{"id":"3fdb3379-4aca-42a5-8a06-b6c109a37bd1","arxiv_id":"2411.18770","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A coordinate-space formulation of the conservative diffuse interface method is proposed that removes ambiguity in choosing interface thickness on curvilinear grids and yields grid-adaptive interface thickness with second-order convergence.","lead":"This paper derives a way to write phase-field equations on curved and skewed computational grids by applying the interface regularization in a uniform computational space and mapping it back to physical space. A smart generalist might read it because the method makes diffuse-interface simulations of two-phase flows usable on the non-Cartesian grids common in engineering geometry, with verified accuracy at second order.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proposed computational-space sharpening flux in Eq. 10 is not aligned with the physical interface normal on non-orthogonal grids, so the claimed physical/grid-independent interface adaptation is not established; RT evidence is only visual.","rationale":"The reader identified the same weakest assumption: the computational-space interface normal is not the physical normal. I agree and sharpen the concern by noting that Eq. 10's sharpening flux and the physical normal differ by the metric Jacobian, and they are equal only for conformal mappings, which the test grids are not. This is a direct attack on the central claim of physical grid-independence, not on peripheral details, because the proposed 'transformation' replaces the isotropic Cartesian regularization with a metric-dependent operator. The numerical convergence results are still valuable: they show second-order self-consistency on the chosen smooth grids, but comparing against the initial condition on the same grid cannot certify that the equilibrium interface is the physical one. The Rayleigh-Taylor test is the only two-way coupled evidence, and it is not quantified. The boundedness proof in Appendix B has a separate discretization mismatch, but I consider the normal-direction issue more load-bearing. A focused steady-interface orientation study would settle the concern, so the appropriate verdict remains CONDITIONAL: the paper should be accepted only after the physical isotropy of the proposed regularization is demonstrated or the claim is narrowed to computational-space regularization.","tokens_in":17246,"tokens_out":16264,"duration_ms":154863,"concrete_test":"On a non-orthogonal grid such as Grid 3 (S = 1/10), initialize a planar interface with zero velocity whose physical normal is oriented at angle θ = 0, 30, 60, and 90 degrees relative to ξ1, using the same local grid spacing. Run Eq. 10 to steady state. For each θ, measure the steady φ = 0.5 contour's physical normal and the physical interface width, e.g. 1/|∇_x φ| at the inflection point. If the normal rotates, the width varies with θ by more than the known local grid anisotropy, or the profile is not a tanh in the physical normal coordinate, then the computational-space regularization is not an isotropic physical interface model.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the proposed transformation accurately adapts interface thickness to the local grid in physical space and preserves the physics of the phase-field regularization. The load-bearing step is Eq. 10, where the nonlinear sharpening term uses the computational-space gradient direction, ∂ϕ/∂ξ_i / |∇_ξ ϕ|. In physical space this vector is proportional to M^{-T} n_ξ, with M_{ji} = ∂ξ_i/∂x_j, whereas the physical interface normal is ∇_x ϕ / |∇_x ϕ|, proportional to M^T n_ξ / |M^T n_ξ|. These two directions coincide only when the mapping is locally conformal, i.e. M^T M is a scalar multiple of the identity. On the sinusoidal grids 2-4 in Table 1, the Jacobian is not conformal, so the sharpening flux in physical space is not normal to the interface. Consequently, the equilibrium profile is a constant-thickness tanh in computational coordinates; after projection it is not the usual isotropic tanh interface in physical space, and its physical width and orientation can depend on the local metric. The convergence tests against the initial condition on the same grid (Figs. 7 and 9) do not test this isotropy, and the Rayleigh-Taylor grid-independence claim rests on visual contour overlays (Fig. 12) without a quantitative error measure. Thus the paper has not demonstrated that the proposed model is the physically correct diffuse-interface regularization on general curvilinear grids.