{"id":"22cd3cb6-3710-492e-ae66-dd54f22d717d","arxiv_id":"2608.08322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"Across four 3D benchmarks, the best eikonal weight for PINN level-set advection tracks how far the exact solution departs from a signed-distance function, spanning 10^-1 to 10^-5, while only two of four weights transfer unchanged from 2D.","lead":"This paper tests whether an eikonal regularisation weight for physics-informed neural networks, calibrated in two dimensions, carries over to three-dimensional level-set advection. It finds the ordering transfers but the exact values do not, and that the regulariser also reduces seed-to-seed variability.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Protocol differences across the four benchmarks align perfectly with the selected weights, so the SDF-departure ordering in §4.6 may be a confound rather than evidence of mechanism.","rationale":"I read the paper as an honest, reproducible empirical study whose headline ordering is plausible but not yet established as a mechanism. The reader's weakest assumption concerns whether the inherited 8x256 tanh architecture and training protocol remain adequate in 3D. My concern is sharper: the protocol is not even held fixed across the four benchmarks being compared. ZD3D and RV3D receive Fourier features, RAD/RAR, and causal weighting; TR3D and RO3D do not, and this partition coincides exactly with the partition between small and large selected weights. This makes the central 'ordering transfers' claim underdetermined. The paper's own §6.3 limitations support this reading. I nevertheless do not reject the paper: the empirical results are reported transparently, the code is promised on Zenodo, the multi-seed protocol is a real improvement, and the WENO comparison is carefully matched. A CONDITIONAL verdict remains appropriate; I would add a requirement for a within-benchmark control that varies SDF departure while holding protocol fixed. My concrete test is one such control. The reader and I partially agree: the capacity concern is real, but the protocol confound is the more load-bearing issue.","tokens_in":20976,"tokens_out":6791,"duration_ms":62458,"concrete_test":"Repeat the six-weight, three-seed sweep on the reversed-vortex benchmark with the velocity field scaled by α ∈ {0, 0.25, 0.5, 0.75, 1}, keeping RAD/RAR, causal weighting, Fourier features, N_f, N_i, and the training budget exactly as in Table 1 for RV3D. Measure the exact solution's SDF departure, e.g., mean ∥∥∇φ_exact∥−1∥ in the band |φ| < 3Δx at T/2, and record the selected w_eik at each α. If the selected weight does not decrease monotonically as α increases, the four-benchmark ordering in §4.6 is confounded by protocol differences rather than governed by SDF departure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central transferability claim — that the selected eikonal weight tracks the exact solution's departure from the signed-distance property (§4.6, Abstract) — is supported by only four benchmarks that are not run under a common protocol. According to Table 1 and §2.3–2.4, ZD3D and RV3D use random Fourier features, RAD/RAR adaptive collocation, and causal chunk weighting, while TR3D and RO3D use none of these; the reported ZD3D configuration also doubles the sampling budget (§4.4). These protocol differences align exactly with the outcome: the two benchmarks with adaptive/causal machinery select the two smallest weights (10^-3, 10^-5), and the two without select 10^-1. Adaptive sampling concentrates residual points near slot edges or stretched folds, and causal weighting changes the temporal loss landscape; both can alter the optimal eikonal weight independently of the SDF-departure mechanism. The falsifiability statement in §6.3 tests the mechanism but not this confound, and §6.3 itself concedes that transfer to different networks/protocols is untested. The observed ordering is therefore a correlation until a controlled comparison separates SDF departure from protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a systematic three-dimensional extension of earlier two-dimensional work on choosing the eikonal regularisation weight in PINNs for level-set advection. Four benchmarks (translating sphere, rotating sphere, Zalesak slotted sphere, reversed vortex) are each swept over six eikonal weights with three seeds under a pre-registered selection rule; the selected weights are then interpreted as tracking the extent to which the exact solution departs from a signed-distance function (10^-1 for rigid smooth advection, 10^-3 for local departure, 10^-5 for global stretching). The paper also documents seed-to-seed variance reduction from the regulariser, compares the PINN against a WENO5 reference on identical grids and metrics, and introduces a feature-restricted classification error for thin geometric features.","tokens_in":21265,"tokens_out":5928,"duration_ms":52559,"significance":"The experimental protocol is a strength: full training budget per run, three