{"id":"b39980af-f22d-42b5-9068-b517f9aba32f","arxiv_id":"2607.25142","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adaptive GSM on a layered hierarchical triangular mesh solves Allen–Cahn/Cahn–Hilliard with overall second-order accuracy and O(N) cost versus O(N²) uniform FDM for thin-interface 2D problems.","lead":"The paper couples the Gradient Smoothing Method with a hierarchical adaptive triangular mesh so phase-field interfaces are refined only in a thin band. For large 2D problems with thin interfaces this yields near-linear cost versus quadratic cost on uniform grids, while keeping second-order accuracy in practice.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The \"negligible error in flat bulk\" justification for global second-order accuracy is validated only on single-interface benchmarks whose bulk is genuinely near-constant; the long-time multi-interface coarsening regime the method targets has O(1/R) bulk chemical-potential gradients computed on the c","rationale":"Good-faith read: this is a methods paper whose contribution is combinatorial (known GSM + hierarchical longest-edge-bisection AMR + interface-uniform refinement), and the core engineering story — confine fine resolution to a uniform interfacial band so GSM keeps its uniform-mesh accuracy while DOFs scale with interface length — is coherent and honestly presented. Several things check out internally: per-triangle telescoping in Eq. (34) (G1+G2+3 ≡ 0) genuinely delivers global conservation of the C–H flux sum regardless of the shakier per-edge antisymmetry assertion in §3; the O(N) element-count argument (4–5 layers across the interface, Fig. 22(a)) is sound in 2D; and the authors are commendably candid that GSM is ~8× more expensive per DOF than FDM, that FDM wins below ~400×400, and that the crossover is morphology-dependent. I checked whether the conservation identity, the stability constraints (Eqs. 22–23), or the MOOSE baseline protocol (same dt imposed on an implicit FEM code handicaps it) were more load-bearing; none undermines the central accuracy/efficiency claim as directly as the untested bulk-gradient regime, because if bulk errors corrupt interface motion in multi-interface coarsening, both the \"overall second-order accuracy\" claim and the efficiency framing (you would have to refine more of the domain, eroding the O(N) advantage) weaken simultaneously. This is exactly the reader's identified weakest assumption, so I agree; the test above would settle whether it lands. No change to CONDITIONAL: the fix is an additional benchmark and scoped language, not a re-derivation.","tokens_in":25934,"tokens_out":3698,"duration_ms":57289,"concrete_test":"Run a two-droplet Cahn–Hilliard Ostwald-ripening test: two circular droplets of slightly different radii separated by a large bulk region (domain ≫ droplet spacing), with adaptive GSM (N0=21 bulk, n_level set to resolve the interface per Eq. 9) versus uniform-mesh FDM at the same finest h. Track the small droplet's radius R(t) and mass-transfer rate until disappearance. Since ripening is driven entirely by bulk diffusion of μ between droplets, if adaptive GSM's shrinkage time deviates from FDM by more than the discretization-error level seen in §5.3's static test (repeat at two refinement depths to separate h-convergence from bulk-coarsening error), the flat-bulk assumption fails in the multi-interface regime and the accuracy claim must be restricted to single-interface morphologies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's accuracy restoration argument (§5.2, end; §5.3; §6) is: mesh non-uniformity lives in a transition band where the order parameter is nearly constant, so the first-order GSM truncation error there is harmless. The numerical evidence (Fig. 17, Fig. 19) uses one shrinking square (A–C) and one relaxing rectangle (C–H) — morphologies whose bulk is essentially flat, so the premise holds by construction. But for the Cahn–Hilliard equation, which the paper itself identifies as the primary efficiency target (\"practical demand for aggressive acceleration is generally more critical for C–H modeling\", §5.4), interface motion in multi-phase systems is *driven* by bulk diffusion: the chemical potential μ is smooth but non-constant in the bulk (Gibbs–Thomson boundary values, O(1/R) curvature-dependent gradients), and interface velocities are set by bulk ∇²μ. In the adaptive mesh, that bulk Laplacian is computed on the coarsest 21×21 mesh with a stencil that is locally first-order wherever the transition band sits — and μ is smooth and extended, not confined