{"id":"cba3cedb-ea96-451d-9497-6d0fca91a0b7","arxiv_id":"2509.05001","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A trajectory-aware reduced-order preconditioner fixes the offline-online residual mismatch in ROMSAD and solves parametric RTE in 2-3 online iterations on a lattice test.","lead":"This paper presents a preconditioner to speed up repeated radiative transfer simulations across many parameter settings, combining reduced-order models with classical synthetic acceleration. The method converges in 2-3 iterations on a test lattice problem, which could cut costs in design, uncertainty, and inverse problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Offline trajectory-aware ROMs may not generalize to unseen test parameters; the 2-3 iteration claim rests on 10 random test cases with no error bars.","rationale":"The paper's trajectory-aware construction is internally coherent: if the offline sequence of ROMs is built with the same initial guess and same previous-ROM corrections as online, then for training parameters the residual trajectories are matched. The derivations in Sec. 4 and Appendix A appear sound, and the reported offline costs are plausible. The central risk is empirical generalization: the method's success depends on low-rank representativeness of correction snapshots across the parameter space, an assumption that is stated but never quantified. With only 10 random test parameters and no spread/error-bar data, the '2-3 iterations' claim could conceal substantial variation; a single outlier would make the headline misleading. The reader's weakest_assumption identified exactly this, and my independent reading agrees. The pin-cell snapshot-accuracy discrepancy is an additional unexplained point that could indicate a deeper issue, but it is secondary to the lattice claim. Therefore the reader's CONDITIONAL verdict is appropriate; no adjustment is needed, but the proposed leave-one-out test would either retire or substantiate the main concern.","tokens_in":29072,"tokens_out":13368,"duration_ms":148283,"concrete_test":"Perform leave-one-out cross-validation on the lattice problem: for each of the 121 training parameters, build TAR-IG with basis and initial-guess ROM from the other 120 samples and run the online stage. Record the distribution (min/median/max, and 90th percentile) of transport sweeps. If the max or 90th percentile exceeds 3 sweeps, the '2-3 iterations' headline is not robust. Additionally, report per-test sweep counts for the existing 10 test parameters instead of only averages.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central lattice claim (Sec. 5.4) requires that for every test parameter, the trajectory-aware correction snapshots δf_µ^(l) = f_µ − f_µ^(l) (Eq. 34) lie in the span of the offline SVD basis U_rl^(l) at each aware level l. The paper provides no theoretical bound linking basis dimension to convergence (explicitly deferred in Sec. 6) and validates on only 10 randomly chosen test parameters, reporting solely averages in Table 8. If a test parameter's residual trajectory leaves this span, the ROM correction is inaccurate, and after switching to DSA at Nw+1, DSA may require many sweeps to remove remaining low-frequency error. The paper's Fig. 9 shows only a single test parameter. Separately, Sec. 5.3 contains an unexplained statement: SI-DSA fails to converge, leading to 'snapshots with limited accuracy,' yet TAR-IG still reaches residuals below 1e-12; if offline reference solutions are not actually converged, the correction snapshots are not the ideal corrections and the algorithm's premise is violated. While this pin-cell discrepancy is a serious open question, the lattice headline is most directly threatened by the representativeness/generalization issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes trajectory-aware reduced-order-model (ROM) enhanced synthetic acceleration (SA) preconditioners for parametric radiative transfer equations. Building on the authors' earlier ROMSAD method, it constructs a sequence of ROMs offline, each conditioned on the correction produced by the previous ROM, so that the residual trajectory seen online matches the offline trajectory. The framework is developed for source iteration and for FGMRES, using a reformulation of the ideal correction equation to avoid directly solving the kinetic correction problem. Numerical tests in 1D slab, 2D variable scattering, pin-cell, and lattice configurations report substantial reductions in transport sweeps and wall-clock time relative to SI-DSA and ROMSAD, with the lattice problem claimed to converge in 2--3 iterations online.","tokens_in":29333,"tokens_out":4433,"duration_ms":48082,"significance":"The trajectory-aware construction is a conceptually clean and plausible remedy for the offline/online residual-trajectory mismatch identified in ROMSAD. The algebraic derivations in Appendix A are correct, and the numerical results are consistently in the claimed direction across four test problems. The additional offline cost over ROMSAD is reported as small relative to the cost of generating training snapshots. If the method generalizes as claimed, it would be a valuable tool for multi-query RTE applications. However, the evidence base is narrow: all online claims rest on 10 (or 20) random test parameters, only averages are reported, no error bars or per-parameter distributions are given, and no code or data are provided. The most serious concern is the pin-cell case in §5.3, where the reference solutions used to form correction snapshots are themselves not converged, directly calling into question the premise of the