seeds, pre-registered selection rule with the eikonal deviation demoted to a tiebreak, disclosed tie artefacts, stratified L-BFGS analysis, quantified reference error for the reversed vortex, and identical error definitions for the WENO comparison. Reproducible code is promised via Zenodo. If the central ordering claim survives a controlled test, the paper would give practitioners a transferable rule for setting w_eik based on the problem geometry rather than per-benchmark sweeps, and the feature-restricted measure is a useful methodological contribution for thin-interface problems.","major_comments":[{"comment":"The central ordering claim is confounded by protocol differences between the benchmarks. TR3D and RO3D use uniform collocation, no causal weighting, no Fourier features, and Nf=10^4; ZD3D and RV3D use RAD/RAR adaptive sampling, causal chunk weighting, and (for ZD3D) Fourier features and Nf=2x10^4 in the reported configuration. The selected weights partition exactly along these lines: the two benchmarks without adaptive/causal machinery select 10^-1, and the two with it select 10^-3 and 10^-5. Since adaptive sampling and causal weighting can alter the optimal eikonal weight independently of the signed-distance departure of the exact solution, the observed correlation is not yet evidence of the proposed mechanism. The authors should run a controlled comparison that holds the protocol fixed across benchmarks (for example, re-sweeping TR3D or RO3D with RAD/RAR and causal weighting, or ZD3D and RV3D without them), or explicitly restrict the ordering claim to the protocol family used here.","section":"§4.6 and Table 1"},{"comment":"The claim that the regulariser reduces seed-to-seed standard deviation by more than an order of magnitude is only demonstrated for TR3D. For RO3D, Table 3 shows the relative L2 error standard deviation falling from 0.593 at w_eik=0 to 0.205 at the selected 10^-1, a factor of about 2.9, and the 10^-3 row has a standard deviation of 0.557. The unqualified statement in the abstract is therefore not supported by the reported data; the variance-reduction result should be presented per benchmark or restricted to the benchmarks where the effect is actually of that size.","section":"Abstract and §4.3"},{"comment":"The proposed falsifiability criterion tests only the SDF-departure mechanism and not the protocol confound. The reversed-vortex prediction (small selected weight) is also what the protocol hypothesis would predict, because RV3D is one of the two benchmarks with adaptive sampling and causal weighting. The authors should state explicitly what observation would distinguish 'the optimum tracks the exact solution's departure from signed distance' from 'the optimum is shifted by the adaptive/causal training machinery'.","section":"§6.3"},{"comment":"The explanatory variable in the ordering, 'the extent to which the exact solution departs from a signed distance function', is never directly measured or reported. The paper reports the eikonal deviation of the trained network, which is not the same quantity as the SDF departure of the exact solution. Reporting a quantitative measure of the exact solution's departure (for example, the mean or maximum of ||grad phi_exact| - 1| in a band around the interface, or the fraction of the domain where it exceeds a threshold) would make the ordering testable and would allow the reader to verify that ZD3D and RV3D are indeed ordered by this property rather than by other differences.","section":"§4.6 and §3.1"}],"minor_comments":[{"comment":"The dimension of the random matrix B is not stated; specify B in R^{128 x 3} (or the number of input spatial coordinates) so that the Fourier feature output dimension is unambiguous.","section":"Equation (6)"},{"comment":"The axis label 'relative L2 volume' is ambiguous and could be read as a single quantity; label the vertical axes explicitly as 'relative L2 error (%)' and 'volume MAPE (%)', with the axis for each line identified in the caption.","section":"Figures 4, 7, 11, 14"},{"comment":"References [7] and [8] are cited as 2026 works without a journal, volume, or DOI; if they are preprints, please state the repository or arXiv identifier.","section":"References [7] and [8]"},{"comment":"The statement that varying w_eik changes relative L2 error by 'up to a factor of 31' should identify the specific weight pair producing that ratio, since the reader must otherwise infer the comparison from Tables 2-5.","section":"§4.8"},{"comment":"The word 'pre-registered' suggests an external time-stamped registration; consider rewording to 'the selection rule was fixed before the sweeps were run' to avoid an unintended stronger implication.