{"id":"8c52f0b9-698b-49e2-88d6-3026d3d56642","arxiv_id":"2508.09589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Treating time as a spatial dimension and preconditioning the all-at-once system with space-time multigrid yields up to 52x speed-up for transient thermal topology optimization, demonstrated up to 4.2 billion degrees of freedom.","lead":"This paper solves transient heat-conduction topology optimization by treating time as a third dimension and solving the whole space-time system at once, accelerated by a multigrid preconditioner. It reports up to 52x wall-clock speed-up over time-stepping and solutions with 4.2 billion unknowns on a supercomputer.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 52× speed-up rests on an extrapolated time-stepping baseline that is not a runnable 1280-step configuration; a direct comparison at full time horizon is needed.","rationale":"The reader's weakest_assumption identifies the same extrapolated time-stepping baseline, and I agree it is the central load-bearing concern. The other flagged issues (semi-coarsening transfer and energy conservation for time-varying designs) are real but secondary: the headline claim is about speed-up, and the 52× figure depends directly on the comparison baseline. The paper is candid about the uncertainty, memory limitations, and the fact that actual 100-iteration runs give lower speed-ups, but that candor does not remove the need for a direct, runnable time-stepping comparison. The proposed test would settle whether the linear scaling assumption holds and whether the advertised speed-up is achievable against a feasible baseline. Thus the correct verdict remains CONDITIONAL: the method is promising and the solver results are valuable, but the central speed-up claim needs stronger empirical support before full acceptance.","tokens_in":28246,"tokens_out":5555,"duration_ms":58608,"concrete_test":"Run the time-stepping reference for Example 1 (640×640 spatial mesh) with Nt = 20, 40, 80, 160, 320, 640, and 1280 time steps, each for 100 design iterations (or at least 10 for the largest if resources are limited), on the minimum number of nodes needed to fit the stored forward/adjoint data. Record wall time and node-hours. Fit time vs Nt and compare the extrapolated value at Nt=1280 with the measured value. Also compare the measured 1280-step/100-iteration time to 855 s and to the reported space-time times. If the measured time is significantly above the linear extrapolation, or requires extra nodes whose overhead is not counted, then the 52× speed-up is not supported; if it matches, the scaling premise is confirmed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest quantitative claim is the 52.1× speed-up over 'traditional time-stepping' reported in Section 4.3. That number compares a 16.4 s space-time run on 400 nodes with an 855 s time-stepping reference. The reference is not a measured 1280-step run: Section 4 states that only 10 time steps and 10 design iterations were run, and the 855 s figure is obtained by multiplying the 10-step time by 128, based on a claimed linear scaling. This is a load-bearing premise for the headline claim, and it is insecure for three reasons. (i) The 1280-step time-stepping job does not fit in the memory of the node(s) used for the reference; the paper admits this. Thus the baseline is an extrapolation to an infeasible configuration of the actual code, and the comparison excludes the memory cost (or the extra nodes/checkpointing) that a real 1280-step run would incur. (ii) Linearity in the number of time steps is asserted, not demonstrated across the relevant range, and per-step cost includes storage/adjoint recomputation effects that can be non-linear. (iii) The comparison uses only 10 design iterations; the paper itself notes that iteration counts grow later in the optimization, and its more realistic 100-iteration estimate in Section 5.1.2 yields 18.6–20.8×, not 52×. The claim is therefore conditional on a reference that has not been run as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a space-time topology optimization framework for transient thermal conduction. Time is treated as an extra spatial dimension, the governing equation is discretized with a stabilized continuous Galerkin space-time finite element method, and the resulting all-at-once system is solved with FGMRES preconditioned by a parallel space-time multigrid method using semi-coarsening. The framework is implemented on distributed-memory supercomputers and demonstrated on two benchmark problems, including time-constant and time-varying design fields, with up to 4.2 billion DOFs. The central quantitative claim is a speed-up of up to 52.1× over traditional time-stepping, at