ROM construction. These issues, together with an unsupported amortization claim, prevent the paper from being accepted in its current form.","major_comments":[{"comment":"SI-DSA fails to converge for the pin-cell problem within the 50-iteration limit (residual 1.09e-09 versus tolerance 1e-11), yet the correction snapshots δf^(l)_µ = f_µ − f^(l)_µ in Eq. (34) require the converged solution f_µ. If f_µ is not converged, the snapshots are not the ideal corrections of Eq. (22), and the offline ROM is built on inaccurate data. The paper explicitly acknowledges 'snapshots with limited accuracy' but does not explain why TAR-IG nevertheless reaches 1e-12. This is a load-bearing issue: the key premise of the method is violated in a test that is used to claim robustness. Please either demonstrate that the inaccuracy is benign (e.g., by showing that the reference solutions are converged for all training parameters, or by quantifying the effect of the unconverged reference on the ROM) or use a properly converged reference to generate snapshots.","section":"§5.3, Table 6"},{"comment":"The central performance claims, including the abstract's '2-3 iterations online', are based on averages over only 10 randomly selected test parameters, with no error bars, standard deviations, minima/maxima, or seed information. For example, Table 8 reports ¯nsweep=2.3 and ¯niter=3.4, but the reader cannot tell whether this is a tight distribution or a few lucky cases. Given the small test sets and the strong claims, the paper should report per-parameter convergence counts or a measure of spread, and ideally a larger or multiple-seed test set. This is needed to support the robustness and generalization claims.","section":"§5.2, §5.4, Tables 3 and 8"},{"comment":"The statement that 'even taking the offline computational time into account, our method starts to achieve computational savings when predicting the solution for the first new parameter' is not supported by Tables 5, 7, or 9. Those tables report only the cost of ROM construction relative to one linear solve; the dominant offline cost is the 50 (or 121) training solves needed to generate solution snapshots. A break-even analysis comparing total offline cost plus k online solves against k full-order solves is required to substantiate the multi-query efficiency claim. As written, the claim appears to overlook the fixed training cost.","section":"§5.2, 'Offline efficiency' paragraph"},{"comment":"No theoretical or a posteriori guarantee is provided that the trajectory-aware reduced basis at each level spans the correction manifold for unseen test parameters. The paper explicitly defers theory on the relation between basis dimension and convergence to future work. While this is understandable for a numerical methods paper, the strong online claim makes the lack of any robustness indicator a practical risk. At minimum, add leave-one-out or cross-validation experiments on the training set, or report performance on test parameters near the boundary of the parameter domain, to give some evidence beyond 10 random interior samples.","section":"§4.1, §4.2, Sec. 6"}],"minor_comments":[{"comment":"There are several typos: 'syntehtic' in Section 3, 'gvien' in Section 4.1, 'Anorldi' in Algorithm 5. The caption of Figure 8 says 'Sec. 5.3' but the lattice problem is in Sec. 5.4.","section":"General"},{"comment":"The notation for the reduced basis is inconsistent: Alg. 3 and Alg. 4 use U^(l)_rl, while the text sometimes writes U^(l)_r or U^(l)_rl. Please unify the subscripts and clarify that r_l denotes the dimension at level l.","section":"Notation"},{"comment":"The derivation of η^(l) in Eq. (46) uses H with subscripts k,l−1 and H_{l,l−1}; the indices appear slightly inconsistent with the standard Arnoldi recurrence in Alg. 5. A short explanation of the notation would improve readability.","section":"§4.2.1, Eq. (46)"},{"comment":"The table reports ¯nsweep=2.3 for TAR-IG but the text later says 'converges in just 3.4 iterations' for FGMRES-TAR-IG; please ensure the reader understands how iterations and sweeps differ (Remark 5.1 partially covers this, but a reminder in the table caption would help).","section":"§5.4, Table 8"},{"comment":"The manuscript provides no statement on data/code availability and no description of the random number generation used for selecting test parameters. Including this information would make the numerical results reproducible.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competently written numerical methods contribution and the central idea is interesting. However, the pin-cell issue in §5.3 is a red flag: the method appears to work even when the offline reference solutions are not converged, which is either a remarkable robustness property requiring explanation or an indication that the reported residuals are not fully trustworthy. The absence of error bars and the small test sets also make it hard to judge the strength of the '2-3 iterations' claim. I would advise the editor to request a revision that addresses the convergence of reference solutions, adds uncertainty/robustness quantification, and supplies a proper amortization analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper does what it says. The trajectory-aware construction is a genuine fix for the offline-online residual mismatch that limited ROMSAD, and the authors are scrupulous about crediting Santo et al. and explaining why a direct extension of [32] to RTE is impractical. The derivations in Appendix A are right; the matrix-free-sweep-compatible form and the FGMRES reformulation are the real contributions, and the reported speedups (roughly 4-10x over SI-DSA in the 2D tests, 2-3 iterations on the lattice problem) are consistent across four test problems.