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is careful and transparent, and the experimental design is well above the field average in its handling of seeds, budgets, and disclosed artefacts. The main obstacle to acceptance is the internal validity of the headline ordering claim: the protocol differences between the two groups of benchmarks align perfectly with the selected weights, so a reader cannot yet tell whether the SDF-departure mechanism or the training protocol is responsible. I would require a controlled comparison or a substantial softening of the causal language before recommending publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper does something genuinely useful: it repeats a 2D eikonal-weight study in 3D with three seeds, full training budgets, a selection rule fixed in advance, and it reports the per-run spread. It gives practitioners a concrete answer about where the eikonal weight matters, and it introduces a sign-based feature-restricted error measure that shows global L2 cannot certify thin-feature preservation. The WENO comparison is honest: same grids, same snapshots, same error definitions, and they disclose the boundary-data asymmetry and the no-reinitialisation choice rather than hiding them. That is real credit.\n\nThe soft spot is the load-bearing ordering claim. The selected weight is supposed to track how far the exact solution departs from a signed distance function: 10^-1 for the two smooth rigid spheres, 10^-3 for the slotted sphere, 10^-5 for the reversed vortex. But the benchmarks do not share a protocol. The slotted sphere and reversed vortex use random Fourier features, RAD/RAR adaptive sampling, and causal chunk weighting; the two smooth spheres use none of those. The slotted sphere also runs at double the sampling budget. The protocol differences line up with the selected weights exactly, so the ordering is a correlation until shown otherwise. Section 6.3's falsifiability statement tests the mechanism, not this confound. A controlled comparison—for example, running a smooth sphere with the adaptive/causal machinery, or running the slotted sphere without it—would settle it. Without that, the headline claim is not yet supported.\n\nMinor points: the 'pre-registered' and 'prediction before the final benchmark' statements are self-reported and cannot be verified from the paper; time-stamped pre-registration would help. The WENO comparison without reinitialisation is a defensible like-for-like choice, but it means the classical baseline is not best practice; the authors should add a WENO+reinitialisation column or at least discuss how it changes the margin. Neither of these is fatal.\n\nThis is a paper for the PINN/level-set community and for people who want an error measure that actually sees thin features. It deserves a serious referee. I'd send it, and I'd ask for the controlled protocol experiment and for a more cautious statement of the ordering claim until the confound is addressed.","headline":"A careful, honest empirical study whose headline mechanism claim is plausible but confounded by protocol differences across the four benchmarks.","tokens_in":21829,"tokens_out":2672,"would_cite":true,"duration_ms":23966,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["47.11.-j","02.70.-c"],"model":"deepseek-v4-flash","headline":"A single tuning rule for PINN level-set solvers survives the move from two to three dimensions, but the specific weight values do not.","keywords":["physics-informed neural networks","level-set method","interface advection","eikonal regularisation","loss weighting","signed distance function","three-dimensional benchmarks","WENO comparison"],"falsifier":"Run the same protocol and six-weight sweep on a rigidly transported smooth ellipsoid, whose exact solution is an exact signed distance function at every time, and check whether any weight at or below $10^{-3}$ wins; if so, the claimed ordering fails. Conversely, a stretching benchmark that selects a weight near $10^{-1}$ would falsify it from the other side. The paper itself names exactly this test in its limitations section.","tokens_in":20721,"feed_emoji":"🌀","tokens_out":10545,"duration_ms":77252,"temperature":0.7,"pith_summary":"This paper tests whether a tuning rule discovered in two dimensions still governs physics-informed neural networks (PINNs) when they solve level-set interface advection in three dimensions. The rule holds: the optimal weight of the eikonal regulariser, the penalty pushing the level-set field toward the signed-distance property $\\|\\nabla\\phi\\|=1$, tracks how far the exact solution departs from that property, spanning four decades from $10^{-1}$ where it holds exactly to $10^{-5}$ where the interface is stretched. The specific weight values, however, transfer only benchmark by benchmark: two of the four three-dimensional benchmarks inherit their two-dimensional optimum unchanged and two do not, so inheritance must be verified rather than assumed. A three-seed sweep additionally shows that the regulariser buys reproducibility, cutting seed-to-seed standard deviation by more than an order of magnitude at small weights. Against a fifth-order WENO solver on identical grids, the classical scheme is more accurate on all four benchmarks, though the margin narrows as geometric difficulty