moderate increases in core-hour cost.","tokens_in":28690,"tokens_out":5176,"duration_ms":60495,"significance":"If the reported performance is representative, this is a significant contribution: it is among the first demonstrations of parallel-in-time multigrid preconditioning inside large-scale transient topology optimization, and it provides measured scalability data on a leadership-class machine. The paper is transparent about many caveats, lists all solver parameters, reports iteration counts and timings for multiple material cases, and includes a clear validation against a prior coarse-mesh result. The reproducibility-oriented details and the honest discussion of limitations (memory, extrapolated baselines, energy conservation for time-varying designs) strengthen the paper. However, the headline speed-up is conditional on an extrapolated and infeasible time-stepping baseline, and the semi-coarsening criterion is justified only by preliminary tests for the new discretization. These points do not invalidate the method, but they must be addressed before the central claims can be taken at face value.","major_comments":[{"comment":"The headline 52.1× speed-up is based on a time-stepping reference that was not run at the full time horizon. Section 4 states that only 10 time steps and 10 design iterations were executed for time-stepping, and the 855-second reference is obtained by scaling linearly to 1280 time steps. The paper itself notes that a real 1280-step run exceeds single-node memory, so the comparison excludes memory pressure and any checkpointing/extra-node costs that a feasible run would incur. Linear scaling in the number of time steps is asserted, not demonstrated across this range, and per-step cost can grow through adjoint storage and later optimization iterations. Notably, when the authors supply a more realistic 100-iteration estimate in Section 5.1.2, the speed-up drops to 18.6–20.8×. The 52× figure should either be replaced by a measured full-horizon comparison or be explicitly presented only as a","section":"Section 4, Section 4.3 (Fig. 9), Section 5.1.2"},{"comment":"The multigrid convergence and all scaling results rely on the semi-coarsening criterion (D_eff, λ_eff, λ_crit=0.5), but the manuscript reports only that the formulas were validated for a finite-element-in-space/backward-Euler discretization in prior work [69] and that 'preliminary tests' indicate they work for the CG-STFE discretization. This is a load-bearing assumption: the reported iteration counts in Tables 3 and 5, and therefore the timing comparisons, depend on the chosen coarsening path. The paper should provide quantitative evidence for the CG-STFE case, e.g., convergence factors for representative meshes and material contrasts, or a sensitivity study with respect to λ_crit.","section":"Section 3.2.1, Algorithm 1"},{"comment":"The scaling and speed-up comparisons in Section 4 are performed with only 10 design iterations. Section 5.1.2 and Figure 21 show that the number of linear iterations increases substantially later in the optimization, especially for the adjoint solve and for high-contrast designs. Therefore, the Section 4 speed-ups do not directly translate to a full optimization run. The authors acknowledge this, but the abstract and highlights nevertheless state the 52× figure without this qualifier. I recommend reporting the full-optimization timings as the primary comparison and clearly labelling the 10-iteration numbers as a solver-level scaling study.","section":"Section 4.1, Section 5.1.2, Figure 21"},{"comment":"For the time-varying design examples, the paper states that energy conservation is not ensured because the time derivative of the volumetric capacity is omitted, and that this likely explains the non-physical temperature behaviour in Example 2C. Since time-varying designs are presented as a full capability of the proposed framework, this limitation should be addressed more prominently in the abstract and results: either implement the missing capacity-derivative term, or explicitly restrict the claims for time-varying designs to 'illustrative' rather than physically validated. At minimum, the potential impact on the optimized designs (Figs. 17–20) should be discussed in the main text, not only in the limitations section.","section":"Section 5.2.2 and Section 6.1"}],"minor_comments":[{"comment":"The caption refers to 'Example 2A', but the surrounding text (Section 5.2.2) describes Example 2B with a time-varying design. The caption should be corrected.","section":"Figure 18"},{"comment":"The artificial diffusion coefficient is introduced as \\hat{k}_{ad} in Eq. (20) but the notation in Eq. (21) and surrounding text is not fully consistent. Please define the symbol