{"id":"15a96e96-109b-4056-9abd-8d51df41c1be","arxiv_id":"2508.04484","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A dynamical low-rank deterministic solver for proton transport reproduces full-rank dose calculations at much lower cost and matches TOPAS Monte Carlo in homogeneous and heterogeneous media.","lead":"This paper presents a fast deterministic method for calculating proton dose in radiation therapy, using a mathematical trick called dynamical low-rank approximation to shrink the problem size. It reproduces full-rank solver results and matches Monte Carlo simulations closely, at a small fraction of the computational cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neglect of straggling for the collided flux (Eq. 16) is the least-secure condition: if energy-loss fluctuations after first collision are significant, the predicted Bragg peak width/depth—and the headline TOPAS-MC agreement—would be wrong, and the paper offers only post-hoc qualitative support.","rationale":"The paper's central contribution is a DLRA evolution for the collided flux, and its strongest validation claim is close agreement with a full-rank reference and with TOPAS MC. The full-rank agreement is convincing as evidence that the low-rank projection does not lose information: ranks are genuinely low, and the algorithm follows the reference solution closely. However, the full-rank reference solves exactly the same approximate model, so it cannot test the physical approximation in Eq. (16). The TOPAS MC comparison is the only external check of that approximation, but it is qualitative (surface plots and profiles, no gamma index or error norm) and the stopping powers themselves were extracted from TOPAS, so the comparison is not fully independent. The paper's own Discussion (§6) concedes that the straggling neglect is an approximation supported mainly by 'good agreement' and future work is needed to include a straggling term in DLRA. In the heterogeneous case, the observed deviations from TOPAS (underestimated peak dose, wrong local minimum behind the first peak) are exactly the kind of signature that energy-loss straggling would produce. A concrete numerical experiment that restores the collided straggling term in a full-rank solver would settle whether this approximation is responsible. Since the reader's weakest assumption identified the same point, I agree with the conditional verdict; no verdict change is needed. I also note the apparent inconsistency in Table 1's absolute memory figures (41.4 TiB for 1.6×10^8 unknowns), but the theoretical system-size comparison already shows a large reduction, so this is a secondary concern.","tokens_in":17440,"tokens_out":9812,"duration_ms":112279,"concrete_test":"Run the full-rank solver (or an equivalent deterministic discretization) on the homogeneous and heterogeneous test cases with the collided straggling term restored in (23) — i.e., dropping assumption (16) and adding the second-order energy derivative discretized with the same SIPG/energy scheme used in the ray tracer — and compare depth-dose/lateral profiles to the current split model and to TOPAS MC. If the dose difference from including collided straggling is within the existing DLRA-vs-full-rank error (or MC statistical uncertainty), the assumption is validated; if it exceeds that tolerance, the headline accuracy claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim ('good accuracy with respect to TOPAS MC') rests on the collided-uncollided split in §3.1, specifically Eq. (16): ∂²(T ψc)/∂E² ≈ 0. The DLRA evolution (23) for the collided flux contains no straggling term; the paper's only justification is the qualitative agreement with TOPAS MC and reference [32] (Discussion, §6: 'neglecting straggling in the collided part of the equation appears to be a reasonable approximation'). This is not a derived asymptotic limit, and it is not benchmarked in isolation. The full-rank reference solver uses the same approximation, so the excellent DLRA/full-rank agreement (Figs. 3,5,7,8) validates only the low-rank projection, not the physics. If collided protons (which have already undergone scattering and travel longer paths) experience non-negligible energy-loss straggling, the energy spread near and beyond the Bragg peak would be mispredicted; the observed deviations in the heterogeneous case (underestimated peak dose, local minimum behind the first peak) are consistent with this possibility. Because the highest-resolution runs are feasible only in low rank, the high-resolution accuracy claim depends entirely on this unverified assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dynamical low-rank approximation (DLRA) method for