to the refined band, so no refinement indicator (|η_K| < u_thr on the order parameter) will ever refine the μ-gradient region. There is no error analysis (only the six-stencil RMSE study of §5.2 and the single-equilibrium-profile test of §5.3) showing that coarse-mesh, first-order errors in bulk μ do not corrupt interfacial fluxes and hence coarsening rates. Note also that at u_thr = 0.925 the transition band sits where the tanh profile still carries ~14% of its maximum gradient, so even in the tested geometry the \"nearly constant\" premise is approximate. The claim may well survive — smooth errors on a coarse mesh can be small — but the load-bearing assumption is currently asserted and tested only in its friendliest regime.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript presents a strong-form phase-field solver coupling the Gradient Smoothing Method (GSM) with a hierarchical adaptive structured triangular mesh (HAMR) based on longest-edge bisection. The key design is a layered refinement strategy that keeps the interfacial band uniformly fine while confining mesh non-uniformity to a narrow transition band in the bulk, with the aim of preserving global second-order accuracy despite GSM being only first-order on non-uniform stencils. The authors validate the GSM Laplacian on uniform meshes (second-order, matching central-difference FDM), document first-order degradation on six representative non-uniform stencils, and present adaptive Allen–Cahn and Cahn–Hilliard benchmarks (equilibrium interface relaxation, curvature-driven shrinkage/relaxation, and layer retraction with concentration-dependent mobility) showing agreement with FDM and the analytical tanh profile. Timing studies report O(N) per-step cost for the adaptive GSM versus O(N²) for uniform FDM, with a crossover near N≈400, and faster runtimes than MOOSE/FEM under matched conditions. A fully vectorized triangle-wise assembly of the GSM Laplacian is described in algorithmic detail.","tokens_in":26414,"tokens_out":5743,"duration_ms":186725,"significance":"If the accuracy-restoration claim holds in the regimes the method targets, the contribution is practically significant: a matrix-free, strong-form adaptive solver with conservation-by-construction fluxes (the antisymmetric edge-coefficient argument in §3), a reproducible vectorized implementation (Eqs. 24–42), and per-step cost growing with interfacial length rather than domain area would be genuinely useful for large-scale thin-interface phase-field work. The mesh-quality parameter study (§4.5, Figs. 8–10) tying n̄_level and u_thr to the analytical interface thickness (Eq. 9) is a concrete, transferable design rule, and the directional-correction checkerboard demonstration (Fig. 20) is instructive. The validation is against independent references (analytical profiles, FDM, MOOSE), not circular. The main gap is that the headline accuracy claim is demonstrated only in flat-bulk configurations, and the efficiency claim is stated per fixed step count under explicit time stepping; both need qualification or additional evidence before the O(N)/second-order conclusions can be taken as general.","major_comments":[{"comment":"The central accuracy claim — that the HAMR design restores global second-order accuracy because mesh non-uniformity sits in a transition band where the order parameter is nearly constant — is validated only on benchmarks whose bulk is essentially flat (single shrinking square for A–C, single relaxing rectangle for C–H, Figs. 17, 19). For the C–H coarsening regime the paper itself identifies as the primary target (§5.4: 'practical demand for aggressive acceleration is generally more critical for C–H modeling'), interface motion is driven by bulk diffusion: the chemical potential μ is smooth, extended, and non-constant in the bulk (Gibbs–Thomson boundary values, O(1/R) gradients). The refinement indicator |η_K| < u_thr acts on c, so the μ-gradient region in the bulk is never refined, and ∇²μ there is computed on the coarsest 21×21 mesh, partly with the locally first-order stencils of §5.2.","section":"§5.2–5.4, §6 (accuracy restoration argument)"},{"comment":"The antisymmetry argument after Eq. (21) establishes exact flux conservation for the GSM operator on a fixed mesh, and the text claims 'strict concentration conservation in GSM simulations.' However, every remeshing event (every 100 steps for C–H) inserts/removes nodes with field values assigned by linear interpolation (Eq. 47) and deletes mid-edge nodes on coarsening; neither operation conserves the discrete integral of c. Over