{"id":"15c82dc0-66b1-4435-8bd0-a49a37f7f425","arxiv_id":"2505.24504","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Jacobian-free multigrid preconditioner built on a low-order finite volume subgrid makes implicit DG time integration competitive for 2D atmospheric flows.","lead":"This paper introduces a preconditioner that lets implicit time stepping work efficiently for high-order discontinuous Galerkin weather models, by solving a cheap low-order finite volume version of the same problem inside a multigrid cycle. The approach cuts Krylov iterations dramatically in 2D atmospheric test cases, and in some high-resolution runs the implicit solver is about twice as fast as an explicit one.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The method's effectiveness hinges on an unexamined spectral-equivalence assumption between the DG Jacobian and the low-order FV subgrid; the two-resolution iteration counts do not establish h-robustness.","rationale":"I read the paper as a numerical-methods contribution whose central claim is that the proposed Jacobian-free multigrid preconditioner, built on a low-order FV subgrid, substantially reduces GMRES iterations and makes implicit DG time integration competitive with explicit time stepping. The three 2D atmospheric test cases provide genuine empirical support: large explicit/implicit time-step ratios, substantial iteration reductions relative to the unpreconditioned solver where it could be run, and wall-clock comparisons on two of the test cases. This is real evidence and I would not dismiss it. The most load-bearing weakness is not the absence of a proof of mesh independence per se, but the unexamined spectral link between the high-order DG operator and the low-order FV operator via the transfer operator T. The paper explicitly leaves the design criterion for T at the level of 'functions are close in some sense' (Section 2.4.1), and all justification is delegated to prior work and to the numerical results themselves. The observed growth of iteration counts with refinement, while modest, is exactly what one would expect if the low-order subgrid captures the dominant error modes only approximately; the experiments stop at two resolutions per test case, so the trend cannot be extrapolated. The reader's verdict of CONDITIONAL is appropriate: the method is credible and worth replicating, but the missing spectral/robustness evidence and the lack of released code mean the strongest form of the claim, that the preconditioner makes the implicit method competitive at high resolution in NWP, is not yet settled. My proposed test is targeted: it directly measures whether the transfer and multigrid combination preserves the relevant error components at the finest resolutions and whether the preconditioner's quality degrades with mesh refinement.","tokens_in":16161,"tokens_out":11078,"duration_ms":154300,"concrete_test":"At a representative SDIRK stage in the rising-bubble test at Delta x = 25 m and Delta x = 12.5 m, assemble (matrix-free) the operators G'_h and g'_h together with the transfer operator T_h, and estimate the spectrum of the preconditioned operator P_h^{-1} G'_h with P_h^{-1} = T_h^{-1} (V-cycle)_h T_h, e.g. by a few hundred Arnoldi iterations on random initial vectors. If the eigenvalues are not clustered away from zero, or if the effective condition estimate grows by more than roughly a factor of two per mesh refinement, then the transfer/multigrid pair is not preserving the dominant DG error modes and the claimed competitiveness at higher resolution is not supported. If the spectrum is tightly clustered at both resolutions, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the FV-multigrid preconditioner makes implicit DG time stepping competitive at high resolution depends on the unquantified assumption that the transfer pair (T, T^{-1}) and the low-order FV Jacobian g' faithfully represent the dominant error dynamics of the DG Jacobian G'. Section 2.4.1 supports this only by the statement that the transfer functions are 'defined such that the functions are close in some sense' and by citations to prior numerical comparisons [16,30]; no estimate of the defect D = G' - T^{-1} g' T is given, and no smoothing analysis is provided for the explicit-Euler pseudo-time iteration on the viscous, gravity-driven systems considered here. The reported iteration counts are consistent with a preconditioner whose quality degrades under refinement: in the rising-bubble case (Section 3.2) the cumulative linear-iteration count grows by a factor below two when the grid width is halved from 25 m to 12.5 m, and the density-current case (Section 3.3) shows a similar trend. Because the conclusion is explicitly about high-resolution NWP, the missing h-uniformity evidence is load-bearing: if the defect D is not small on the coarse error components at finer meshes, the observed iteration reductions will not persist and the wall-clock advantage over explicit stepping would disappear.