{"id":"0fb98977-7982-4e26-812d-52bae6dc70c0","arxiv_id":"2507.18200","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A tolerance-controlled 'blocks' method accelerates the non-history temporal superposition algorithm for finite line sources, cutting precomputation cost by orders of magnitude while keeping errors below a user-set threshold.","lead":"This paper presents a numerical acceleration for heat-flow simulations of borehole heat exchanger fields, cutting the expensive precomputation phase by up to four orders of magnitude for line-to-line interactions. The method ignores heat loads that have not yet influenced a target beyond a user-set error tolerance, enabling faster quadrature and, potentially, simulations of hundreds of boreholes over decades of hourly time steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Error-tolerance claim rests on a surrogate integrand (Eq. 29) that selects Gauss-Kronrod nodes; no bound connects surrogate oscillation structure to true quadrature error, so the a priori ε/K guarantee is unproven for untested geometries.","rationale":"The reader's weakest assumption is the same one I would flag: Eq. (29) is the only link between the requested tolerance and the actual quadrature mesh. Section 3.4 propagates ε/K into the adaptive algorithm, but quadrature error control is only meaningful for the integrand actually sampled; the true v(ζ) is evaluated later at nodes chosen from the surrogate. The internal algebra (block recurrence, change of variables, Appendix B cutoffs) checks out, and the paper is transparent about the surrogate's roughness, so this is not a soundness accusation. It is a correctness-risk concern: the headline claim 'for all tolerances tested, ∥ε∥∞ < ε' is empirical, not a priori, and the tested space is narrow (aligned equal-length lines, σ/r_b from 10 to 1000, ε ≥ 1e-12). Real borefield configurations include partially overlapping and offset lines, where the piecewise α/β structure in (26) can produce oscillation bands absent from (29). The proposed test would settle whether the surrogate-selected nodes are numerically sufficient. Secondary limitations (first block uses the method of the unpublished preprint [12]; the combined-time speedup statement in the Conclusions is contradicted by Tables 1-2 for some cases) are real but do not change the central verdict: the method is plausible and well-engineered, but the error-control guarantee should remain conditional pending a broader test or a rigorous surrogate-to-truth error bound. Thus UNCHANGED relative to the reader's CONDITIONAL verdict.","tokens_in":26650,"tokens_out":13789,"duration_ms":141218,"concrete_test":"Reproduce the Table 1 / Figure 13 precomputation error test for line-to-line configurations outside the tested parameter range: (a) offset depths D_s=0, H_s=1500, D_t=300, H_t=600, σ=1 m; (b) non-overlapping lines D_s=0, H_s=1500, D_t=1600, H_t=300, σ=10 m; at ε=10^-10 and ε=10^-14. Use the same surrogate-selected nodes and weights from (29)-(30), and compare the computed temperature to the direct convolution (32). If any ∥ε∥∞ exceeds the requested tolerance, the a priori error-control claim is falsified; if all pass, the surrogate concern is empirically bounded for these cases. As a second leg, run an adaptive integrator directly on the true v(ζ)/ζ^2 to see whether it selects materially different nodes for the same tolerance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an a priori error-controlled precomputation (Section 3.4: each block contributes at most ε/K). The nodes ζ_s used in (30) are generated by adaptive Gauss-Kronrod applied to the surrogate I_k(ζ) in (28), where v is replaced by the two-frequency trapezoidal approximation (29), which the paper explicitly calls 'a rough estimation.' The adaptive algorithm's error estimate bounds only the surrogate's quadrature error; it says nothing about the error of the true integrand evaluated at the same nodes, namely the product of the Gaussian source-block envelope, the full geometry factor v(ζ)/ζ^2, and s_k(ζ). The true line-to-line geometry factor (26) contains a piecewise-constant α(˜r)/sqrt(˜r^2−σ̃^2) term plus the β term, with up to five pieces (46); the surrogate (29) captures only endpoint sine frequencies and can miss interior oscillation bands. Consequently, the claimed ε/K tolerance is not a proven bound on the actual quadrature error. The empirical validation (Figures 11–13) covers aligned lines with