deterministic proton transport dose calculations. The authors split the solution into uncollided and collided parts, treat the uncollided part with a ray tracer that includes straggling, and evolve the collided part with a rank-adaptive DLRA integrator while neglecting straggling in that part (Eq. (16)). They combine high-order spatial upwinding and spherical-harmonics angular discretizations with a mixture model for heterogeneous materials, and test both the Boltzmann operator and a Fokker-Planck approximation. Numerical experiments compare DLRA against an equivalent full-rank solver on coarse grids and against TOPAS MC at higher resolutions, for homogeneous, heterogeneous, and two-beam cases. The central claims are that DLRA reproduces full-rank results with significantly lower rank and cost, enables finer resolutions, and achieves 'good accuracy' relative to TOPAS MC.","tokens_in":17766,"tokens_out":2997,"duration_ms":36564,"significance":"If fully validated, the method would be a significant step toward practical deterministic proton dose calculation, potentially enabling high-resolution simulations on clinical workstations and joint treatment of multiple beams. The paper has notable strengths: the DLRA/full-rank agreement is demonstrated on independent coarse-grid experiments for both Boltzmann and Fokker-Planck, the implementation is openly available, and the multi-beam rank behavior is an interesting and useful observation. However, the physical accuracy claim relative to TOPAS MC currently rests on a partly circular comparison (stopping powers extracted from TOPAS in §2.1.1) and on an unverified neglect of straggling in the collided flux (Eq. (16)). The absence of quantitative error metrics further weakens the central claim. These issues are addressable but require additional benchmarking and reporting.","major_comments":[{"comment":"The assumption ∂²(T ψ_c)/∂E² ≈ 0 is load-bearing for the paper's accuracy claim against TOPAS MC. The full-rank reference solver uses the same approximation, so the excellent DLRA/full-rank agreement in Figs. 3, 5, 7, 8 validates only the low-rank projection, not the physics. The paper's only justification is qualitative agreement with TOPAS and reference [32] (Section 6). Please provide a quantitative test of this assumption, e.g., a one-dimensional or simplified version of the collided equation with the straggling term included, compared against the current model, or a TOPAS run with collided-part straggling disabled. Without such evidence, the high-resolution accuracy claim is conditional on an unverified approximation.","section":"3.1, Eq. (16)"},{"comment":"The comparison to TOPAS MC is partly circular. Section 2.1.1 states that stopping powers are extracted from TOPAS MC, and Section 4 compares the computed dose to TOPAS MC. Agreement with TOPAS is therefore partly built in. Please state explicitly which physical parameters come from TOPAS and which are independently modeled, and quantify the sensitivity of the dose to the stopping-power source. An independent benchmark using ICRU/NIST stopping-power data for at least one test case would remove this concern.","section":"2.1.1 and Section 4"},{"comment":"No quantitative error metrics are reported. Claims such as 'good accuracy' or 'reasonable agreement' are supported only by visual inspection of Figs. 3, 5, 7, 8. Please provide, for each test case, quantitative measures such as relative L2 error or gamma-index pass rates (e.g., 1%/1 mm or 2%/2 mm) against TOPAS MC, and for the DLRA/full-rank comparison. This is necessary to support the central claim and to make the Boltzmann vs. Fokker-Planck comparison meaningful rather than anecdotal.","section":"Section 4"},{"comment":"The paper reports negative dose values behind the Bragg peak in several tests. These negative regions are dismissed as 'slightly negative', but their magnitude and spatial extent are not reported. Since dose is a non-negative physical quantity, negative values of clinically relevant magnitude would undermine the clinical applicability claim. Please quantify these artifacts (e.g., minimum value relative to peak dose) and discuss whether they can be eliminated or bounded by the chosen discretizations or rank-adaptivity.","section":"4.1, 4.2, Figs. 3(e), 5(e), 8(e)"}],"minor_comments":[{"comment":"The text says 'Γ_out;i and Γ_out;i' where the second should presumably be Γ_in;i. Please correct the notation.","section":"3.1, after Eq. (17)"},{"comment":"The theoretical complexity for the Fokker-Planck case is reported as 101% of full rank, yet runtime is only 1.2% because of GPU acceleration. This makes the 'theoretical' column misleading without an explicit statement that the comparison mixes GPU and CPU implementations. Please clarify or report CPU-only DLRA times.","section":"Section 5, Table 1"},{"comment":"The 'average rank' used for theoretical cost estimates is not precisely defined (average over all time/energy steps? weighted by step size?). Please define it precisely.","section":"Section 5"},{"comment":"The statement that full-rank high-resolution is infeasible is clear, but the resulting dependence of the fine-grid accuracy claim on extrapolation should be stated again in the abstract or conclusion, where the claim 'feasible at much higher resolutions' might be read as implying validated accuracy at those resolutions.