a long C–H coarsening run with many remesh cycles, cumulative mass drift is possible and would directly undermine the use of the method for conservation-sensitive coarsening. No mass time series is reported for any adaptive C–H run. Please report the evolution of total mass error for the Fig. 19(b) and Fig. 24 cases, and state whether a conservative remapping (or a correction step) is needed.","section":"§3 (conservation claim), §4.3 Eq. (47), §4.4"},{"comment":"The O(N) complexity claim is measured over a fixed count of 10,000 time steps. With explicit integration, Eq. (23) gives Δt_CH = O(h⁴/κ), so the number of steps to reach a fixed physical time grows as N⁴; time-to-solution therefore scales roughly as O(N⁵) for GSM versus O(N⁶) for FDM, not O(N) versus O(N²). The relative advantage survives, but the abstract's unqualified 'O(N)' overstates it. Relatedly, the MOOSE comparison is performed at the same time-step size and remeshing frequency as GSM, which strips MOOSE of its main practical advantage (implicit stepping with much larger Δt); the conclusion that GSM beats FEM 'under the same conditions' should be explicitly scoped to explicit/matched-step conditions, as §6 partially acknowledges. Please qualify the complexity statements in the abstract and §5.4 accordingly.","section":"Abstract, §5.4 Figs. 21–23 (complexity claim)"},{"comment":"Case V appears internally inconsistent. With N0=21 and n̄_level=10, the finest grid size per the paper's own level-size relation (grandchild = half grid size at +2 levels, §4.2) is (1/20)/2⁵ = 1/640, not the stated h=1/1280. Separately, with κ=1.5625×10⁻⁷, Eq. (9) gives δ≈2.05×10⁻³, i.e., only ~2.6 grid layers across the interface at h=1/1280 — below the 4–5 layers the paper itself requires (§4.5). The pattern of cases I–IV (κ quartered each time, δ/h≈8.3 throughout) suggests κ for case V should be 1.5625×10⁻⁶ and/or n̄_level=12. Since case V is the extreme thin-interface case featured in Fig. 18 and underpins the 'very thin interface' selling point, please verify and correct the table and, if case V was actually run under-resolved, re-examine the associated accuracy/efficiency statements.","section":"Table 1, case V"}],"minor_comments":[{"comment":"The Laplacian is written as ∂²u/∂x² + ∂²u/∂x²; the second term should be ∂²u/∂y².","section":"Eq. (42)"},{"comment":"V_e = (h_c²/6) × 2^{l_e} grows with refinement level, whereas element area should halve per level; presumably this should be 2^{−l_e}. Please check.","section":"Eq. (31)"},{"comment":"Typos: 'Allan–Cahn' and 'Cahn–Hillard' in keywords; 'Intial mesh' (§4.1 heading); 'tra ks' (§4.5); 'where is far from the evolving interface' (missing 'the element', §4.4 after Eq. 48); Fig. 15 caption says 'compared with GDM' — should be FDM. Several lines in §5.1–5.2 contain garbled characters (e.g., ' he performan e'), apparently from a font/encoding problem; please proofread the final typeset file.","section":"Keywords, §4.1, §4.4, §4.5, Fig. 15 caption"},{"comment":"Zhijie (Jay) Xu is listed in the CRediT contributions but does not appear in the author list on the title page; please reconcile.","section":"CRediT statement"},{"comment":"'The results obtained from GSM and FDM agree exactly' overstates it for two distinct discretizations on different meshes; 'agree closely' or a quantitative discrepancy measure (e.g., interface position error vs. time) would be more accurate and more informative.","section":"§5.4, Fig. 19"},{"comment":"No code or data availability statement is given. Given that the vectorized assembly (Eqs. 24–42) and the HAMR algorithms (Figs. 6–7) are described in near-implementable detail, releasing the GSM/HAMR code and benchmark inputs would materially increase the paper's impact; please add an availability statement.","section":"Reproducibility"},{"comment":"The growing remeshing share at large N is acknowledged, but since remeshing is the non-vectorized component, some discussion of its asymptotic cost (and whether it eventually dominates the O(N) scaling) would help readers extrapolate to the 'hundreds of millions of elements' regime mentioned in §6.","section":"§5.4, Fig. 21(b)"}],"recommendation":"major_revision","confidential_remarks":"The method is competently executed and the paper is unusually explicit about implementation, which I view positively. Two things the editor may wish to weigh: (i) the accuracy-restoration premise is currently untested in precisely the regime (bulk-diffusion-driven C–H coarsening) the paper says matters most — my major comment 1 suggests a cheap decisive test, and the outcome genuinely determines whether the headline claim survives; (ii) the MOOSE benchmark is run under conditions (matched explicit-like time steps) that favor the proposed method, so the comparative efficiency language should be toned down regardless of the revision outcome. The GSM citations (refs. 24–26, 42) are the authors' own prior work, which is natural here but means the method's community track record is thin; the numerical evidence in this paper largely stands on its own, however."