σ̃ ≥ 10 and ε down to 10^-12; it does not exercise partially overlapping or offset line pairs, non-uniform line lengths, or tolerances below 10^-12. If a missed frequency band aliases into the ζ integral, the error can exceed the requested tolerance without the adaptive routine detecting it, which would break the headline claim of enforcing a user-set error tolerance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a \"blocks\" method to accelerate the non-history temporal superposition algorithm for finite line source (and point source) geothermal heat-transfer calculations. The load history is partitioned into time blocks; each block is associated with a spatial influence region determined by a user-set error tolerance ε, exploiting the approximate causality of heat diffusion. The resulting integrands are less oscillatory, enabling a combination of adaptive Gauss-Kronrod quadrature and asymptotic expansions for the ζ integral and for the geometry factors. The authors claim that the method enforces a user-set error tolerance, reduces precomputation cost by up to four orders of magnitude compared with the original non-history implementation in the line-to-line case, and maintains linear complexity in the number of time steps. The method is validated by comparison with direct discrete convolution for point-to-point, line-to-point, and line-to-line configurations over a range of distances and tolerances.","tokens_in":26785,"tokens_out":6999,"duration_ms":69142,"significance":"If correct, the claimed speedup would materially advance borefield-scale simulations by removing the precomputation bottleneck that made the non-history method impractical for hundreds of sources and hundreds of thousands of hourly steps. The manuscript has several genuine strengths: the block recurrence (7) is cleanly derived, the change of variables leading to (16) removes the endpoint singularity, the asymptotic/Gauss split (23) is sensible, and Appendix B provides concrete cutoff bounds. The error tolerance is an input rather than a fitted constant, and the empirical comparison uses an independent reference (the direct discrete convolution (32)), which strengthens the honesty of the error study. The complexity analysis and timing tables give a concrete picture of where the method helps and where it does not (the per-step simulation phase is more expensive than the original). However, the central error-control claim is not fully proven: adaptive node selection for blocks k>1 is based on a surrogate integrand, and the first block falls back on the very method the paper criticizes for lacking a priori error control.","major_comments":[{"comment":"The claimed a priori error tolerance of ε/K per block is not established for the true integrand. The Gauss-Kronrod nodes ζ_s are generated by applying the adaptive algorithm to the surrogate I_k(ζ) built from the trapezoidal two-frequency approximation (29), which the manuscript itself calls \"a rough estimation.\" The actual precomputed quantities H_{k,s} in (30), however, use the true geometry factor v(ζ)/ζ² evaluated at those nodes. An adaptive quadrature error estimate controls only the error of the surrogate integrand; it provides no bound on the error of the true integrand at the same nodes. Since the true line-to-line factor contains a piecewise-constant α-term and β-term with up to five pieces (Eqs. (26), (45)–(46), Appendix C.2), while (29) has only two endpoint sine frequencies, a missed interior frequency band could push the true quadrature error above ε/K without the adaptive routine detecting it. The manuscript needs either a rigorous bound relating the surrogate and true integrands, a post-adaptive check on the true integrand with node refinement, or a clearly stated restriction of the guarantee to the surrogate with supporting evidence for a wider geometry class.","section":"Sections 3.3–3.4, Eqs. (28)–(30), Eq. (29)"},{"comment":"The empirical validation is too narrow to support the unrestricted claim \"for all tolerances ϵ tested, ‖ε‖∞ < ϵ.