{"id":"3a5c08dd-5f62-4469-b696-9229de05a311","arxiv_id":"2607.04680","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"3D-PRIMME learns a local grain-boundary update from two time steps on 100³ voxels and extrapolates to 1024³ domains while preserving coarsening kinetics and topology.","lead":"A neural model learns 3D grain-growth rules from only two consecutive simulation frames on a small grid, then runs the same rule on domains a thousand times larger without retraining. That scale jump could make long-horizon microstructure prediction practical for materials design where full physics solvers are too expensive.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Fixed windows may not stay scale-invariant once mean grain size grows far beyond the training receptive field, even if domain size alone is scaled.","rationale":"The reader correctly isolates the fixed receptive-field assumption as the soft spot under the strongest claim. The paper’s own sensitivity study (Fig. 2) and Discussion show that N_o strongly controls growth rate and that windows should track microstructural length scales; the large-domain experiments (Figs. 4–5) do not stress-test that dependence after substantial coarsening or with mismatched initial grain size. That is a genuine, concrete limitation on interpreting the result as a fully scale-independent local rule, but it does not invalidate the empirical domain-size extrapolation that is actually shown. The appropriate verdict therefore remains CONDITIONAL: useful and largely sound as a surrogate-methods result on MF data, with the fixed-window / length-scale caveat still needing to be closed (or more carefully scoped) before treating scale independence as settled for arbitrary 3D microstructures. No stronger internal inconsistency is required to keep the verdict where the reader placed it.","tokens_in":15397,"tokens_out":639,"duration_ms":5872,"concrete_test":"On the 512^{3} or 1024^{3} rollout, recompute ⟨r⟩^{2} slope, mean faces, and grain-size distributions after the mean grain diameter exceeds ~2–3× the training value (or after N_G drops well below N_G,0/4), comparing fixed N_o=N_a=9 against a re-tuned larger observation window (e.g. N_o=15–21) or an adaptive schedule. If kinetics or topology diverge by more than the ~3% relative ⟨r⟩^{2} error reported for the default windows, the fixed-window scale-independence claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a local evolution operator trained on two consecutive frames of a 100^{3}/512-grain Voronoi microstructure is a scale-independent rule that remains kinetically and topologically consistent when applied to 256^{3}–1024^{3} domains (§3.3; Abstract). The supporting experiments keep initial mean grain size matched to training (same Voronoi density) and use fixed N_o = N_a = 9 chosen for that length scale (Fig. 2; Table S1). As coarsening proceeds, ⟨r⟩ grows and the fixed observation window covers a shrinking fraction of each grain’s neighborhood; the Discussion itself flags that optimal windows depend on characteristic length scales and lists adaptive windows as future work. Thus the reported “scale independence” is primarily domain-size extrapolation at fixed initial grain size, not a demonstration that the same fixed receptive field remains adequate after substantial coarsening or under different initial grain densities. If the local interface-site metric (Eq. 5) becomes under-resolved relative to the true curvature scale, the claimed transferable local rule can degrade even while early-time statistics still look linear.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces 3D-PRIMME, a local-window neural surrogate for three-dimensional grain growth. Microstructures are mapped to an interface-site representation (Eq. 5) over an observation window of size N_o; a network then predicts state-flip probabilities inside an action window of size N_a from two consecutive MF time steps. Trained on 100³ domains with 512 grains (often a single sequence of two steps), the operator is applied autoregressively and is reported to recover linear ⟨r⟩² coarsening, average face numbers near 13–14, collapsed topology-vs-size statistics, and grain-size distributions, including on domains up to 1024³ with ~550000 grains and on inclination-dependent MF data. Window-size ablations, multi-seed uncertainty, and data-efficiency studies are provided.","tokens_in":15665,"tokens_out":1322,"duration_ms":35149,"significance":"If the local operator is genuinely transferable, the work addresses a real bottleneck: full-field 3D grain-growth surrogates that do not explode in memory with domain size. Strengths include extreme data efficiency (two consecutive frames), explicit multi-seed uncertainty bands (Fig. 3, S1–S2), quantitative window ablations (Table S1), and successful inclination-dependent tests without an explicit inclination feature. The parallel local-update design is a practical contribution for large 3D microstructure simulation. The central scientific claim—scale-independent local evolution—is interesting but currently demonstrated mainly as domain-size extrapolation at matched initial grain density rather than full length-scale invariance.","major_comments":[{"comment":"Abstract and §3.3 claim a “scale-independent” local evolution rule because a model trained on 100³/512 grains applies to 256³–1024³ domains without retraining. In those experiments the