{"id":"df9d1a45-5a9c-4522-bb97-ff97b8804ed0","arxiv_id":"2607.22455","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For the slotted-cylinder advection benchmark, quasi-monotone limiters best control extrema at the cost of diffusion, while SUPG does not improve over the unstabilized baseline on the tested final-time error metrics.","lead":"This report benchmarks streamline-upwind (SU) and streamline-upwind Petrov–Galerkin (SUPG) stabilization against hyperdiffusion and a quasi-monotone limiter for a slotted-cylinder tracer advection test on the sphere, implemented in ClimaCore.jl and Firedrake. On this test, the limiter best suppresses unphysical extrema but smears the tracer, while SUPG preserves global structure best among stabilized runs yet still does not beat the unstabilized baseline on the tested metrics","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SUPG's failure to beat unstabilized may stem from the lagged residual time-derivative (and its Q-vs-q inconsistency), not from inherent SUPG behavior; implementation is unverified.","rationale":"The reader identified the lagged time-derivative as the weakest assumption, and I agree it is the most load-bearing threat to the headline negative result. I add a further refinement: the text is ambiguous about whether the residual uses ∂_t Q or ∂_t q, and these differ at the discrete level unless the density equation is satisfied exactly. The paper's own §5.2 flags the proxy as inaccurate for large time steps, which is explicit in-scope evidence that the central comparison may be compromised. The negative result is also not accompanied by a reproducibility artifact (no code/commit hash), so the implementation cannot be audited. However, the paper is otherwise honest and well-scoped: it explicitly limits the claim to 'this test case in the computed error metrics' and acknowledges that final-time metrics miss transient behavior. Therefore the reader's CONDITIONAL verdict remains appropriate. I would not escalate to REJECT because the theoretical analysis in §5.1 is standard and the qualitative SU comparison between two independent frameworks provides some confidence; the concern is specific and testable, not a fundamental flaw in the argument. The test I propose would either confirm the artifact or clear SUPG, so the verdict should stay CONDITIONAL pending that check.","tokens_in":13822,"tokens_out":4087,"duration_ms":44451,"concrete_test":"Inspect the ClimaCore.jl branch used for Table 4 (or request a code/commit hash) and verify the exact assembled residual in the SUPG correction: is it ∫ τ (u·∇v) ρ R_adv(q) with R_adv = ∂_t q + u·∇q, and is ∂_t q computed from the previous stage/time step? Then re-run the SUPG case with a time step Δt'=Δt/2 (and update τ accordingly) and compare the ℓ2 and q_over metrics. If the lagged derivative is the cause, halving Δt should move SUPG's errors toward the unstabilized baseline or below it; if it does not, the negative result is robust to this concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (SUPG not improving on unstabilized, Table 4) depends on the SUPG residual being correctly implemented. Section 5.2 states that ∂Q/∂t in the stabilization term is 'computed from the previous time step as a proxy' and admits this 'suffers from inaccuracy for large time steps.' This proxy introduces an O(Δt) error into the correction term. If this error is comparable to the stabilization benefit, the negative result is an artifact. Moreover, Eq. (6) defines the SUPG correction using ρ_h R_adv(q_h) with R_adv = ∂_t q + u·∇q (Table 1), while §5.2's residual and the asymptotic derivation use ∇·(Qu)+∂_t Q. At the discrete level, where ρ_h does not satisfy continuity exactly (noted in §3.3), these two forms differ by a term proportional to the density residual; the code's actual choice is not stated. If the code accidentally mixes them—and the lagged variable is Q rather than q—the correction is inconsistent with the factored form. The conclusions in §4 and §6 do not carry the caveat that the result is conditional on this proxy. Because no code or commit hash is provided, the measured SUPG behavior cannot be independently checked. This is the weakest load-bearing step.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript compares stabilization strategies for transport-dominated spherical advection using the slotted-cylinder benchmark. SU and SUPG are implemented in ClimaCore.jl, and SU also in Firedrake; these are compared against hyperdiffusion, quasi-monotone limiters, and combinations. The central empirical claim is that quasi-monotone limiters best control unphysical extrema (q_over/q_under ~ 1e-7) at the cost of larger global errors, while SUPG gives the lowest global errors among stabilized runs but does not improve on the unstabilized baseline in the computed metrics. A theoretical section proves semidiscrete conservation for the conservative SUPG formulation, explains why standard linear SUPG is not bound-preserving, and