\" All line-to-point and line-to-line tests use aligned source and target (D̃_s=D̃_t, H̃_s=H̃_t) with σ̃ ≥ 10; there are no tests with partially overlapping or offset line pairs, unequal line lengths, or tolerances below 10⁻¹². These are precisely the configurations where the surrogate approximation (29) is most likely to miss oscillation bands. Please add such tests, or qualify the error-tolerance claim accordingly; if no guarantee can be provided for these cases, the conclusions should state the applicable domain.","section":"Section 5.1, Figures 11–13"},{"comment":"The method as presented does not achieve a priori error control for the first block. Section 3.5 sends k=1 back to the Bakhalov-Vasil'eva method of [12], which the Introduction (Section 2.1) identifies as lacking a priori error control; Section 5 then explicitly excludes s1 from the error study. For self-interactions and close interactions, where s1 dominates, the user-set tolerance is therefore not enforced by the new machinery. The authors should either extend error control to the first block, or state in the Abstract and Conclusions that the tolerance guarantee applies only when s1 is negligible (and quantify when this holds).","section":"Section 3.5 and Section 5, first paragraph"}],"minor_comments":[{"comment":"Standardize the transliteration of the Russian authors' names: \"Bakhalov-Vasil'eva\" in the Abstract, \"Bakhalov-Vasil'eva\" in Sections 3.5 and Appendix A, and \"Bakhvalov\" in reference [13] should be unified.","section":"Throughout"},{"comment":"The bullet heading \"Line to point (Section 3.2.3)\" should be \"Line to line\"; the discussion refers to the line-to-line integral in Section 3.2.3.","section":"Section 5.2, third bullet"},{"comment":"Typos: \"compelxity\" (Section 5.2), \"intergrals\" (Introduction), \"tollerance\" (Figure 1 caption), \"naïve\" (Introduction), and \"Reisz-Fréchet\" (Appendix C.2) should be corrected.","section":"Throughout"},{"comment":"Section 5.1 uses ε both for the user tolerance and for machine precision (\"ε ≃ 2 · 10⁻¹⁶\"); rename the latter, e.g., ε_mach, to avoid ambiguity.","section":"Section 5.1"},{"comment":"Figure captions 12 and 13 state D̃=0, H̃=1500 while Section 4 uses D=0, H=150 m; clarify whether these are dimensionless with respect to r_b and define the normalization.","section":"Figures 12 and 13"},{"comment":"Equation (36) defines a synthetic load without units; state that q' is in W/m (or another unit) or that the scaling is arbitrary.","section":"Equation (36)"},{"comment":"Tables 1–3 should state whether the reported times are medians over repeated runs and specify the CPU/compiler environment; this aids reproducibility of the timing comparison.","section":"Tables 1–3"},{"comment":"The sentence \"for all tolerances ϵ tested, ‖ε‖∞ < ϵ is kept below the specified tolerance\" is redundant and should be rephrased, e.g., \"in all tested cases, the maximum error stayed below the prescribed tolerance.\"","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The companion paper [12] is under review and is used for the first block and for the baseline comparison; if [12] is not accessible to the editor, the reproducibility of the first-block claims and the comparison baseline is affected. This is not a blocker for the k>1 analysis, but the editor may want to verify that the preprint is publicly available. The main technical risk is the surrogate-based node selection; I would like to see the error-control issue addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real methods paper with a usable speedup and one loosely supported error-control claim. It deserves a serious referee; I'd send it out, with the tolerance guarantee as the main revision point.\n\nThe genuinely new material: the block decomposition of load history with tolerance-based causal truncation, the per-block reduction of the effective line-source region, the replacement of Bakhvalov-Vasil'eva quadrature by asymptotic plus adaptive Gaussian quadrature, and the coarea-formula derivation of the line-to-line geometry factor. The core algebra checks out: block recurrence (7) is right, the change of variables to (16) is right, and the Appendix B cutoff bounds go in the conservative direction. The empirical study is honest and useful: errors are measured against direct discrete convolution, not against a fitted constant, and Figures 11-13 show errors below the requested tolerance over a good range of distances and tolerances. The precomputation speedups for line-to-line, three to four orders of magnitude, are credible and genuinely matter for borehole field simulation.