\n\nThe soft spots are about evidence, not method. There is no code or data, so independence of the numbers is unverifiable. Test averages are reported without error bars or min/max over the 10 out-of-sample parameters, which matters mainly for the headline claim. That is a fixable weakness, not a fatal one. The more interesting gap is the pin-cell case (Sec. 5.3): SI-DSA fails to converge within 50 iterations, so the paper concedes the offline snapshots have 'limited accuracy,' yet TAR-IG still converges below 1e-12. If the reference solutions used to form delta-f are not converged, then the correction snapshots are not ideal corrections, and the method's premise is violated. The paper gives no explanation. That needs a clear paragraph and probably a re-run with converged references.\n\nOn generalization: the worry that an unseen test parameter's residual trajectory may leave the span of the offline bases is real but is the usual ROM assumption, not a load-bearing flaw. The paper does not offer theory - it explicitly says so - and the empirical out-of-sample results are consistent, but without error bars we cannot tell whether 2.3 mean iterations hides a 5-iteration tail.\n\nWho it's for: anyone working on RTE solvers, synthetic acceleration, or parametric ROMs. It deserves a serious referee. My recommendation: send to review, but require code/data release and per-case statistics, and an explicit account of the pin-cell training snapshots. Fix those and this is a solid accept.","headline":"Trajectory-aware ROM acceleration for RTE is a real, well-scoped improvement over the authors' earlier ROMSAD; the empirical case is plausible but needs code-release and a fix to the pin-cell training-data issue before I'd trust the '2-3 iterations' headline.","tokens_in":29852,"tokens_out":2443,"would_cite":true,"duration_ms":25646,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F10","65F08","65N22","85A25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Trajectory-aware ROM preconditioners solve parametric radiative transfer in just a few online iterations.","keywords":["parametric radiative transfer","synthetic acceleration","reduced order model","trajectory-aware ROM","source iteration","FGMRES","low-rank structure","transport sweeps"],"falsifier":"Choose a test parameter at the edge of (or slightly outside) the training range of the lattice problem, run TAR-IG, and record the residual after each ROM correction and after the switch to DSA; if the residual after the first ROM correction is not smaller than DSA's and convergence takes more than 2–3 sweeps, the paper's central speed claim fails.","tokens_in":28921,"feed_emoji":"⚛️","tokens_out":7548,"duration_ms":76824,"temperature":0.7,"pith_summary":"Solving radiative transfer equations repeatedly for many parameter values is a bottleneck in design, imaging, and uncertainty studies. This paper argues that the usual acceleration strategy—diffusion synthetic acceleration plus a reduced-order model (the ROMSAD scheme)—loses efficiency after the first iteration because the reduced model is built from residual trajectories generated by a fixed preconditioner, while online residuals are generated by the reduced model itself. The proposed trajectory-aware framework removes this mismatch by constructing a sequence of ROMs offline, each one using corrections from the previous ROM-based iteration, so the offline and online residual paths agree. In numerical tests the resulting preconditioner converges in 2–3 online iterations for a parametric lattice problem and substantially beats both DSA and ROMSAD, with only marginal extra offline cost.","feed_headline":"Radiation solves reach 2-3 iterations with trajectory-aware ROMs","feed_subtitle":"By matching offline reduced models to the solver's own residual path, acceleration no longer fades after the first iteration.","key_machinery":"The central object is the trajectory-aware ROM sequence {U_rl^{(l)}}_{l=1}^{N_w}: one reduced basis per iteration, each obtained by SVD of correction snapshots δf_μ^{(l)} = f_μ − f_μ^{(l)} that were produced by running the earlier ROM-based corrections in the offline loop. The associated ROM synthetic-acceleration preconditioner for iteration l is M^{-1}_{ROM,l} = I + C^{-1}_{ROM,l} Σ_s, with C^{-1}_{ROM,l} built from the projected reduced operator (U_rl^{(l)})^T A_μ U_rl^{(l)} and the angular sums U_iso and U_ρ. After N_w iterations the solver switches to DSA. This construction keeps the reduced spaces consistent with the actual residual path of the preconditioned iteration; inside FGMRES,","core_discovery":"The paper's central claim is that the efficiency degradation of ROM-enhanced synthetic acceleration is caused by a trajectory mismatch, not by insufficient basis accuracy. In ROMSAD, correction snapshots δf_μ^{(l)} = f_μ − f_μ^{(l)} are generated offline with a fixed DSA preconditioner, so the basis encodes residuals evolving under DSA; online, the residual path is set by the ROM preconditioner itself, and the two diverge after the first correction. The trajectory-aware framework builds a separate reduced basis for each of the first N_w iterations, computing each new basis from snapshots taken with the preceding ROM-based corrections. This sequential construction makes the offline and online","pith_inferences":["The same sequential-consistency idea could transfer to other parameterized kinetic equations, such as frequency-dependent or anisotropic-scattering transport, where the ideal correction equation