once and use it uniformly. Also, the text contains typographical artifacts such as 'P ´eclet' and 'di ffusion'.","section":"Section 3.1.1, Eq. (21)"},{"comment":"Reference [65] contains a typo in the title ('fro' should be 'for'). Please also check the rendering of accented characters throughout the reference list.","section":"References"},{"comment":"The abstract and highlights state 'up to 52× speed-up' without the caveats that Section 4 itself acknowledges. Consider adding a qualifier such as 'projected' or 'for a 10-iteration comparison' so that the summary matches the body's more cautious statements.","section":"Abstract and Highlights"},{"comment":"The unusual sub-node scaling and the performance drops at 140/160 nodes are left unexplained. Even a brief discussion of possible partitioning/load-balancing effects would help the reader interpret the scaling curve.","section":"Section 4.1 and Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of CMAME and builds on a coherent line of the authors' own prior work [65, 69]. The novelty—applying the previously developed semi-coarsening STMG strategy to a (2+1)D CG-STFE topology optimization framework with large-scale runs—is adequate if the performance claims are reframed. The main risk is that the abstract/highlights advertise 52×, while the body's own more realistic estimate is 18.6–20.8× and the baseline is extrapolated. I would encourage the editor to request the full-horizon or clearly qualified comparison described in the major comments before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real message in this paper is that a parallel space-time multigrid-preconditioned CG formulation can solve transient topology optimization problems at an unprecedented scale. The 4.2-billion-DOF demonstration on LUMI is genuine, the semi-coarsening hierarchy is integrated and tested across several material contrasts, and the authors are unusually honest about their own limitations: they flag the energy-conservation issue for time-varying designs, they cite the \"twelve ways to fool the masses\" paper, and they give conservative estimates in Section 5.1.2. That is more transparency than most HPC papers offer. The core claim is new, and the engineering is real.\n\nThe soft spots are all centered on the headline number. The 52.1x speed-up compares a 16.4 s space-time run against a time-stepping reference that was linearly extrapolated from 10 time steps to 1280. The 1280-step run does not fit on the node used for the baseline, so the comparison excludes the memory cost or extra nodes a real run would need. The paper admits this but still leads with the number. The 100-iteration estimate in Section 5.1.2 gives 18.6-20.8x, which is a much more defensible claim for a full optimization, and it is still a strong result. The paper would be better if that became the headline and the 52x figure appeared as a best-case scaling data point with the extrapolation clearly labeled.\n\nTwo secondary issues. First, the semi-coarsening criterion is imported from previous work by the same authors with only \"preliminary tests\" for the CG-STFE discretization. In practice the solver works well, so this is a minor concern. Second, the code is not released; the data statement promises it later. For a paper with a strong computational claim, source availability matters, and I would want at least a reference implementation before taking the 4.2B DOF numbers at face value.\n\nWho should read this: researchers in topology optimization, parallel-in-time methods, and anyone who needs transient thermal design at scale. The methodology section is well written, and the comparison with time-stepping is instructive even where it is imperfect. This deserves a serious referee and attention in revision, not a desk rejection. I would suggest accepting peer review but requiring a rerun or relabeling of the baseline before publication.\n\nFor your reading group: maybe, once the code is out. I would not cite the 52x number in my own work yet, but I would cite the framework as evidence that parallel-in-time approaches for topology optimization of transient diffusion can work at scale.","headline":"Solid proof-of-concept with a defensible ~20x speed-up; the 52x headline rests on a baseline that was never actually run.","tokens_in":29075,"tokens_out":1877,"would_cite":false,"duration_ms":23027,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65F10","65Y05","74P15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating time as a spatial dimension and solving the whole transient heat-conduction problem at once with a space-time multigrid preconditioner cuts time-to-solution for thermal topology optimization by up to 52x and scales to 4.2 billion d","keywords":["topology