initial mean grain size is deliberately matched to training (same Voronoi density); only domain size and grain count change. Fixed N_o = N_a = 9 were chosen for that training length scale (Fig. 2, Table S1). As coarsening proceeds, ⟨r⟩ grows and the fixed observation window covers a shrinking fraction of each grain; the Discussion itself notes that optimal windows depend on characteristic length scales and lists adaptive windows as future work. The reported “scale independence” is therefore primarily domain-size extrapolation at fixed initial grain size, not a demonstration that the same fixed receptive field remains adequate after substantial coarsening or under different initial grain densities. Please e","section":null},{"comment":"All training and evaluation use the mode-filter (MF) model (§2.1). MF is a convenient stochastic surrogate for curvature-driven growth, but the paper’s broader motivation includes experimental 3D microstructures and classical physics-based models (phase-field, Potts). Without at least one cross-model or experimental transfer test, it remains open whether 3D-PRIMME learns transferable grain-boundary physics or primarily the MF update rule. A limited transfer experiment (e.g., train on MF, evaluate kinetics/topology against a phase-field or Potts trajectory with comparable isotropic mobility) would substantially strengthen the central claim; if that is out of scope, the abstract and conclusion should state clearly that the surrogate is validated against MF.","section":null}],"minor_comments":[{"comment":"Eq. (1) and surrounding text mix MF Hamiltonian notation with the learned operator; a short paragraph clarifying that MF generates labels while 3D-PRIMME never sees Γ̂ or u would reduce confusion.","section":null},{"comment":"Fig. 2d and related text correctly note that voxel accuracy decays under stochastic MF vs deterministic PRIMME; still, reporting a topology- or boundary-focused accuracy (e.g., interface Dice) would better separate physical fidelity from trajectory divergence.","section":null},{"comment":"In §3.4, accuracy decreases when going from 10 to 50 sequences; the redundancy explanation is plausible but speculative—briefly state whether early stopping, learning-rate schedule, or batch composition were held fixed across M.","section":null},{"comment":"Table 1 and the architecture description refer to “the architecture from Ref. [7]” without restating layer widths or activation choices; a one-line summary or SI table would aid reproducibility.","section":null},{"comment":"Minor typos and spacing: “Graingrowthisgoverned”-style run-ons appear in the abstract/intro PDF text; “shows successful results” (§1); “The model is operates” (§1).","section":null},{"comment":"Data/code availability is promised “upon publication”; for a methods paper, a temporary anonymous repository or SI checklist of hyperparameters would help reviewers verify the large-domain claims.","section":null}],"recommendation":"major_revision","confidential_remarks":"The technical contribution is real and the experiments are carefully done within the MF setting; the main risk is over-claiming “scale-independent physics” from domain-size extrapolation at fixed grain density. If the authors tighten the claim language and either add a modest cross-model check or clearly scope the validation to MF, this is a solid methods paper for a computational materials / ML-for-science venue. I do not see a fatal circularity; the circularity concern in the reader note is ordinary supervised surrogate evaluation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The load-bearing result is real and useful: a local interface-site operator trained on two frames of a 100^{3}/512-grain MF run preserves linear ⟨r⟩^{2} coarsening, face-number topology near 13–14, and grain-size distributions when rolled out autoregressively on 256^{3}–1024^{3} domains (same initial grain density) and on inclination-dependent MF data. That is the new empirical fact, and the multi-seed uncertainty bands plus window ablations (Fig. 2, Table S1) make it more than a single cherry-picked plot.\n\nWhat they did well is the engineering of locality. By encoding neighborhoods as interface-site counts (Eq. 5) and predicting only action-window flips, they avoid the memory wall that hits full-field or latent-global 3D models. Removing the explicit GB-site regularizer from 2D PRIMME and still getting stable long rollouts is a clean design win; the windows themselves supply the physics constraint. Data efficiency is also genuine—one or ten sequences of two steps is enough—because a single 3D volume already supplies huge numbers of local transition examples.\n\nSoft spots are real but proportionate. The “scale-independent” claim is mainly domain-size extrapolation at matched initial grain size; as ⟨r⟩ grows the fixed N_o = N_a = 9 covers a shrinking fraction of each grain, and the Discussion itself flags adaptive windows as future work. Validation stays inside the MF family, so transfer to phase-field or experiment is untested. Code/data are promised but not yet public, and head-to-head numbers against 3D GNN/FNO baselines are thin. None of these break the central demonstration; they just bound how far the claim currently travels.