\n\nSoft spots, in rough order of size. First, the a priori error guarantee is not actually a proven guarantee for the true integrand. The zeta nodes are chosen by adaptive Gauss-Kronrod on the surrogate (29), a two-frequency trapezoidal approximation the authors themselves call rough; the true geometry factor is then evaluated at those nodes. Nothing bounds the difference between surrogate quadrature error and true quadrature error. The tests cover aligned lines with sigma-tilde >= 10 and tolerances down to 1e-12; they don't cover partially overlapping or offset line pairs, non-uniform lengths, or tighter tolerances. This is a real gap, though not one that makes the empirical speedup suspicious.\n\nSecond, the first block is excluded from the validation and still uses the unpublished [12] machinery. This is disclosed, but it means the claim \"enforces a user-set tolerance\" is really about blocks k>1 plus an inherited method. Third, the conclusion that combined precomputation plus simulation time is lower in the blocks method is not supported by Tables 1-2 for point-to-point and line-to-point at loose tolerances; the original is faster there. Minor, but the sentence should be fixed. Fourth, no code is shipped and the comparison baseline is the authors' own implementation of [12], which is understandable but worth noting.\n\nNone of this changes the core verdict. The method is internally consistent, the error study is extensive, and the speedup is load-bearing for multi-source, multi-decade simulations. The right outcome is peer review, not desk rejection, with a request to either prove or visibly soften the epsilon/K guarantee and to test the off-aligned geometries where the surrogate could miss oscillations.","headline":"Genuine speedup and honest benchmarks, but the a priori error guarantee is proven only for a surrogate integrand; send to review with that as the main revision point.","tokens_in":27471,"tokens_out":3258,"would_cite":true,"duration_ms":34223,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D30","80A19"],"pacs":["02.60.Jh","44.10.+i"],"model":"deepseek-v4-flash","headline":"A tolerance-based notion of causality removes the precomputation bottleneck in non-history time superposition for finite line source thermal simulations, cutting line-to-line precomputation cost by up to four orders of magnitude.","keywords":["finite line source","non-history temporal superposition","blocks method","borehole heat exchangers","highly oscillatory integrals","error tolerance","heat conduction simulation","precomputation acceleration"],"falsifier":"Run the line-to-line blocks method on a geometry outside the paper's test set, for example with lateral distance $\\sigma = 2$ m, a 20-year hourly simulation, and tolerance $10^{-14}$, then compare against the direct discrete convolution of the loads with the step response evaluated in double precision. If the maximum absolute error $\\|\\varepsilon\\|_\\infty$ exceeds the requested tolerance, the claimed a priori error control does not hold in that regime.","tokens_in":26246,"feed_emoji":"🔥","tokens_out":6254,"duration_ms":65264,"temperature":0.7,"pith_summary":"This paper tries to remove the precomputation bottleneck in the non-history temporal superposition method for finite line source thermal simulations used in borehole heat exchanger fields. Its argument is that although the heat equation's fundamental solution is non-causal, for any chosen error tolerance one can treat a load as not yet felt at a target until it has had time to propagate; the method, called the blocks method, splits each source's load history into blocks separated by such effective causal time lags. Because each older block carries a Gaussian decay in the spectral variable $\\zeta$ and only a smaller region of a line source needs to be integrated, the oscillatory integrands become tractable for adaptive Gauss–Kronrod quadrature and an asymptotic expansion, replacing the Bakhalov–Vasil'eva integrator that was slow and lacked a priori error control. The paper reports that the resulting precomputation is several orders of magnitude faster, up to about four in the line-to-line case, while keeping the maximum error below a user-supplied tolerance, which would make simulations with hundreds of boreholes and hundreds of thousands of hourly time steps practical.","feed_headline":"Causality cuts borehole heat precomputation by 10,000x","feed_subtitle":"A blocks method preserves error control while turning the costly line-to-line precomputation from minutes into milliseconds.","key_machinery":"The load-bearing object is the block decomposition of the source function combined with the effective causality lag. A block $s_k$ is