proposes an asymptotic-expansion-based time-dependent SUPG algorithm as future work.","tokens_in":14114,"tokens_out":5244,"duration_ms":58998,"significance":"If the negative SUPG result is robust, it is a useful caution for climate-dycore practitioners: streamline stabilization may not pay off on this benchmark under final-time global-error metrics. The paper has clear strengths: the semidiscrete conservation proof in §5.1 is clean; the stabilization parameter uses the standard Tezduyar–Osawa formula with no fitted parameters; the comparison includes two independent implementations; and the authors explicitly acknowledge limitations regarding final-time-only metrics and the lack of higher-order tests. However, the central empirical claim rests on an implementation detail that the paper itself flags as inaccurate, and the conclusions do not carry that caveat. The theoretical analysis is largely independent of this issue and is sound in its main claims, though the asymptotic algorithm is only proposed, not implemented or tested.","major_comments":[{"comment":"The SUPG negative result depends on the lagged residual time derivative. The text states that ∂Q/∂t in the SUPG correction is 'computed from the previous time step as a proxy' and that this 'suffers from inaccuracy for large time steps.' At the reported Δt = 345.6 s, there is no convergence study in Δt and no comparison with a non-lagged or implicit SUPG implementation. Without evidence that the lagged term is small compared with the stabilization effect, the conclusion 'SUPG was not found to improve upon the no stabilization case' cannot be distinguished from an artifact of the proxy. This caveat is absent from the Abstract and §4/§6 conclusions. The authors should either verify the result with a non-lagged residual or explicitly state that the result is conditional on this implementation.","section":"§5.2, Table 4, Abstract"},{"comment":"There is an inconsistency between the residual forms used in the discrete implementation and in the theoretical derivation. Equation (6) defines the SUPG correction using ρ_h R_adv(q_h), with the 'consistent' option in Table 1 being ∂_t q + u·∇q. In contrast, §5.2 describes the implemented stabilizing term as involving ∇·(Qu) + ∂Q/∂t, with ∂Q/∂t lagged. At the discrete level, where ρ_h does not satisfy the continuity equation exactly (as acknowledged in §3.3), these two forms differ by a term proportional to q_h times the density residual. The code's actual choice is not stated. If the lagged variable is Q rather than q, the correction is not consistent with the factored form in Eq. (6). The manuscript should clarify which residual is actually used and, if necessary, rerun the SUPG case with the consistent advective residual.","section":"§3.3, Eq. (6), Table 1, §5.2"},{"comment":"The central conclusions overstate the reliability of the SUPG comparison. The statements that SUPG 'does not improve on the no stabilization baseline for any of the metrics' and 'was not found to improve upon the no stabilization case' are presented without the caveat that the result is conditional on the lagged-residual proxy. Because no code or commit hash is provided, the implementation cannot be independently checked. The manuscript should either provide the implementation details needed for reproducibility or soften the claims so that they are explicitly scoped to the implemented lagged form.","section":"§4, §6"}],"minor_comments":[{"comment":"The Firedrake/ClimaCore SU comparison reports ℓ∞ = 4.50 versus 6.17×10⁻¹, a factor of about 7.3. The text says these differ by 'a single order of magnitude'; this is borderline. More importantly, the cause of this discrepancy is not investigated, which weakens the claim that the two implementations provide a strong consistency check.","section":"Table 4, §4"},{"comment":"The finite-difference formula (32) writes ∂Q_i^n/∂t ≈ (Q_i^n − Q_i^{n−1})/Δt, but Step 2 of the algorithm uses ∂Q_0^{n+1}/∂t computed from Q_0^n and Q_0^{n+1}, which corresponds to (Q_0^{n+1} − Q_0^n)/Δt. The indexing should be made consistent.","section":"§5.2, Eq. (32)"},{"comment":"The sentence 'Substituting into (19)' appears to refer to the nondimensionalized SUPG equation (29), not to the exact mixing-ratio equation (19). This is a typographical error that should be corrected.","section":"§5.2"},{"comment":"The initial-condition formula uses logical notation such as '∧' without definition and defines D(s) without specifying its domain for negative arguments. This makes the initialization harder to parse than necessary.","section":"§2"},{"comment":"Reference [9] is cited as 'Stabilization Parameters in SUPG and PSPG Formulations' but the year is given as 2003; the actual Tezduyar–Sathe reference should be checked and the full bibliographic details completed. Minor formatting issues appear in several URLs (e.g., 'doi:https://doi.org/...' with line breaks).