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid numerical-methods contribution with a useful DLRA formulation, but the physics-accuracy claim needs strengthening. The TOPAS circularity and the unvalidated straggling assumption are the main concerns; both are fixable with additional experiments and quantitative reporting. I would not recommend rejection because the DLRA/full-rank validation is genuine and the method is novel. The journal's scope fits, though the clinical-application framing should be tempered until independent benchmarking is provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, useful extension of the group's earlier DLRA work to protons. The new pieces—ray-traced uncollided transport with straggling, the mixture material model, higher-order spatial discretization—are real and the numerical experiments show DLRA matches the full-rank solver on coarse grids with genuinely low ranks. The cost table is striking: Boltzmann DLRA at under 0.05% of full-rank runtime and 0.004% memory. That part is convincing.\n\nThe soft spot is the physics validation. The collided equation drops the straggling term (Eq. 16) and the paper's only justification is the qualitative TOPAS MC agreement. The full-rank reference uses the same approximation, so the DLRA-vs-full-rank agreement only tells you the low-rank projection is faithful, not that the model is right. Since the high-resolution runs are only feasible in low rank, the 'good accuracy vs TOPAS' claim depends entirely on that unverified assumption. The observed underestimation of peak dose and the local minimum behind the first peak in the heterogeneous case are consistent with missing straggling in the collided part. That doesn't sink the paper—the assumption is physically plausible and the agreement is reasonably close—but it means the accuracy claim is conditional, not established.\n\nAlso worth flagging: stopping powers are taken from TOPAS MC and then compared to TOPAS MC, so the comparison validates the transport solver, not the stopping-power model. There are no quantitative error metrics (gamma index, for example), only profiles and surfaces. Negative dose artifacts appear and are attributed to angular discretization. Code is claimed openly available but no repository link or hash is given, which is a reproducibility gap.\n\nWho should read it: people working on deterministic transport in medical physics and on DLRA for kinetic equations. It's a solid numerical methods paper that deserves a serious referee. I'd recommend sending it out. The referee should ask for quantitative error metrics and a sensitivity test on the collided-straggling assumption—maybe a small comparison where straggling is included in a simplified way, or a test with a different straggling model—before the physics claim is taken as established.","headline":"Solid extension of DLRA to proton transport with real computational gains, but the physics claim rests on an unverified straggling assumption that the paper only supports post hoc.","tokens_in":18230,"tokens_out":2144,"would_cite":true,"duration_ms":25254,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M70","82C70","65F55","92C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"Deterministic proton dose calculations can match Monte Carlo accuracy at clinical resolution by evolving the collided flux as a low-rank matrix with energy as pseudo-time, cutting memory by orders of magnitude.","keywords":["dynamical low-rank approximation","proton therapy","deterministic transport","Boltzmann equation","Fokker-Planck approximation","model order reduction","ray tracing","dose calculation"],"falsifier":"In the homogeneous water phantom at 0.25 mm resolution, run a full-rank deterministic reference that keeps the full straggling term in the collided equation and compare depth-dose profiles with the low-rank code's output; if the difference in the Bragg-peak region exceeds a few percent of peak dose, or if a 30 MeV beam (where straggling is proportionally larger) shows low-rank