the sum of loads in a window of the past weighted by Gaussians, and it can be advanced by a recurrence that only requires the boundary loads entering and leaving the block. The lag $N_{i\\to j}(\\epsilon)$ is computed by solving $f(r_{ij}, N_{i\\to j}\\Delta t) \\leq \\epsilon$, where $f$ is the error of truncating the step response to a limited line region. For each block, the $\\zeta$-integration nodes are fixed before simulation by running adaptive Gauss–Kronrod quadrature on a two-frequency trapezoidal surrogate of the oscillatory line factor, and the true geometry factor is then evaluated at those nodes to build the precomputed weights. The line-geometry integrals themselves are computed with a combination of Gaussian quadrature for slowly oscillating regions and an asymptotic expansion in powers of $1/\\omega$ for rapidly oscillating regions, with the split point chosen by an a priori error bound; only the first, most recent block keeps the original Bakhalov–Vasil'eva treatment.","core_discovery":"The paper claims that causality, understood as a tolerance-based cutoff, can be built directly into the non-history marching scheme. It defines blocks $s_k(\\zeta, n\\Delta t)$ of the Gaussian-smoothed load history, where block $k$ spans the time steps from $n-N_k$ to $n-N_{k-1}$ and satisfies a recurrence that lets each block be updated cheaply at every time step. For each source–target pair, the interaction only needs blocks older than the causal lag $N_{i\\to j}(\\epsilon)$, and each block's integrand is proportional to $e^{-\\zeta^2 N_{k-1}\\Delta \\tilde{t}}$, suppressing oscillations in $\\zeta$; for line sources the integration region over the source line is also reduced to a block-dependent interval. This makes the $\\zeta$ integrals amenable to fixed Gauss–Kronrod nodes chosen once per simulation from a cheap surrogate integrand, with the true geometry factor evaluated at those nodes. The authors demonstrate numerically for point-to-point, line-to-point, and line-to-line geometries that for all tolerances tested down to $10^{-12}$ the infinity-norm error stays below the requested tolerance, at precomputation costs around four orders of magnitude lower than the original method in the line-to-line case.","pith_inferences":["An implementation that wants the tolerance guarantee for arbitrary geometries should add a cheap validation step: compare the surrogate against the exact geometry factor at a few $\\zeta$ values after node selection, and refine or flag when they disagree in oscillation content, for example near a line's endpoint with small lateral distance.","The same block-and-lag idea transfers to other diffusion problems with oscillatory spatial kernels, such as multi-source heat diffusion in layered or anisotropic media, whenever the kernel admits a bound that defines an influence region.","Because the blocks method is a true marching scheme, it can be used for online simulation and control where future loads are unknown; the higher per-step cost reported in the simulation phase compared with an FFT convolution is the price of that causality and matters less as the number of interacting sources grows."],"forward_implications":["Precomputation for a single line-to-line interaction drops from hundreds of seconds to tens of milliseconds at tolerance $10^{-12}$, reducing the cost that scales with the square of the number of boreholes.","Users set an error tolerance and the method returns a marching scheme whose maximum error is numerically shown to stay below that tolerance, instead of tuning the number of integration points blindly.","The marching structure is preserved: updating source functions costs $O(N_s)$ per time step while temperature evaluation is per interaction, keeping overall time complexity linear in the number of time steps.","Combined, these gains make non-history temporal superposition feasible for systems with hundreds of sources and hundreds of thousands of hourly steps, the regime of multi-decade borefield simulations.","At tolerances tighter than about $10^{-12}$, the number of Gaussian nodes grows rapidly, so the advertised speedup weakens for extreme precision."],"supporting_citations":[{"why":"Defines the original non-history temporal superposition algorithm specialized to point and line sources; the blocks method is an acceleration and error-controlled replacement for its precomputation.","marker":"[12]"},{"why":"Introduces the non-history dependent