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a usable engineering methods paper, not a theory result. They marry known GSM with a conforming longest-edge hierarchical triangular AMR that keeps the interface band locally uniform, so the GSM Laplacian stays second-order where it matters and the DOF count tracks interfacial length rather than domain area.\n\nWhat is actually new is the packaging: interface-uniform layered refinement on structured right-isosceles triangles, parent–child–sister rollback coarsening, and a fully vectorized triangle-wise GSM Laplacian with the directional correction that kills checkerboard. Prior art already has GSM, block AMR phase-field FDM/FVM, and adaptive FEM (MOOSE, PRISMS-PF, Alamo). The combination and the design choice to force uniform fine layers across the diffuse interface are the contribution.\n\nThey do the numerics carefully. Uniform-mesh GSM matches second-order FDM on manufactured Laplacians; non-uniform stencils drop to first order as expected; adaptive runs recover ~second-order error vs the tanh equilibrium and track FDM interfaces on A–C shrink and C–H relax; interfacial mobility retraction looks right. Complexity plots vs uniform FDM and a MOOSE adaptive baseline are honest enough: GSM is slower per DOF, wins when the interface is thin or short, and they state the ~8× crossover rule rather than hiding it. Conservation of the C–H flux via antisymmetric edge coefficients is a nice strong-form detail.\n\nSoft spots, in proportion. The load-bearing accuracy story—“non-uniformity lives in nearly flat bulk, so first-order GSM error is harmless”—is supported only on single-interface cases whose bulk really is flat. For multi-particle C–H coarsening, bulk μ carries O(1/R) Gibbs–Thomson gradients that set interface velocity, and those live on the coarse 21×21 mesh with no order-parameter trigger to refine them. That is a genuine gap in the evidence, not a proof the method fails; smooth coarse errors may still be small, but they did not stress-test it. Explicit-only, 2D-only, no code, free knobs (u_thr, n_level, remesh stride), and “better than existing solvers” is broader than the two baselines they actually ran. Remeshing cost share grows with N and is not fully vectorized.\n\nWho it is for: people building large 2D phase-field codes who want a strong-form adaptive option without FEM assembly. Worth a serious referee. I would send it to peer review with a request to narrow the complexity claim, add a multi-droplet C–H bulk-μ check, and release the mesh/operator code.","headline":"Solid adaptive strong-form phase-field methods paper: GSM + layered hierarchical triangles delivers practical O(N) 2D scaling and recovered second-order interface accuracy on the tests they ran; the bulk-μ coarsening caveat is real but not a desk-reject issue.","tokens_in":25136,"tokens_out":679,"would_cite":false,"duration_ms":23722,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M50","65M06","35K55","74N20"],"pacs":[],"model":"grok-4.5","headline":"An adaptive gradient-smoothing solver keeps second-order accuracy on thin interfaces while scaling as O(N) instead of O(N²).","keywords":["phase field method","Allen-Cahn equation","Cahn-Hilliard equation","gradient smoothing method","adaptive mesh","hierarchical AMR","strong-form discretization"],"falsifier":"Measure global convergence rates on a problem whose bulk fields have strong gradients or whose interface topology changes rapidly; if the observed order drops from two to one, or if mass/energy drifts appear after repeated coarsening, the accuracy claim fails.","tokens_in":24785,"feed_emoji":"△","tokens_out":919,"duration_ms":15060,"temperature":0.7,"pith_summary":"Phase-field models track moving interfaces by spreading them into thin diffuse bands, but resolving those bands everywhere on a uniform grid quickly becomes too expensive for large domains or very thin interfaces. This paper couples a strong-form Gradient Smoothing Method with a hierarchical adaptive triangular mesh that keeps fine, uniform resolution only inside the interfacial band