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a curvilinear-coordinate formulation of conservative diffuse-interface (CDI/ACDI) phase-field equations. Instead of directly transforming the Cartesian regularization operator, the authors define the regularization in computational space with the standard thickness parameter ε_ξ = 0.75 Δξ and project the resulting flux back to physical space, yielding Eq. (10). They verify the formulation with three test problems — drop advection, drop in shear flow, and a two-way coupled compressible Rayleigh-Taylor instability — on four grids with varying skewness. The reported results show second-order L1 convergence for the advection and shear tests and visual grid independence for the Rayleigh-Taylor case.","tokens_in":17478,"tokens_out":8863,"duration_ms":81355,"significance":"If the claims hold, the paper provides a parameter-unambiguous extension of CDI/ACDI to structured curvilinear grids, with no fitted constants and a simple computational-space guideline for the interface thickness. The verification suite is appropriate in scope: three canonical cases, four grids, quantitative L1 convergence for two of them, and a genuinely coupled compressible problem. The paper also makes a credible effort to extend the boundedness argument from prior Cartesian work. The main weaknesses are that the physical-space meaning of the computational-space interface normal is not established, the Appendix B boundedness proof discretizes a different operator than Eq. (10) and a different advection scheme than the KEEP scheme used in Section 3, and the Rayleigh-Taylor grid-independence claim rests on visual contour overlays rather than a quantitative metric. These issues are local and fixable, but they are load-bearing for the central claims of accurate physical-space interface adaptation, boundedness without oscillations, and grid-independent convergence.","major_comments":[{"comment":"The sharpening direction in physical space is not the physical interface normal on non-orthogonal grids. In Eq. (10) the nonlinear flux is proportional to ∂φ/∂ξ_i / |∇_ξ φ|; when interpreted as a physical vector this is (∂x_j/∂ξ_i)∂φ/∂ξ_i / |∇_ξ φ|, whereas the physical normal to the level sets is (∂ξ_i/∂x_j)∂φ/∂ξ_i / |∇_x φ|. These coincide only when the mapping is locally conformal, which the grids in Table 1 are not. Consequently the equilibrium profile is a constant-thickness tanh in computational coordinates, and in physical space its width and orientation depend on the local metric; it is not the usual isotropic diffuse interface. This is not necessarily a flaw if the method is intended as a computational-space regularization, but the paper should state that explicitly and justify it, because the abstract and Section 2.2 claim the transformation accurately adapts the interface thickness in physical space. The verification tests do not resolve the ambiguity: Eq. (14) initializes a circle in computational coordinates, so the advection and shear tests measure self-consistency on the same grid, and the Rayleigh-Taylor comparison in Fig. 12 is visual only. I recommend adding a quantitative test on a non-orthogonal grid with a flat interface whose physical normal is not aligned with a coordinate line, measuring the physical-space profile and normal, and reporting an interface-width error or a grid-to-grid difference metric for the RT case.","section":"2.2, Eq. (10)"},{"comment":"The boundedness proof does not apply to the discretization actually used in the code. The diffusion term in Eq. (B.3) is written as Γ_ξ ε_ξ (φ_{i-1}/J_{i-1} - 2φ_i/J_i + φ_{i+1}/J_{i+1})/Δξ^2, which is a discretization of ∂^2(φ/J)/∂ξ^2, whereas Eq. (10) contains ∂/∂ξ_i [Γ_ξ(...)/J]; a consistent finite-volume discretization would evaluate 1/J at faces. The advection term is likewise a second-order central difference, not the sixth-order KEEP scheme described in Section 3. The proof therefore does not establish boundedness for the implemented scheme, and the statements in Section 2.2 and the Conclusion that boundedness has been extended to the curvilinear domain are stronger than what is shown. The authors already note that no formal proof exists for higher-order schemes, but the mismatch even at the second-order level should be corrected, or the boundedness claim should be downgraded to an empirical observation supported only by the test cases.","section":"Appendix B, Eqs. (B.3)-(B.8)"},{"comment":"The claim of grid-independent convergence for the Rayleigh-Taylor case is supported only by visual contour overlays. The text attributes the Grid 4 discrepancy to earlier wall interaction without a quantitative measure, so it is not possible to assess whether the four solutions converge to the same physical solution. Please provide a quantitative metric, for example the L1 or L2 difference of φ or of the interface position between grids