dose deviating from TOPAS MC by more than the full-rank code does, the Eq. (16) assumption fails.","tokens_in":17338,"feed_emoji":"⚛️","tokens_out":10276,"duration_ms":97392,"temperature":0.7,"pith_summary":"This paper claims that the linear Boltzmann equation governing proton transport in tissue can be solved deterministically at clinically relevant spatial resolution by evolving a low-rank representation of the collided particle flux with energy as pseudo-time. Uncollided protons are handled by a ray tracer that retains energy straggling, while the collided part advances without straggling on the manifold of low-rank matrices. The method reproduces the output of a full-rank deterministic code with ranks as small as 5–20, reducing complexity from $O(n\\cdot m)$ to $O(r^2(n+m))$ and memory from $O(n\\cdot m)$ to $O(r(n+m))$. At high resolution it agrees reasonably with TOPAS Monte Carlo in homogeneous and heterogeneous materials, and several beam angles can be computed together with little extra cost.","feed_headline":"Low-rank solver reproduces proton dose with 0.004% of memory","feed_subtitle":"Deterministic transport at clinical resolution runs on one GPU, and multi-angle beams add little rank.","key_machinery":"The engine is the rank-adaptive augmented basis-update & Galerkin (BUG) integrator for dynamical low-rank approximation. The discretized flux matrix $u(t)\\in\\mathbb{R}^{n\\times m}$ is kept in factorized form $X(t)S(t)V(t)^\\top$ with orthonormal bases, and the dynamics are projected onto the tangent space of the rank-$r$ manifold; the augmented BUG integrator updates the basis matrices in parallel, forms a Galerkin system for $S$, and truncates singular values after each step. The enabling split is the collided–uncollided decomposition: the uncollided flux is advanced by a deterministic ray tracer that keeps the straggling term, while the collided equation, in which straggling is dropped via","core_discovery":"What the paper establishes is that the cost of deterministic proton dose calculation is set not by the physics of transport but by the representation of the solution. By splitting the angular flux into an uncollided part that is ray-traced (with straggling) and a collided part that is advanced without straggling on the manifold of rank-$r$ matrices, the authors turn the energy variable into a pseudo-time and apply a rank-adaptive augmented BUG integrator. In the homogeneous 90 MeV beam test, the Boltzmann-based solver reproduces the full-rank reference with an average rank of 4.77, and at 0.25 mm spatial / $P_{115}$ angular resolution it agrees with TOPAS Monte Carlo within a few percent exc","pith_inferences":["The authors validate the no-straggling collided assumption only by agreement with TOPAS MC; a quantitative a priori estimate of the neglected $\\partial^2(T\\psi_c)/\\partial E^2$ term would show whether the method transfers to lower energies or higher-Z materials where straggling is relatively larger.","The reported speedups mix algorithm gains with GPU implementation effects (the full-rank baseline runs on CPU); a constant-error comparison against optimized Monte Carlo on the same hardware would better locate the method's clinical value.","The appearance of small negative dose regions behind the Bragg peak suggests the modal $P_N$ expansion plus upwinding is not entirely positivity-preserving; a limit on the angular expansion or a correction could close this gap without a major redesign.","If the rank stays this low for anatomically realistic CT-based geometries, DLRA could turn time-dependent treatment-planning questions (e.g., 4D dose with organ motion) into tractable deterministic simulations."],"forward_implications":["Dose distributions with full angle and energy resolution become affordable for treatment-planning problems, enabling Monte-Carlo-free uncertainty quantification and beam configuration optimization.","The Boltzmann test case's theoretical memory drops from 41.4 TiB to 1.69 GiB (0.0039%), so a single workstation GPU can run what previously required a supercomputer-scale dense solve.","Because ranks rise only slightly when a second beam at a different angle is added (roughly +2 for Boltzmann), many beamlets can be bundled into a few angle groups instead of thousands of pencil-beam runs.","The Boltzmann operator with the extended transport correction outperforms the Fokker-Planck approximation in accuracy and rank at high resolution, suggesting the cheaper operator is not the right trade once DLRA removes the cost penalty.","The same code structure can absorb nuclear interactions and absorption without changing the low-rank machinery, since those effects enter only the collision