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Jacobian-free multigrid preconditioner for implicit time integration of high-order discontinuous Galerkin (DG) discretizations of compressible atmospheric flow equations. The preconditioner transfers the DG solution to a low-order finite volume subgrid with the same number of degrees of freedom, applies a geometric multigrid V- or W-cycle with explicit Runge-Kutta pseudo-time smoothers, and transfers back with a mass-conservative mapping. The method is embedded in a Jacobian-free Newton-GMRES framework with SDIRK2 time stepping and is tested on three two-dimensional benchmark problems: non-hydrostatic inertia gravity, rising warm air bubble, and density current. The paper reports GMRES iteration counts and wall-clock times, showing significant reductions relative to unpreconditioned GMRES and, in some high-resolution runs, CPU times competitive with an explicit strong-stability-preserving Runge-Kutta method.","tokens_in":16474,"tokens_out":6879,"duration_ms":77719,"significance":"If the results hold, the paper offers a practical path toward making implicit high-order DG methods affordable for numerical weather prediction, a setting where stiffness from gravity and low Mach numbers makes explicit time stepping unattractive. The strengths are the use of standard benchmark test cases, direct comparison of iteration counts and wall-clock times, the mass-conservative transfer construction, and the availability of the implementation within the open DUNE-FEM framework. The main weaknesses are that several strong claims—parallel scalability, low memory footprint, and near-optimal arithmetic intensity—are not supported by measurements, the h-dependence of iteration counts is only demonstrated over one refinement step per test case, and solution accuracy is only assessed visually rather than through quantitative error norms.","major_comments":[{"comment":"The pseudo-CFL number used for the explicit Euler smoother is never reported. The text states that the pseudo time step is 'computed based on a pseudo CFL number' and that the parameters are 'chosen based on [18]', but no value or tuning procedure is given. Since the iteration counts in Figures 5, 7, and 10 depend directly on this smoother parameter, the experiments cannot be reproduced without it. Please provide the pseudo-CFL value(s) used, and clarify how the choice of one-stage explicit Euler relates to the optimized multi-stage Runge-Kutta smoothers in [18].","section":"Section 2.4.4 and Section 3"},{"comment":"The central claim that the preconditioned implicit solver becomes competitive at high resolution is supported by only two resolutions in each test case, and the iteration counts grow under refinement. The paper explicitly acknowledges that the rising-bubble iteration count grows by less than a factor of two when the grid width is halved from 25 m to 12.5 m and that the density-current count grows by less than 50% from 320 m to 160 m. This is insufficient to establish that the preconditioner will remain effective at the much finer resolutions typical of NWP. Please either add experiments at one or two additional refinement levels, or revise the conclusions to state the demonstrated resolution regime.","section":"Section 3.2, Fig. 7(c); Section 3.3, Fig. 10(b); Section 4"},{"comment":"The statements that the preconditioner 'achieves close to optimal arithmetic intensity for large problems on parallel machines' and 'uses little extra memory' are not supported by any measurement in the manuscript. There are no FLOP-to-byte ratios, roofline data, memory footprint measurements, or parallel scaling results. The abstract also promises a 'fast, highly parallel, low-memory iterative solver.' These claims should either be demonstrated with appropriate experiments or removed/tempered, as they currently go beyond what the paper shows.","section":"Section 4 (Summary and Conclusions)"},{"comment":"Solution accuracy is only assessed visually. Figures 3, 4, and 9 show line plots, and the text uses qualitative language such as 'broadly follow' and 'some deviation.' No L1, L2, or L-infinity error norms relative to the explicit reference solution are reported. Since the conclusion states that 'the implicit high order solver produces accurate results even for large CFL numbers,' quantitative accuracy metrics are needed to support this claim.","section":"Section 3 and Section 4"}],"minor_comments":[{"comment":"There are typos: 'precondtioner' in the abstract and 'straighforward' in Section 2.4.2. Please correct them.","section":"Abstract and Section 2.4.2"},{"comment":"In the sentence 'at ∆x=25m we the DG solution is computed' the verb is missing, and later the configuration 'mg0011111V' appears with an extra '1'; it should be 'mg001111V'.","section":"Section 3.2"},{"comment":"The domain notation '[0, 300 000]×[0, 10 000]km^2' is dimensionally inconsistent; presumably