\n\nThis is for people who need large-domain 3D grain-growth surrogates or who are building local-patch microstructure models. The math and citation pattern look solid (self-cites are methodological background, not circular). I would send it to peer review; a serious referee can tighten the window-scale language and ask for the artifacts. Worth engaging if you work in this space.","headline":"Solid 3D extension of PRIMME: data-efficient local rule that really does roll out to 1024^{3} on MF data, with the usual surrogate caveats.","tokens_in":16367,"tokens_out":557,"would_cite":true,"duration_ms":4968,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A neural model trained on two time steps of a 100-cubed grain-growth simulation extrapolates to million-grain domains without retraining.","keywords":["3D grain growth","microstructure evolution","physics-regulated machine learning","local evolution operator","scale-independent surrogate","inclination-dependent anisotropy","data-efficient learning"],"falsifier":"Run the trained operator on a domain whose average grain diameter has grown well beyond the fixed observation window (or start from a microstructure whose initial grain size is far from the training size) and check whether the linear coarsening rate and the steady average number of faces still match an independent physics-based reference.","tokens_in":16205,"feed_emoji":"🔬","tokens_out":667,"duration_ms":6078,"temperature":0.7,"pith_summary":"Grain growth in metals and ceramics reshapes the internal crystal mosaic that controls strength and corrosion resistance, but accurate three-dimensional simulations become expensive as the domain grows. This paper shows that a neural network can learn a purely local update rule for grain-boundary motion from only two consecutive snapshots on a modest 100-cubed grid containing 512 grains. Once trained, the same rule can be applied, without any retraining, to domains up to 1024 cubed that contain hundreds of thousands of grains, while still recovering the classic linear coarsening law and the expected topological statistics. The authors further show that the same local rule can capture inclination-dependent (anisotropic) boundary migration when the training data themselves are anisotropic. The practical upshot is a data-efficient, memory-light surrogate that can explore large-scale three-dimensional microstructure evolution that would otherwise be prohibitive.","feed_headline":"Two snapshots train a grain-growth model that scales to 550k grains","feed_subtitle":"A local neural rule learned on 100^{3} grids reproduces coarsening kinetics on 1024^{3} domains without retraining.","key_machinery":"The interface-site representation: each voxel is replaced by a local count of neighboring sites that belong to different grains, computed inside a fixed observation window; a second, fixed action window then supplies the neural-network input that predicts which neighboring grain label will occupy the central site at the next step. Because both windows are local, the learned update can be tiled over arbitrarily large grids.","core_discovery":"3D-PRIMME learns a scale-independent, temporally stable local evolution operator for three-dimensional grain growth. Trained on only two consecutive time steps of a 100-cubed microstructure with 512 grains, the operator reproduces linear mean-square-radius coarsening and preserves average face counts and grain-size distributions when applied autoregressively for a hundred steps on domains as large as 1024 cubed containing roughly 550 000 grains.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Two snapshots teach a 3D grain-growth rule that scales to 550k grains","Physics-regulated net learns local grain evolution from 100³ to 1024³","Local operator trained on 512 grains runs autoregressively on 550000","3D-PRIMME: two timesteps yield scale-independent coarsening kinetics","Tiny-domain training yields temporally stable 3D microstructure evolution"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that fixed observation and action windows chosen for the training grain size remain an adequate local receptive field even after grains coarsen by large factors and the domain size jumps by orders of magnitude.","fun_headline_variants_meta":{"raw":{"variants":["Two snapshots teach a 3D grain-growth rule that scales to 550k grains","Physics-regulated net learns local grain evolution from 100³ to 1024³","Local operator trained on 512 grains runs autoregressively on 550000","3D-PRIMME: two timesteps yield scale-independent coarsening kinetics","Tiny-domain training yields temporally stable 3D microstructure evolution"]},"model":"grok-4.5","effort":"low","cost_usd":0.00612,"raw_usage":{"total_tokens":1554,"prompt_tokens":753,"num_sources_used":0,"completion_tokens":105,"cost_in_usd_ticks":61200000,"prompt_tokens_details":{"text_tokens":753,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":696,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":753,"tokens_out":105,"duration_ms":5503,"temperature":1.0,"reasoning_tokens":696,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T15:17:43.367537+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the trained operator on a domain whose average grain diameter has grown well beyond the fixed observation window (or start from a microstructure whose initial grain size is far from the training size) and check whether the linear coarsening rate and the steady average number of faces still match an independent physics-based reference.","supporting_citations":[],"review_version":1}