and leaves the bulk coarse. A layered refinement rule confines mesh irregularity to a narrow transition zone where the order parameter is nearly constant, so the method still shows overall second-order accuracy. Because the work then grows with interface length rather than domain area, runtime scales near-linearly in system size, beating both uniform finite-difference grids and adaptive weak-form finite-element solvers on the Allen–Cahn and Cahn–Hilliard problems once the grid is large enough or the interface is thin enough.","feed_headline":"Phase-field solver scales as O(N) by refining only the interface","feed_subtitle":"Layered adaptive mesh keeps second-order accuracy while cutting cost for thin diffuse fronts","key_machinery":"Hierarchical Adaptive Mesh Refinement (HAMR) with longest-edge bisection on right-isosceles triangles, driven by an element-average indicator that forces a prescribed number of uniform fine layers across the interface; this mesh supplies the GSM Laplace operator (with directional correction) so that second-order accuracy is retained where gradients matter.","core_discovery":"Coupling the Gradient Smoothing Method to a hierarchical adaptive moving structured triangular mesh that enforces locally uniform fine layers across the diffuse interface yields overall second-order accuracy for Allen–Cahn and Cahn–Hilliard evolution while reducing computational complexity from the O(N²) of uniform grids to O(N), making the solver markedly cheaper for large-scale problems whose interfacial area fraction is small.","pith_inferences":["Once GPU-parallelized, the local node-wise GSM and tree-based remeshing should map cleanly onto thousands of cores, potentially pushing the practical crossover well below the reported 400×400 threshold.","The same framework could be tested on coupled phase-field/elasticity or fluid problems where bulk gradients are no longer negligible, directly probing the load-bearing bulk-flatness assumption.","If the O(N) scaling survives topology changes and three dimensions, adaptive GSM becomes a competitive matrix-free alternative to existing block-structured AMR codes for microstructure evolution."],"forward_implications":["For thin-interface or large-domain phase-field runs the dominant cost becomes proportional to interfacial length rather than domain area.","Memory footprint shrinks with the number of adaptive elements, enabling larger three-dimensional simulations on the same hardware.","The same layered-refinement idea can be reused for any strong-form stencil method that loses order on irregular meshes.","Explicit time-step limits remain the same as finite differences, so the efficiency gain is purely spatial until implicit schemes are added."],"fun_headline_variants":["Adaptive GSM phase-field solver reaches O(N) via interface-only mesh","Layered hierarchical mesh yields O(N) strong-form phase-field evolution","GSM with moving adaptive layers keeps second-order accuracy at O(N) cost","Interface-localized GSM cuts phase-field complexity from O(N²) to O(N)","Hierarchical adaptive mesh enables linear-scaling Allen-Cahn and Cahn-Hilliard"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The first-order truncation error that GSM suffers on non-uniform stencils stays negligible because those stencils sit only in bulk regions where the order parameter is nearly flat.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive GSM phase-field solver reaches O(N) via interface-only mesh","Layered hierarchical mesh yields O(N) strong-form phase-field evolution","GSM with moving adaptive layers keeps second-order accuracy at O(N) cost","Interface-localized GSM cuts phase-field complexity from O(N²) to O(N)","Hierarchical adaptive mesh enables linear-scaling Allen-Cahn and Cahn-Hilliard"]},"model":"grok-4.5","effort":"low","cost_usd":0.004796,"raw_usage":{"total_tokens":1375,"prompt_tokens":809,"num_sources_used":0,"completion_tokens":113,"cost_in_usd_ticks":47964000,"prompt_tokens_details":{"text_tokens":809,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":453,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":809,"tokens_out":113,"duration_ms":8092,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T00:12:13.467718+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Measure global convergence rates on a problem whose bulk fields have strong gradients or whose interface topology changes rapidly; if the observed order drops from two to one, or if mass/energy drifts appear after repeated coarsening, the accuracy claim fails.","supporting_citations":[],"review_version":1}