the lengths are in meters.","section":"Section 3.1"},{"comment":"The value 'cp = 1, 004J/(kg K)' uses a comma as a decimal separator inconsistently with the rest of the text; consider writing 1004 J/(kg K).","section":"Section 3.3"},{"comment":"The phrase 'defined such that the functions are close in some sense' is vague. Please specify the relevant norm or cite a precise approximation result from [16, 30] for the interpolation-based transfer operator.","section":"Section 2.4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid engineering-oriented numerical study, but the gap between the abstract's promises (parallel scalability, low memory, near-optimal arithmetic intensity) and the actual evidence (serial 2D runs, no memory measurements) is significant. The h-robustness and reproducibility concerns are fixable within the scope of a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a solid engineering extension of a known subcell-FV multigrid preconditioner to 2D viscous compressible flow with gravity, with honest benchmark data on three standard NWP test cases. The central claim—that the preconditioner makes implicit DG competitive with explicit SSP-RK at high resolution—is supported for the cases shown, though the concluding generalizations about parallel efficiency outrun the evidence.\n\nWhat's new and good: the paper takes the preconditioner from [15–17] and adds the viscous terms, gravity source, well-balanced formulation, and a mass-fix variant of the transfer operator. The three test cases (inertia-gravity, rising bubble, density current) are standard, and the comparison against explicit SSP-RK with wall-clock times is useful. The iteration counts in Figures 5, 7, and 10 are the real evidence, and they look credible: the preconditioner cuts GMRES iterations dramatically versus unpreconditioned, and the growth under grid refinement is reported honestly—less than a factor of two when the mesh is halved—rather than hidden.\n\nSoft spots, in proportion. First, the paper releases no code or data; for a method whose entire claim is performance, this is a real reproducibility gap. DUNE-FEM is open source, so an artifact should be feasible. Second, the accuracy assessment is visual only; there are no error norms or convergence tables, so the reader cannot quantify what is lost at the largest time steps. Third, the last section claims the method 'achieves close to optimal arithmetic intensity for large problems on parallel machines,' but nothing in the paper measures arithmetic intensity or parallel scaling; that sentence should go or be backed by data. Fourth, the transfer operator's spectral equivalence is asserted by citation and 'close in some sense' (Section 2.4.1); no analysis of the defect G' - T^{-1}g'T is given. The stress-test worry about h-robustness does not land as fatal—the two-resolution data actually show modest iteration growth—but it does mean the conclusion is only established for the tested resolutions, and the authors should say so plainly. A serious referee would want either one more refinement level, a Fourier/smoothing analysis, or a clear statement that h-independence is not claimed.\n\nWho it is for: numerical analysts and NWP solver developers working on matrix-free implicit DG. It deserves a serious referee; with an artifact and trimmed conclusions, it would be a useful contribution. I would not desk-reject it.","headline":"Solid empirical extension of a known FV-multigrid preconditioner to 2D viscous atmospheric flows; the benchmark data support the main claim, but the conclusions overreach and the missing artifacts weaken reproducibility.","tokens_in":16963,"tokens_out":2430,"would_cite":true,"duration_ms":29960,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65M55","65F08","76M10","86A10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that a Jacobian-free multigrid preconditioner built on a low-order finite-volume subgrid makes implicit, high-order discontinuous Galerkin time stepping competitive for two-dimensional atmospheric flows.","keywords":["discontinuous Galerkin","finite volume subgrid","multigrid preconditioner","Jacobian-free Newton-Krylov","implicit time integration","numerical weather prediction","well-balanced scheme","pseudo-time smoother"],"falsifier":"Run the same preconditioner on a problem whose solution error is dominated by components that oscillate inside each DG cell, or refine the grid while holding the time step fixed and record the multigrid convergence factor: if the preconditioned GMRES iteration count grows proportionally to the number of levels, or if the V-cycle convergence factor approaches one, the finite-volume subgrid is failing to represent the error modes that the solver must remove. A more direct calculation is to compare the energy of representative high-order error modes before and after transfer through $T^{-1}T$; modes whose energy is significantly reduced by the transfer are not being