at the common highest resolution, or a Richardson-type comparison of a scalar quantity such as the bubble tip position.","section":"4.3, Fig. 12"}],"minor_comments":[{"comment":"The y-mapping for Grid 3 contains 'sin(λπξ1) sin(λπξ1)'; the second factor should probably be sin(λπξ2).","section":"Table 1, Grid 3"},{"comment":"The text says the radius is '0.25% of the computational domain width,' while Eq. (14) appears to use R = 0.25; please clarify which value is intended.","section":"Section 4.1"},{"comment":"The definition of A_l uses the free index i without specifying the coordinate direction; since A_l is used as a flux component in Eq. (12), clarify whether A_l denotes a vector of components or the ξ1-component.","section":"Eq. (13)"},{"comment":"The HTR solver and Legion framework are mentioned but no code or data availability statement is provided; adding one would strengthen the reproducibility of the verification results.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable after revision if the authors clarify whether the proposed regularization is intended to be isotropic in computational space or in physical space, and if the boundedness proof is aligned with the actual discretization. The physical-normal issue is the deepest conceptual point; if the method is a computational-space regularization, the claims in the abstract and Section 2.2 should be reframed accordingly. The boundedness proof mismatch is a rigor/documentation issue that should be fixed in the appendix. The Rayleigh-Taylor comparison would also benefit from a quantitative grid-convergence measure rather than visual overlays only."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: the central idea is genuinely new and the verification is appropriate. The authors define the phase-field regularization in computational space and project the flux back to physical space, avoiding the parameter ambiguity of a direct transformation. The derivation is compact—the metric identity collapses cleanly—and the same recipe is given for CDI, ACDI, Allen-Cahn, and Cahn-Hilliard. On the numerical side, the advection and shear tests on four grids show clean second-order convergence, and the Rayleigh-Taylor simulation converges visually to the same physical solution across grids. There are no fitted constants; epsilon_xi = 0.75 Delta_xi is the standard criterion. That is a solid package for a method paper.\n\nSoft spots, in order of importance. First, the boundedness proof in Appendix B discretizes a diffusion term that is not the one in Eq. 10. The proof works with a central difference of phi_i/J_i, while Eq. 10 is a divergence of (1/J) dphi/dxi. Those differ when J varies, so the proof as written does not establish boundedness for the actual discrete operator. The numerical results are not in question, but the proof needs to be redone or corrected. Second, the Rayleigh-Taylor 'grid independence' claim rests on visual contour overlays; a quantitative error norm on the interface position would make it convincing. That is a minor ask. Third, the stress-test point that the sharpening flux in physical space is not aligned with the physical interface normal on non-orthogonal grids is correct as a statement about the construction. The authors should say plainly that the model produces a tanh profile in computational coordinates, not an isotropic physical tanh, and explain why that is the intended definition of grid-adapting thickness. They may well have a good answer, but the paper does not currently address it. Finally, no code or data are shipped, so exact reproduction requires re-implementation.\n\nNone of this sinks the central claim. The method is well-motivated, the tests verify what they claim, and the paper is written clearly. This is a paper for CFD/phase-field practitioners working on curvilinear and body-fitted grids. It deserves a serious referee. I would send it to peer review and ask the authors to fix the boundedness proof, add a quantitative RT error measure, and address the interface-normal question head-on.","headline":"A genuinely new computational-space formulation with solid verification; fix the boundedness proof and address the interface-normal question before publication.","tokens_in":18069,"tokens_out":3715,"would_cite":true,"duration_ms":34270,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a parameter-free transformation of phase-field regularization into generalized curvilinear coordinates, making the diffuse interface thickness adapt to the local grid size while preserving Cartesian-order convergence.","keywords":["phase field","diffuse interface","curvilinear coordinates","conservative diffuse interface","ACDI","grid-adapting interface thickness","boundedness","two-phase