), and these corrections lead to morphologically different predictions such as fewer, larger islands and extended AdGa-decorated step edges. The value of the work is in showing that environment-aware barriers are both computable and consequential for GaN growth morphology.","feed_headline":"On-the-fly barriers reshape simulated GaN islands","feed_subtitle":"Nearest-neighbor-only catalogs miss 0.14 eV step-edge corrections; machine-learned barriers predict fewer, larger islands.","key_machinery":"The central mechanism is the nn1×nn2 species-resolved coordination fingerprint plus the on-the-fly barrier-evaluation workflow. The fingerprint encodes the three tetrahedral bonding sites (nn1) and six same-sublattice lateral neighbors (nn2) around a diffusing atom, allowing the framework to distinguish terrace, step-edge, and kink geometries that share the same nn1 coordination. When a previously unseen fingerprint is encountered, the activation barrier is estimated via the Brønsted–Evans–Polanyi relation E_a = E_0 + αΔE, with α = 0.5 and ΔE computed by the machine-learned potential CHGNet, then cached in an in-memory dictionary and persistent database for reuse. This BEP-plus-MLIP approach","core_discovery":"The paper's central claim is that extending the local coordination descriptor from the first-neighbor shell (nn1) to include the six same-sublattice lateral neighbors (nn2) creates a configuration space of up to 140 or 497 distinct fingerprints, and that evaluating these barriers on the fly—rather than precomputing them all—makes environment-aware simulation feasible. When novel configurations arise, the framework computes activation barriers using a Brønsted–Evans–Polanyi relation E_a = E_0 + αΔE with α = 0.5, using CHGNet machine-learned reaction energies ΔE and reference barriers E_0 from the DFT catalog. This reveals that most novel barriers cluster within ±0.01 eV of the predefined valu","pith_inferences":["The BEP sensitivity analysis shows that the predicted barrier corrections vary by 0.086–0.14 eV across α = 0.25–0.75, which is the same order as the corrections themselves; direct NEB validation on the novel step-edge configurations would sharpen the quantitative claims.","The caching strategy effectively makes the framework a self-learning KMC for compound semiconductors, where the barrier database grows as the simulation roughens—this suggests a path toward predictive growth simulation without exhaustive precomputation.","An experimentally testable extension is to compare the predicted island density and size distribution against in situ STM or post-growth AFM on GaN(0001) grown under slightly Ga-rich conditions; the framework predicts measurably larger, more elongated islands than a nearest-neighbor-only model."],"forward_implications":["Predefined nn1-only barrier catalogs systematically miss environment-dependent effects at step edges and kinks; the on-the-fly extension resolves these, producing quantitatively different step-edge kinetics.","If the OTF corrections are right, GaN growth simulations will predict fewer and larger islands with more elongated perimeters under Ga-rich conditions, and more extended AdGa accumulation when ES barriers are active.","The framework can model realistic MBE shutter sequences by explicitly simulating growth interruptions, capturing Ostwald ripening and post-deposition relaxation that continuous-flux models omit.","At high temperature, desorption-driven island walking emerges as a distinct morphological regime, with island translation rates set by the local balance of N–Ga exchange and N2 desorption.","Because the fingerprint and OTF machinery are species-generic, the approach should transfer to other compound semiconductors and nonequilibrium growth problems beyond GaN."],"fun_headline_variants":["Adaptive barrier evaluation yields larger GaN islands","On-the-fly barriers make GaN simulation more accurate","Second-neighbor barriers change GaN island growth","Machine-learned barriers refine GaN epitaxy model","Fewer, larger islands from adaptive GaN barriers"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the BEP estimate E_a = E_0 + αΔE with α = 0.5, using CHGNet reaction energies, gives accurate activation barriers for previously unencountered step-edge configurations; if the machine-learned reaction energies or the linear BEP form are unreliable for these GaN(0001) edge environments, the OTF corrections and the resulting morphology changes are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive barrier evaluation yields larger GaN islands","On-the-fly barriers make GaN simulation more accurate","Second-neighbor barriers change GaN island growth","Machine-learned barriers refine GaN epitaxy model","Fewer, larger islands from adaptive GaN barriers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1560,"prompt_tokens":794,"completion_tokens":766,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":691}},"tokens_in":538,"tokens_out":766,"duration_ms":7686,"temperature":1.0,"reasoning_tokens":691,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:34:17.416873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute full nudged elastic band barriers directly for the Ga2N step-edge configuration (nn2Ga = 2) using a reliable DFT functional and compare with the predicted 1.59 eV; a discrepancy larger than roughly 0.1 eV would indicate the BEP-CHGNet estimate is not capturing the step-edge stabilization. Alternatively, measure island size distributions in GaN MBE growth under slightly Ga-rich conditions and compare against the predicted shift to fewer, larger islands.","supporting_citations":[],"review_version":1}