optimization","space-time finite elements","space-time multigrid","parallel-in-time","transient heat conduction","semi-coarsening","large-scale computing","adjoint sensitivity analysis"],"falsifier":"Run the comparison problem (Example 1) with the reference time-stepping solver at the full 1280 time steps and 100 or more design iterations, with enough memory and identical tolerances, and compare wall-clock time at 120 and 400 nodes. If the ratio of space-time to time-stepping time is not close to the quoted 25x and 52x, the extrapolated speed-up was the artifact. Separately, a sweep of the threshold $\\lambda_{\\mathrm{crit}}$ for the CG-STFE discretisation would test whether the semi-coarsening rule is transferable; poor convergence at $\\lambda_{\\mathrm{crit}}=0.5$ would weaken the method's","tokens_in":28203,"feed_emoji":"🔥","tokens_out":10507,"duration_ms":95696,"temperature":0.7,"pith_summary":"This paper is a proof-of-concept for a parallel-in-time route to transient thermal topology optimization: instead of walking through time one step at a time, it discretizes the transient heat equation on a space-time mesh and solves the entire (2+1)-dimensional problem as one large linear system. The solve is handled by a Krylov method preconditioned with a space-time multigrid V-cycle whose coarsening direction is chosen by the local anisotropy of the diffusion problem. On two benchmark problems, the authors report wall-clock reductions over a backward-Euler time-stepping reference of up to 52x, at 2.4x to 7.7x the core-hour cost, and they solve systems with up to 4.2 billion unknowns in roughly 17 minutes on a distributed-memory machine. If these numbers transfer beyond the two benchmarks, the method offers a credible way to optimize time-dependent designs at scales previously reserved for static problems.","feed_headline":"52x speed-up: space-time solve cuts heat design time to minutes","feed_subtitle":"Treating time as an extra dimension replaces serial time steps with one parallel solve, at a few times the core-hour cost.","key_machinery":"The central object is the all-at-once discrete system $\\mathbf{J}\\mathbf{s}=\\mathbf{f}$ from a stabilised continuous Galerkin discretisation with time as a third coordinate and artificial time-direction diffusion $\\hat{k}_{\\mathrm{ad}}=\\tfrac12 C\\Delta t$. The solver is FGMRES preconditioned by a space-time multigrid V-cycle. The hierarchy follows a semi-coarsening rule: compute $\\lambda_{\\mathrm{eff},l-1}=D_{\\mathrm{eff}}\\Delta t_{l-1}/\\Delta x_{l-1}^2$; if below $0.5$, double $\\Delta t$ (coarsen in time), else double $\\Delta x$ (coarsen in space). This keeps cheap pointwise-style smoothers effective on the anisotropic parabolic operator, which is what makes the all-at-once solve fast enoug","core_discovery":"Central claim: a stabilised continuous Galerkin space-time finite element method, solved all-at-once by a Krylov solver preconditioned with a semi-coarsened space-time multigrid, cuts time-to-solution for transient thermal topology optimisation by over an order of magnitude. Evidence: 52.1x speed-up over a backward-Euler reference (16.4 s vs 855 s), 4.2 billion DOFs solved in ~17 minutes, near-linear scaling from 8 to 400 nodes. Semi-coarsening is load-bearing: state iterations fall 9.7 to 5.6, adjoint 55.6 to 9.4. Time-varying designs are demonstrated, with the authors noting energy conservation is not yet enforced.","pith_inferences":["Extension: the authors' own Section 6.1 caveat that time-varying designs do not conserve energy means the morphing-structure results are best read as demonstrations of solver capability, not yet as physically predictive designs; adding the missing $\\partial C/\\partial t$ term is a natural next step.","Extension: the comparison normalizes by wall-clock time, not by energy or monetary cost; if a user's real constraint is core-hours or budget, the optimum operating point may be the 25x-at-4.8x-cost configuration rather than the 52x-at-7.7x-cost one.","Extension: the 52x figure depends on a time-stepping baseline that saturates at around 1000 cores because the 2D spatial problem is small; on problems with genuinely large 3D spatial meshes, time-stepping would scale better, so the space-time advantage might shrink or move to different core counts.","Extension: a direct extension suggested by the paper's own framing is to tune coarsening strategies specifically for the adjoint equation, which the results show is harder to converge than the state equation; that could cut the dominant cost in optimisation."],"forward_implications":["A transient thermal optimisation that takes roughly 15 hours with time-stepping is claimed to finish in under an hour on the same class of