flow"],"falsifier":"On a non-uniform, non-orthogonal grid with sharply varying Jacobian, initialize a stationary planar interface and measure the equilibrium profile thickness in physical space as a function of position; if the number of cells across the interface varies more than the local grid-size ratio, or if oscillations or values of $\\phi$ outside $[0,1]$ appear for parameters satisfying the stated boundedness constraints, the central claim fails. Alternatively, run the drop-advection test on a grid whose mapping has discontinuous second derivatives and check whether the near-second-order convergence persists.","tokens_in":1742,"feed_emoji":"🌊","tokens_out":3255,"duration_ms":161775,"temperature":0.7,"pith_summary":"The paper tries to establish that phase-field two-phase methods, specifically the conservative diffuse interface (CDI) family and its ACDI variant, can be extended to generalized curvilinear grids without introducing a free parameter for interface thickness. The proposed move is to define the regularization flux in computational space, where the grid is uniform, and then project that flux back to physical space, so the interface automatically keeps the same number of cells across it everywhere. If true, this removes a known obstacle: direct transformations either make the interface too thick in refined regions or distort the interface through direction-dependent thickness parameters. Verification on drop advection, drop-in-shear, and a two-way coupled compressible Rayleigh-Taylor instability shows the same near-second-order convergence as on Cartesian grids, including on highly skewed grids. The practical payoff is that curvilinear and multi-block structured grids keep their efficiency advantages while phase-field interfaces remain stable, bounded, and grid-adaptive.","feed_headline":"Phase-field interfaces now self-adapt on skewed grids","feed_subtitle":"A computational-space flux projection makes interface thickness track local resolution and preserves second-order convergence.","key_machinery":"The central object is the computational-space regularization-flux projection. Starting from the CDI flux written in computational coordinates, the physical flux is obtained by contraction with $\\partial x_j/\\partial \\xi_i$, and the divergence is re-expressed in computational space via the metric Jacobian. The identity $(\\partial x_j/\\partial \\xi_k)(\\partial \\xi_i/\\partial x_j) = I$ collapses the projected equation to Eq. (10), where the sharpening normal uses computational-space gradients normalized by $|\\nabla_\\xi \\phi|$. This is what makes the interface thickness constant in computational coordinates and therefore grid-adapting in physical space, and it also supplies the boundedness constraints $\\epsilon_\\xi/\\Delta\\xi \\ge |U|_{\\max}/\\Gamma_\\xi + 1/2$ and $\\Delta t \\le \\Delta\\xi^2/(2\\Gamma_\\xi \\epsilon_\\xi)$.","core_discovery":"The paper claims that the ambiguity in curvilinear phase-field regularization—how to choose the interface-thickness parameter $\\epsilon$ when the grid size varies—can be removed by formulating the regularization flux entirely in computational space and projecting it back to physical space. In the proposed form, $\\epsilon_\\xi$ is simply proportional to the computational grid spacing $\\Delta\\xi$ and $\\Gamma_\\xi$ to the maximum contravariant velocity, and the chain-rule projection reduces the final equation to the same shape as the Cartesian CDI operator with an additional $1/J$ factor. The resulting equilibrium interface has constant thickness in computational space, which projects to an interface that adapts to the local physical grid size. On the advection, shear-flow, and Rayleigh-Taylor tests, the curvilinear formulation converges at the same near-second-order rate as the Cartesian one on all grids tested, and the volume fraction remains bounded under the stated parameter and time-step constraints.","pith_inferences":["Because the transformation is derived purely from chain-rule projection, the same computational-space prescription could apply to other divergence-form regularization terms, including sharpening terms in level-set or volume-of-fluid contexts, though the paper does not test that.","The formulation naturally pairs with multi-block or locally refined structured grids, where each block carries its own $\\Delta\\xi$ and interface thickness would stay grid-adaptive across block boundaries; this is not demonstrated in the paper.","The constant-thickness-in-computational-space design implies that physical-space interface normals and curvatures derived directly from $\\phi$ may reflect the computational-space profile rather than the physical one, so