is also an expensive kinetic problem.","For non-affine parameter dependence, standard row-sampling hyper-reduction is incompatible with matrix-free transport sweeps, so a practical extension would need an oblique or sampling-free approximation of the parametric operator.","The 2–3 iteration behavior is empirical rather than proven; a scan of test parameters near the boundary of the training region would reveal how quickly the benefit degrades when the correction snapshots leave the offline span."],"forward_implications":["For multi-query RTE workloads such as optimization, inverse problems, and uncertainty quantification, the dominant per-query cost can drop from tens of transport sweeps to a few sweeps.","Because the ROMs are tailored to the preconditioner's own trajectory, the method remains effective under loose POD truncation tolerances, where ROMSAD loses robustness.","The same trajectory-aware construction works for both source iteration and FGMRES, so robust Krylov solvers inherit the acceleration without extra theoretical assumptions.","Offline cost stays dominated by the training solves; the extra N_w transport sweeps per training parameter are reported to be small relative to one full solve, so online savings begin almost immediately in multi-query settings."],"supporting_citations":[{"why":"Defines diffusion synthetic acceleration, the baseline preconditioner that this paper accelerates and then switches to after N_w iterations.","marker":"[9]"},{"why":"Establishes source iteration with synthetic acceleration for discrete-ordinates transport, the iterative framework being preconditioned.","marker":"[5]"},{"why":"Shows Krylov methods with synthetic acceleration are needed for problems with material discontinuities, motivating the FGMRES extension.","marker":"[14]"},{"why":"Introduces ROMSAD, the ROM-enhanced SA preconditioner whose first-iteration-only efficiency motivates the trajectory-aware fix.","marker":"[16]"},{"why":"Extends ROMSAD to a flexible GMRES solver and supplies the reformulation of the ideal correction equation used to build snapshots.","marker":"[17]"},{"why":"Provides the trajectory-aware multi-space reduced-basis preconditioner concept that this paper adapts to RTE, with explicit comparisons.","marker":"[32]"},{"why":"Documents why hyper-reduction is incompatible with matrix-free transport sweeps, shaping the method's online-efficient design.","marker":"[33]"},{"why":"Gives FGMRES, the Krylov method used because it permits changing preconditioners across iterations.","marker":"[53]"}],"fun_headline_variants":["Trajectory-aware ROMs cut radiation solves to 2-3 iterations","Fix residual mismatch: radiation preconditioner hits 2-3 iterations","Matching residual paths accelerates parametric radiation solves","Offline-online alignment boosts radiation solver to 2-3 iterations","Preconditioner fix stops degradation after first radiation iteration"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The correction snapshots built from training parameters must stay representative for unseen test parameters; the paper gives no theoretical bound tying reduced-basis dimension to convergence rate.","fun_headline_variants_meta":{"raw":{"variants":["Trajectory-aware ROMs cut radiation solves to 2-3 iterations","Fix residual mismatch: radiation preconditioner hits 2-3 iterations","Matching residual paths accelerates parametric radiation solves","Offline-online alignment boosts radiation solver to 2-3 iterations","Preconditioner fix stops degradation after first radiation iteration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1305,"prompt_tokens":780,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":439}},"tokens_in":524,"tokens_out":525,"duration_ms":5646,"temperature":1.0,"reasoning_tokens":439,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:42:07.722998+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a test parameter at the edge of (or slightly outside) the training range of the lattice problem, run TAR-IG, and record the residual after each ROM correction and after the switch to DSA; if the residual after the first ROM correction is not smaller than DSA's and convergence takes more than 2–3 sweeps, the paper's central speed claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines diffusion synthetic acceleration, the baseline preconditioner that this paper accelerates and then switches to after N_w iterations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes source iteration with synthetic acceleration for discrete-ordinates transport, the iterative framework being preconditioned."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows Krylov methods with synthetic acceleration are needed for problems with material discontinuities, motivating the FGMRES extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends ROMSAD to a flexible GMRES solver and supplies the reformulation of the ideal correction equation used to build snapshots."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the trajectory-aware multi-space reduced-basis preconditioner concept that this paper adapts to RTE, with explicit comparisons."},{"cited_title":"Behne, J","cited_arxiv_id":null,"evidence_quote":"Documents why hyper-reduction is incompatible with matrix-free transport sweeps, shaping the method's online-efficient design."},{"cited_title":"Saad, A flexible inner-outer preconditioned GMRES al gorithm, SIAM Journal on Scientific Computing 14 (2) (1993) 461–469","cited_arxiv_id":null,"evidence_quote":"Gives FGMRES, the Krylov method used because it permits changing preconditioners across iterations."}],"review_version":1}