operator."],"supporting_citations":[{"why":"Establishes the low-rank manifold evolution and tangent-space projection that the whole reduction is built on.","marker":"[24]"},{"why":"Supplies the rank-adaptive augmented BUG integrator used for the energy/pseudo-time stepping.","marker":"[7]"},{"why":"Prior DLRA solver for electron therapy; the collision-source split and energy-as-time trick that this paper extends to protons.","marker":"[27]"},{"why":"Low-rank power-iteration method for neutron criticality; its material decomposition is adapted here for tissue mixtures.","marker":"[28]"},{"why":"Deterministic proton transport with ray tracing; the basis for the uncollided-part treatment and the argument that neglecting collided straggling is reasonable.","marker":"[32]"},{"why":"TOPAS, the Monte Carlo code against which the method's dose accuracy is validated.","marker":"[40]"},{"why":"Validates the Molière scattering model against TOPAS; justifies the differential cross section used in the physics model.","marker":"[3]"},{"why":"The extended transport correction applied to the Boltzmann scattering moments.","marker":"[11]"},{"why":"Angular Fokker-Planck decomposition whose correction stabilizes the FP moments (together with [37]).","marker":"[30]"},{"why":"The deterministic adjoint-based algorithm and tissue mixture model used for the ray tracer and material composition.","marker":"[2]"}],"fun_headline_variants":["Proton dose at 0.004% memory via low-rank","Low-rank cuts proton dose memory 25,000x","Rank-5 transport matches full-rank dose","Low-rank transport handles heterogeneous proton dose"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method rests on the assumption that energy-loss straggling can be dropped from the collided equation, $\\partial^2(T\\psi_c)/\\partial E^2\\approx 0$ (Eq. (16)); if the term is non-negligible, the energy spread of scattered protons and hence the Bragg peak will be biased.","fun_headline_variants_meta":{"raw":{"variants":["Proton dose at 0.004% memory via low-rank","Low-rank cuts proton dose memory 25,000x","Rank-5 transport matches full-rank dose","Low-rank transport handles heterogeneous proton dose"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001673,"raw_usage":{"total_tokens":6476,"prompt_tokens":753,"completion_tokens":5723,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":5670}},"tokens_in":497,"tokens_out":5723,"duration_ms":47211,"temperature":1.0,"reasoning_tokens":5670,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:55:53.698959+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the homogeneous water phantom at 0.25 mm resolution, run a full-rank deterministic reference that keeps the full straggling term in the collided equation and compare depth-dose profiles with the low-rank code's output; if the difference in the Bragg-peak region exceeds a few percent of peak dose, or if a 30 MeV beam (where straggling is proportionally larger) shows low-rank dose deviating from TOPAS MC by more than the full-rank code does, the Eq. (16) assumption fails.","supporting_citations":[{"cited_title":"Koch and C","cited_arxiv_id":null,"evidence_quote":"Establishes the low-rank manifold evolution and tangent-space projection that the whole reduction is built on."},{"cited_title":"Ceruti, J","cited_arxiv_id":null,"evidence_quote":"Supplies the rank-adaptive augmented BUG integrator used for the energy/pseudo-time stepping."},{"cited_title":"Kusch and P","cited_arxiv_id":null,"evidence_quote":"Prior DLRA solver for electron therapy; the collision-source split and energy-as-time trick that this paper extends to protons."},{"cited_title":"Lathouwers","cited_arxiv_id":null,"evidence_quote":"Deterministic proton transport with ray tracing; the basis for the uncollided-part treatment and the argument that neglecting collided straggling is reasonable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TOPAS, the Monte Carlo code against which the method's dose accuracy is validated."},{"cited_title":"Burlacu, D","cited_arxiv_id":null,"evidence_quote":"Validates the Molière scattering model against TOPAS; justifies the differential cross section used in the physics model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The extended transport correction applied to the Boltzmann scattering moments."},{"cited_title":"Landesman and J","cited_arxiv_id":null,"evidence_quote":"Angular Fokker-Planck decomposition whose correction stabilizes the FP moments (together with [37])."},{"cited_title":"Burlacu, D","cited_arxiv_id":null,"evidence_quote":"The deterministic adjoint-based algorithm and tissue mixture model used for the ray tracer and material composition."}],"review_version":1}