machine at about 2.4x the core-hours, changing how many design iterations a developer can afford per day.","Spatiotemporal resolutions of billions of unknowns (e.g., 1280x1280x2560 elements) become solvable in minutes, bringing time-dependent design closer to the scale of static large-scale topology optimisation.","Time-varying design fields, not just time-constant ones, become optimisable, as shown by conductive spirals tracking a moving heat source.","The semi-coarsening choice is worth roughly two-fold fewer state iterations and six-fold fewer adjoint iterations than full coarsening on the fine benchmark, making the preconditioner's hierarchy design the practical key to the speed-up.","Because the underlying equation is a linear diffusion equation, the authors expect the same machinery to extend to mass diffusion, charge diffusion, and porous flow, where transient optimisation faces the same time-to-solution bottleneck."],"supporting_citations":[{"why":"Supplies the semi-coarsening strategy and effective-anisotropy formulas on which the multigrid hierarchy is built.","marker":"[69]"},{"why":"Defines the benchmark problem, material parameters, and the one-shot Parareal baseline that motivates the comparison.","marker":"[65]"},{"why":"Introduces space-time multigrid, the method class used as preconditioner here.","marker":"[66]"},{"why":"Documents semi-coarsening as the remedy for slow space-time multigrid convergence on parabolic problems.","marker":"[68]"},{"why":"Provides the parallel topology-optimisation code infrastructure that this implementation extends.","marker":"[10]"},{"why":"Defines FGMRES, the flexible Krylov solver wrapped around the multigrid preconditioner.","marker":"[87]"},{"why":"Supplies GMRES, used as both smoother and coarse-grid solver inside the V-cycle.","marker":"[88]"},{"why":"Provides the stabilisation principle (Peclet-unity artificial diffusion) used in the space-time discretisation.","marker":"[83]"}],"fun_headline_variants":["Time as extra dimension gives 52x faster heat optimization","52x speed-up: space-time method replaces serial time steps with one parallel solve","Parallel space-time multigrid solves 4.2B DOFs, 52x faster","All-at-once space-time solve: 52x faster thermal topology optimization","Space-time finite elements deliver 52x speed-up in transient heat design"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The headline speed-ups assume that the time-stepping reference cost grows linearly from the measured 10 steps / 10 iterations to 1280 steps / 100 iterations, and that the semi-coarsening threshold chosen for an earlier discretisation remains the right one for the new space-time Galerkin method; if either gives way, the quoted ratios overstate the gain.","fun_headline_variants_meta":{"raw":{"variants":["Time as extra dimension gives 52x faster heat optimization","52x speed-up: space-time method replaces serial time steps with one parallel solve","Parallel space-time multigrid solves 4.2B DOFs, 52x faster","All-at-once space-time solve: 52x faster thermal topology optimization","Space-time finite elements deliver 52x speed-up in transient heat design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001286,"raw_usage":{"total_tokens":5093,"prompt_tokens":749,"completion_tokens":4344,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":4243}},"tokens_in":493,"tokens_out":4344,"duration_ms":32586,"temperature":1.0,"reasoning_tokens":4243,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:57:03.843701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the comparison problem (Example 1) with the reference time-stepping solver at the full 1280 time steps and 100 or more design iterations, with enough memory and identical tolerances, and compare wall-clock time at 120 and 400 nodes. If the ratio of space-time to time-stepping time is not close to the quoted 25x and 52x, the extrapolated speed-up was the artifact. Separately, a sweep of the threshold $\\lambda_{\\mathrm{crit}}$ for the CG-STFE discretisation would test whether the semi-coarsening rule is transferable; poor convergence at $\\lambda_{\\mathrm{crit}}=0.5$ would weaken the method's","supporting_citations":[{"cited_title":"Space-Time Multigrid Methods Suitable for Topology Optimisation of Transient Heat Conduction","cited_arxiv_id":"2505.10168","evidence_quote":"Supplies the semi-coarsening strategy and effective-anisotropy formulas on which the multigrid hierarchy is built."},{"cited_title":"One-shot Parareal Approach for Topology Optimisation of Transient Heat Flow","cited_arxiv_id":"2411.19030","evidence_quote":"Defines the benchmark problem, material parameters, and the one-shot Parareal baseline that motivates the comparison."}],"review_version":1}