fast algorithm for ground-coupled heat pump simulations that the new method builds on.","marker":"[10]"},{"why":"Extends the non-history algorithm to arbitrary response factors, establishing the recursive update that the block recurrence generalizes.","marker":"[11]"},{"why":"Provides the Bakhalov–Vasil'eva oscillatory integration method that the new method replaces for all but the first block.","marker":"[13]"},{"why":"Supplies the adaptive Gauss–Kronrod quadrature used to fix the $\\zeta$-integration nodes.","marker":"[14]"},{"why":"Supplies the asymptotic expansion method for highly oscillatory integrals used for the line-to-point and line-to-line geometry factors.","marker":"[18]"},{"why":"Supplies the moment-free framework for highly oscillatory integrals that complements the asymptotic expansion strategy.","marker":"[19]"},{"why":"Provides the coarea formula used to rewrite the line-to-line double integral over the rectangle as a single integral in the distance variable.","marker":"[20]"}],"fun_headline_variants":["Heat-wave causal cutoff speeds borehole sims 10,000x","Causality-inspired blocks slash precompute time","Tolerance-based causality accelerates thermal sims","Fast temporal superposition via causal blocks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The promised per-block accuracy is guaranteed for an approximate two-frequency stand-in for the true line-geometry factor, not for the exact integrand, so a geometry whose true oscillatory structure the approximation misses could violate the stated error tolerance.","fun_headline_variants_meta":{"raw":{"variants":["Heat-wave causal cutoff speeds borehole sims 10,000x","Causality-inspired blocks slash precompute time","Tolerance-based causality accelerates thermal sims","Fast temporal superposition via causal blocks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1594,"prompt_tokens":1102,"completion_tokens":492,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":718,"tokens_out":492,"duration_ms":5335,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:42:14.928561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the line-to-line blocks method on a geometry outside the paper's test set, for example with lateral distance $\\sigma = 2$ m, a 20-year hourly simulation, and tolerance $10^{-14}$, then compare against the direct discrete convolution of the loads with the step response evaluated in double precision. If the maximum absolute error $\\|\\varepsilon\\|_\\infty$ exceeds the requested tolerance, the claimed a priori error control does not hold in that regime.","supporting_citations":[{"cited_title":"A non-history dependent temporal superposition algorithm for the finite line source solution","cited_arxiv_id":"2501.16027","evidence_quote":"Defines the original non-history temporal superposition algorithm specialized to point and line sources; the blocks method is an acceleration and error-controlled replacement for its precomputation."},{"cited_title":"A fast algorithm for the simulation of gchp systems,","cited_arxiv_id":null,"evidence_quote":"Introduces the non-history dependent fast algorithm for ground-coupled heat pump simulations that the new method builds on."},{"cited_title":"A fast algorithm for the hourly simulations of ground-source heat pumps using arbitrary response factors,","cited_arxiv_id":null,"evidence_quote":"Extends the non-history algorithm to arbitrary response factors, establishing the recursive update that the block recurrence generalizes."},{"cited_title":"Integration with control of accuracy,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive Gauss–Kronrod quadrature used to fix the $\\zeta$-integration nodes."},{"cited_title":"Efficient quadrature of highly oscillatory integrals using deriva- tives,","cited_arxiv_id":null,"evidence_quote":"Supplies the asymptotic expansion method for highly oscillatory integrals used for the line-to-point and line-to-line geometry factors."},{"cited_title":"Moment-free numerical integration of highly oscillatory functions,","cited_arxiv_id":null,"evidence_quote":"Supplies the moment-free framework for highly oscillatory integrals that complements the asymptotic expansion strategy."},{"cited_title":"The co-area formula for sobolev mappings,","cited_arxiv_id":null,"evidence_quote":"Provides the coarea formula used to rewrite the line-to-line double integral over the rectangle as a single integral in the distance variable."}],"review_version":1}