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a well-organized project report with a clean theoretical section, but the main empirical claim is not yet fully supported because of the lagged-residual SUPG implementation. If the authors can supply a non-lagged SUPG run, a Δt-convergence check, or an explicit demonstration that the proxy error is negligible, the negative result would be persuasive. Without that, the paper should be substantially qualified. The scope is more of a technical report than a traditional research article, but the negative result is of potential interest to the community if the implementation issue is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-scoped computational study from a graduate modeling camp, with a clean theoretical section and a useful cross-framework consistency check. But the central negative result—SUPG not improving on unstabilized advection—rests on an implementation detail the paper itself flags as inaccurate, and the conclusions don't carry that caveat. Take the empirical claims conditionally until the code is available or the residual is fixed.\n\nWhat's actually new: the specific comparison of SU/SUPG against limiters and hyperdiffusion in ClimaCore on the slotted-cylinder test, the Firedrake SU cross-check, and a proposed asymptotic SUPG algorithm (not implemented). The theoretical analysis in §5.1 is standard but correct: the conservation proof via the constant test function is clean, and the distinction between conservation and pointwise boundedness is worth stating. The authors also honestly acknowledge that the metrics are final-time only and that higher-order tests are missing.\n\nWhere the paper gets shaky: the SUPG result. Section 5.2 explicitly says the time derivative in the residual is 'computed from the previous time step as a proxy' and that this 'suffers from inaccuracy for large time steps.' That is exactly the setup used for Table 4. If the proxy error is the same order as the stabilization benefit, the measured non-improvement could be an artifact. The conclusions in §4 and §6 do not mention this caveat. In addition, the paper never states precisely which residual form the ClimaCore code uses—the factored ρ R_adv(q) with ∂_t q from §3.3, or the conservative ∇·(Qu) + ∂_t Q used in the §5.2 asymptotic derivation. At the discrete level these differ by a density-residual term, and the lagged variable appears to be Q, not q. No code or commit hash is provided, so a reader cannot check which one was implemented. That makes the headlining empirical claim load-bearing but unverified.\n\nThe asymptotic algorithm is a reasonable idea but unvalidated, and the expansion parameter ε ≈ 4×10⁻³ is only loosely justified. Fine as future-work speculation.\n\nFor whom: researchers working on FEM-based atmospheric dycores who want a quick map of trade-offs on one benchmark. It is not a definitive study. A serious referee could engage with this, but only with the request to supply code, correct the SUPG residual, and rerun the benchmark. As is, the empirical conclusion should not be cited as established.","headline":"A competent camp report whose headline negative result about SUPG is provisional because the implementation uses an acknowledged lagged time-derivative proxy and no code is provided.","tokens_in":14612,"tokens_out":3638,"would_cite":false,"duration_ms":40259,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65M12","76M10","86A10"],"pacs":[],"model":"deepseek-v4-flash","headline":"On the slotted-cylinder advection benchmark, streamline-upwind Petrov-Galerkin stabilization does not improve final-time errors over the unstabilized finite-element baseline, while quasi-monotone limiters nearly eliminate overshoots and und","keywords":["SUPG stabilization","streamline-upwind","transport-dominated flows","slotted-cylinder advection","quasi-monotone limiters","hyperdiffusion","finite element method","asymptotic expansion"],"falsifier":"Run the same slotted-cylinder benchmark with a fully implicit SUPG (or sufficiently small Δt so the time-derivative proxy is accurate) and recompute the error table. If SUPG then improves on the unstabilized baseline in ℓ1 or ℓ2, the paper's negative verdict is implementation-driven; if it still does not, the verdict is robust.","tokens_in":13731,"feed_emoji":"🌊","tokens_out":5061,"duration_ms":53542,"temperature":0.7,"pith_summary":"This paper tries to settle how different stabilization strategies behave on a standard deformational-flow benchmark: advection of slotted cylinders on a sphere. Its empirical claim is that SUPG, despite being the theoretically favored streamline-diffusion method, shows no error-metric improvement over no stabilization at final time in this test, while quasi-monotone limiters reduce spurious extrema to near machine precision at the price of smearing the tracer field. It also proves that the conservative semidiscrete SUPG formulation conserves total mass but does not preserve pointwise bounds or TVD, so the observed ringing is structural. Finally it proposes an