geometric postprocessing may need metric-corrected normals, which the paper leaves implicit.","A testable extension is to use the boundedness constraints to choose time steps adaptively in regions of strong grid stretching, since the relevant CFL condition depends on $\\Gamma_\\xi$ and $\\Delta\\xi$ rather than the physical velocity and grid size alone."],"forward_implications":["Phase-field parameters have a unique prescription on any curvilinear grid: $\\epsilon_\\xi \\propto \\Delta\\xi$ and $\\Gamma_\\xi = \\max_i |U_i|$, removing the option-1/option-2 ambiguity.","Interface thickness in physical space follows the local grid resolution, so refined regions get sharper interfaces without user intervention.","Boundedness of the volume fraction ($0 \\le \\phi \\le 1$) carries over from Cartesian grids under the stated time-step and $\\epsilon_\\xi$ constraints for second-order centered discretizations.","The same derivation applies to CDI, ACDI, Allen-Cahn, and Cahn-Hilliard forms, as summarized in the paper's appendix.","On the tested flows, convergence is near second order and identical between Cartesian and curvilinear grids, with error magnitude growing with skewness but the convergence rate unchanged."],"supporting_citations":[{"why":"Introduces the conservative phase-field (CDI) model that the curvilinear transformation generalizes.","marker":"[5]"},{"why":"Supplies the CDI boundedness analysis and parameter criteria that the curvilinear version extends.","marker":"[9]"},{"why":"Establishes the conservative diffuse-interface formulation for compressible two-phase flows used in the coupled Rayleigh-Taylor test.","marker":"[10]"},{"why":"Demonstrates the direct alternative choice $\\epsilon_i = \\Delta_i$ on unstructured grids, whose distortion motivates the proposed projection.","marker":"[17]"},{"why":"Introduces the ACDI signed-distance reformulation adopted for the numerical tests.","marker":"[21]"},{"why":"Provides the signed-distance/logarithmic transformation used to define $\\psi$ and improve interface interpolation.","marker":"[20]"},{"why":"Supplies the kinetic-energy and entropy-preserving curvilinear-grid discretization used for convection.","marker":"[25]"},{"why":"Provides the consistent two-phase discretization that lets the phase-field fluxes retain the KEEP properties.","marker":"[26]"}],"fun_headline_variants":["Phase-field thickness now follows grid skewness","Curvilinear phase-field automates interface thickness","No epsilon tuning on skewed meshes","Grid-adaptive phase-field for curved coordinates"],"cache_read_input_tokens":20096,"weakest_assumption_plain":"The method assumes that the physically correct diffuse interface on a general grid is the one that has constant thickness in computational coordinates after the flux is projected back, an assumption verified here only on smooth, analytically defined grids.","fun_headline_variants_meta":{"raw":{"variants":["Phase-field thickness now follows grid skewness","Curvilinear phase-field automates interface thickness","No epsilon tuning on skewed meshes","Grid-adaptive phase-field for curved coordinates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001126,"raw_usage":{"total_tokens":4642,"prompt_tokens":868,"completion_tokens":3774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":3719}},"tokens_in":484,"tokens_out":3774,"duration_ms":30053,"temperature":1.0,"reasoning_tokens":3719,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:54:28.891392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a non-uniform, non-orthogonal grid with sharply varying Jacobian, initialize a stationary planar interface and measure the equilibrium profile thickness in physical space as a function of position; if the number of cells across the interface varies more than the local grid-size ratio, or if oscillations or values of $\\phi$ outside $[0,1]$ appear for parameters satisfying the stated boundedness constraints, the central claim fails. Alternatively, run the drop-advection test on a grid whose mapping has discontinuous second derivatives and check whether the near-second-order convergence persists.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the conservative diffuse-interface formulation for compressible two-phase flows used in the coupled Rayleigh-Taylor test."},{"cited_title":"Hwang, S","cited_arxiv_id":null,"evidence_quote":"Demonstrates the direct alternative choice $\\epsilon_i = \\Delta_i$ on unstructured grids, whose distortion motivates the proposed projection."},{"cited_title":"Chiodi, O","cited_arxiv_id":null,"evidence_quote":"Provides the signed-distance/logarithmic transformation used to define $\\psi$ and improve interface interpolation."}],"review_version":1}