smoothed, and their presence in the residual would falsify the assumption.","tokens_in":15979,"feed_emoji":"🌤️","tokens_out":11089,"duration_ms":125993,"temperature":0.7,"pith_summary":"This paper claims that implicit time stepping for high-order discontinuous Galerkin (DG) discretizations of atmospheric flow can be made competitive with explicit time stepping by preconditioning the Newton--Krylov solver with a multigrid method defined on a low-order finite-volume subgrid. In three 2D test cases, the preconditioner sharply reduces the number of GMRES iterations and lets the implicit method take time steps hundreds to thousands of times larger than the explicit stability limit. The practical stake is that DG methods are attractive for numerical weather prediction but have lacked efficient, matrix-free, parallel solvers for the large algebraic systems produced by implicit time integration. The paper's solver is Jacobian-free and uses little extra memory, and on the finest grids tested its wall-clock time is roughly half that of the explicit reference scheme.","feed_headline":"Jacobian-free multigrid makes implicit DG weather solvers competitive","feed_subtitle":"On fine grids it runs about twice as fast as explicit time stepping while taking steps roughly 500 times larger.","key_machinery":"The load-bearing object is the composite preconditioner $Q^{-1}=T^{-1}q^{-1}T$, where $T$ interpolates the high-order DG solution at the cell centers of a finite-volume subgrid and $q^{-1}$ is a geometric multigrid cycle for the first-order finite-volume discretization built on that subgrid. The multigrid restriction is volume averaging of child-quad averages, prolongation is injection of the agglomerate average, and the smoother is a one-stage explicit Euler pseudo-time iteration on the finite-volume levels, with optional smoothing on the DG data before transfer. This machinery matters because it approximates the inverse of the high-order DG Jacobian using only operators that are local, parallel, and matrix-free; the outer GMRES iteration then sees a better-conditioned system without the memory cost of forming the full Jacobian.","core_discovery":"The central claim is that the linear systems from an SDIRK2 time discretization of a DG space discretization of the 2D Euler or viscous flow equations with gravity can be solved efficiently by preconditioned Jacobian-free Newton-GMRES, provided the preconditioner is an agglomeration multigrid V-cycle applied to a first-order finite-volume discretization on a subgrid with the same number of degrees of freedom. Transfer between the DG polynomial space and the finite-volume cell averages is done by interpolation, with an optional mass-conservative correction, and the multigrid smoother is an explicit Runge-Kutta pseudo-time iteration, so no Jacobian matrix is ever formed. The stated result is that the preconditioner increases the stability of the solver while substantially reducing the number of iterations, making the implicit method competitive; on the fine-grid runs, the implicit method finished in about half the wall-clock time of the explicit SSP time stepping while using time steps around 500 times larger.","pith_inferences":["Because the preconditioner acts on the residual operator rather than on any specific flux or geometry, the same transfer-plus-subgrid-multigrid construction should carry over to three-dimensional DG-SEM and simplicial DG discretizations, with the node layout and subgrid subdivision adjusted accordingly.","A sharper test of the transfer assumption would be to measure, for a fixed error mode, the ratio of its energy after $T^{-1}T$ to its original energy; modes that the finite-volume subgrid annihilates are exactly the ones the smoother must handle, and their loss would predict when iteration reductions fail to generalize.","In flows where global mass conservation is a hard constraint, such as moist or tracer transport, the optional mass-conservative correction in the transfer may become essential even though it made no difference in the dry test cases reported here.","The observed growth of iterations under mesh refinement indicates the scheme is not yet an optimal-order solver; a natural next step would be to identify whether the limiting factor is the pseudo-time smoother or the coarse-grid correction of the finite-volume subgrid."],"forward_implications":["Unpreconditioned Jacobian-free GMRES is prohibitively slow or fails to converge within the allocated time on the inertia-gravity and rising-bubble tests; the multigrid preconditioner is what makes those runs finish.","Implicit time steps of roughly 500 to 2000 times the explicit stability limit reproduce the reference solution closely in the rising-bubble and density-current cases, with larger steps degrading accuracy.","On the finer grids tested, the preconditioned implicit method used roughly half the wall-clock time of the explicit SSP reference, for example 7915 seconds