asymptotic-expansion SUPG algorithm that replaces the explicit previous-step proxy for the time derivative with a series of explicit Galerkin solves.","feed_headline":"SUPG does not beat unstabilized advection on slotted-cylinder test","feed_subtitle":"Quasi-monotone limiters tame overshoots but smear shape; SUPG holds sharpest stabilized structure with no metric gain.","key_machinery":"The key mechanism is the Petrov-Galerkin test-function perturbation v + τ u·∇v, where τ is a stabilization parameter assembled from advective, temporal, and diffusive timescales in an inverse-norm formula. The SUPG correction adds residual-based diffusion along streamlines only; the tracer equation is stabilized in mass form with the advective residual weighted by density. The paper's asymptotic proposal replaces the previous-time-step proxy for ∂Q/∂t with an expansion Q = Q0 + εQ1, so each time step solves two explicit Galerkin problems instead of one implicit SUPG problem.","core_discovery":"The central discovery is comparative and partly negative: for the slotted-cylinder test with the supplied resolution and error metrics, the SUPG stabilization term (added mass-weighted to the tracer residual) does not beat the unstabilized Galerkin baseline in ℓ1, ℓ2, ℓ∞, overshoot, or undershoot; it does, however, give the smallest global errors among stabilized runs and preserves sharp features. Quasi-monotone limiters alone or in combination drive overshoot and undershoot down to about 1e-7 but roughly double the ℓ1 and ℓ2 errors. The paper's theoretical section explains why: SUPG is consistent residual-based dissipation that conserves total mass but is not a monotone or bound-preserving","pith_inferences":["If the explicit previous-step proxy for ∂Q/∂t introduces O(Δt) error, a fairer test of SUPG would use small time steps or an implicit treatment; the no-improvement verdict may be implementation-limited rather than intrinsic.","Tracking error at every time step (not just final) would likely show transient differences, since the benchmark returns to the initial state at final time, and may reveal when each method's bias enters.","The nondimensional stabilization size ε ≈ 4e-3 suggests SUPG's correction is genuinely weak at atmospheric scales; this predicts its effect stays small on this test regardless of implementation, and higher-order polynomial degree or sharper initial data would be needed to expose its benefit.","A testable extension: implement the asymptotic Q0 + εQ1 algorithm and compare against implicit SUPG; if the series converges quickly, it offers near-implicit accuracy at explicit cost."],"forward_implications":["On this benchmark, the quasi-monotone limiter's extrema control comes with large global shape error; no tested method wins on every metric.","Standard linear SUPG conserves total density and tracer mass on the closed sphere, but does not guarantee nonnegativity, a discrete maximum principle, or TVD behavior; the observed overshoots are consistent with the method's structure.","Combining the limiter with hyperdiffusion, SU, or SUPG leaves extrema metrics at the limiter-only level, so the limiter dominates extrema behavior.","An asymptotic-expansion SUPG scheme can in principle include a non-delayed time derivative in the residual at the cost of solving one extra explicit Galerkin problem per expansion term."],"fun_headline_variants":["SUPG fails to beat unstabilized advection in slotted-cylinder test","Quasi-monotone limiters tame overshoots but diffuse structure","SUPG preserves sharp features but yields no error metric gain","Slotted-cylinder test: SUPG no better than plain Galerkin"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The empirical comparison rests on using the previous time step's ∂Q/∂t as a proxy in the SUPG residual; if that lag injects error, the measured non-improvement of SUPG may be an artifact rather than a property of the method.","fun_headline_variants_meta":{"raw":{"variants":["SUPG fails to beat unstabilized advection in slotted-cylinder test","Quasi-monotone limiters tame overshoots but diffuse structure","SUPG preserves sharp features but yields no error metric gain","Slotted-cylinder test: SUPG no better than plain Galerkin"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001242,"raw_usage":{"total_tokens":4921,"prompt_tokens":721,"completion_tokens":4200,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":4121}},"tokens_in":465,"tokens_out":4200,"duration_ms":32792,"temperature":1.0,"reasoning_tokens":4121,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:40:26.490734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same slotted-cylinder benchmark with a fully implicit SUPG (or sufficiently small Δt so the time-derivative proxy is accurate) and recompute the error table. If SUPG then improves on the unstabilized baseline in ℓ1 or ℓ2, the paper's negative verdict is implementation-driven; if it still does not, the verdict is robust.","supporting_citations":[],"review_version":1}