versus 14000 seconds for the rising bubble at 12.5 meter spacing.","GMRES iteration counts grow sublinearly with increasing time step and by less than a factor of two when the grid is refined, although they are not fully grid independent.","The most efficient configuration was a V-cycle with one pre- and post-smoothing step on the DG data and all finite-volume levels; W-cycles improve iteration counts but roughly double the run time."],"supporting_citations":[{"why":"introduces the subcell low-order finite-volume multigrid preconditioner for high-order DG that this paper extends.","marker":"[15]"},{"why":"extends the approach to convection-diffusion and compares transfer operators, supporting the interpolation-based choice.","marker":"[16]"},{"why":"supplies the optimized coefficients for the explicit pseudo-time smoother used inside the multigrid cycle.","marker":"[18]"},{"why":"provides the high-order implicit time-integration and inexact Newton framework with adaptive tolerances that the solver follows.","marker":"[19]"},{"why":"supplies the forcing-term criterion used to select the GMRES tolerance adaptively.","marker":"[29]"},{"why":"introduces the well-balanced perturbation formulation used to handle the gravity source term in the atmospheric test cases.","marker":"[34]"},{"why":"defines the non-hydrostatic inertia-gravity test case used for the first validation.","marker":"[36]"},{"why":"defines the rising warm bubble test case used for the second validation.","marker":"[37]"},{"why":"defines the density current benchmark used for the third validation.","marker":"[39]"}],"fun_headline_variants":["Jacobian-free multigrid speeds implicit DG weather solvers","Multigrid preconditioner halves DG weather solver runtime","Implicit DG weather model gains speed from subgrid multigrid","DG weather implicit steps get 500x larger with new preconditioner","Subgrid multigrid makes implicit DG weather runs twice as fast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a coarse cell-averaged version of the same equations, connected to the high-order DG discretization by simple interpolation, retains the error modes that dominate the high-order Jacobian, a property the paper relies on from earlier empirical comparisons rather than proving here.","fun_headline_variants_meta":{"raw":{"variants":["Jacobian-free multigrid speeds implicit DG weather solvers","Multigrid preconditioner halves DG weather solver runtime","Implicit DG weather model gains speed from subgrid multigrid","DG weather implicit steps get 500x larger with new preconditioner","Subgrid multigrid makes implicit DG weather runs twice as fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000411,"raw_usage":{"total_tokens":2170,"prompt_tokens":1028,"completion_tokens":1142,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1055}},"tokens_in":644,"tokens_out":1142,"duration_ms":12230,"temperature":1.0,"reasoning_tokens":1055,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:19:59.143627+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same preconditioner on a problem whose solution error is dominated by components that oscillate inside each DG cell, or refine the grid while holding the time step fixed and record the multigrid convergence factor: if the preconditioned GMRES iteration count grows proportionally to the number of levels, or if the V-cycle convergence factor approaches one, the finite-volume subgrid is failing to represent the error modes that the solver must remove. A more direct calculation is to compare the energy of representative high-order error modes before and after transfer through $T^{-1}T$; modes whose energy is significantly reduced by the transfer are not being smoothed, and their presence in the residual would falsify the assumption.","supporting_citations":[{"cited_title":"and Versbach, L","cited_arxiv_id":null,"evidence_quote":"extends the approach to convection-diffusion and compares transfer operators, supporting the interpolation-based choice."},{"cited_title":"ETNA 39, 298–312 (2012)","cited_arxiv_id":null,"evidence_quote":"supplies the optimized coefficients for the explicit pseudo-time smoother used inside the multigrid cycle."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the high-order implicit time-integration and inexact Newton framework with adaptive tolerances that the solver follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the forcing-term criterion used to select the GMRES tolerance adaptively."},{"cited_title":"PhD thesis, Universit¨ at Freiburg (2003)","cited_arxiv_id":null,"evidence_quote":"introduces the well-balanced perturbation formulation used to handle the gravity source term in the atmospheric test cases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the rising warm bubble test case used for the second validation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the density current benchmark used for the third validation."}],"review_version":1}