increases.","feed_headline":"Eikonal weight rule transfers to 3D, values don't","feed_subtitle":"The selected weight tracks signed-distance departure across four decades; only two of four values carry over unchanged.","key_machinery":"The eikonal regulariser is the loss term $L_{\\mathrm{eik}} = \\frac{1}{N_f}\\sum_i (\\|\\nabla\\hat\\phi(x_i,\\tau_i)\\|-1)^2$ added to the composite PINN loss with weight $w_{\\mathrm{eik}}$, and it is the mechanism that carries the argument. It enforces the signed-distance property that advection destroys, and its weight is the single dominant hyperparameter of the protocol. The paper's core evidence is the monotone relation between the selected weight and a geometrically diagnosable quantity, namely how far the exact solution departs from $\\|\\nabla\\phi\\|=1$, a departure that can be assessed from the initial geometry and the velocity field before any training is performed. The supporting machinery is the fixed protocol, eight hidden layers of 256 tanh units with Adam followed by L-BFGS, residual-based adaptive sampling, and causal weighting, which makes the across-benchmark comparison meaningful.","core_discovery":"The central claim is that the design principle for eikonal regularisation established in two dimensions transfers to three: the selected weight $w_{\\mathrm{eik}}$ is governed not by the spatial dimension but by how far the exact solution departs from a signed-distance function, and the same four-decade ordering appears in both settings. Concretely, the sweep selects $10^{-1}$ for the translating and rotating spheres, where $\\|\\nabla\\phi\\|=1$ holds exactly for all time; $10^{-3}$ for the slotted sphere, where the property fails locally near the slot's reentrant edges; and $10^{-5}$ for the reversed vortex, where stretching makes the property fail globally. The ordering is preserved while the values are not: the translating-sphere optimum moves a decade from its two-dimensional value and the reversed-vortex optimum shifts from $10^{-4}$ to $10^{-5}$. The paper further claims that the weight remains the dominant hyperparameter in three dimensions, since varying it over four decades changes the relative $L_2$ error by up to a factor of 31 and the volume conservation error by up to a factor of 40, with a non-monotone response and a well-defined interior optimum on every benchmark, and that the regulariser reduces seed-to-seed variance by more than an order of magnitude at small weights. A like-for-like comparison against fifth-order WENO finds the classical solver more accurate on all four benchmarks, by two orders of magnitude on smooth rigid advection, with a margin that narrows monotonically with geometric difficulty and is considerably smaller in volume conservation than in the field norm.","pith_inferences":["Because the signed-distance departure is computable from the problem specification, the ordering suggests a cheap pre-training diagnostic: estimate the departure from the initial geometry and velocity field, pick a starting $w_{\\mathrm{eik}}$ from the four-decade ladder, and run a single verification sweep instead of the full grid.","The narrowing WENO margin hints that the network's relative standing improves exactly where the level-set solution is least smooth; a stretching benchmark with more reversals or thinner filamentation would test whether the trend continues or reverses.","The variance-reduction result implies that even where accuracy does not demand a large eikonal weight, reproducibility does, a consideration that adaptive weighting schemes should be checked against for recovered optima rather than only for final error.","The feature-restricted measure is a natural default report for any physics-informed study claiming thin-structure preservation, including the sparse and noisy interface-data regimes this paper explicitly leaves untested."],"forward_implications":["The eikonal weight must be re-selected per problem: with only two of four benchmarks inheriting their two-dimensional optimum unchanged, assuming transfer is unreliable and verification is required.","The weight sweep is worth its cost: since varying $w_{\\mathrm{eik}}$ over four decades changes the relative $L_2$ error by up to 31 times and the volume conservation error by up to 40 times, the 18-run sweep per benchmark is the price of a trustworthy configuration.","Single-seed experiments should be treated with suspicion in this setting: at small weights the seed-to-seed standard deviation equals the error itself, and one nominal tie between weights reported in the paper is an artefact of that variance rather than a genuine equivalence.","A neural solver is not the efficient choice for prescribed-flow advection on a uniform grid: the matched WENO5 comparison is more accurate everywhere, with the margin falling from a factor of 122 to a factor of 4.3 as geometric difficulty increases.","Global norms cannot certify thin features: on the slotted sphere, substituting a sphere with no slot at all changes the relative $L_2$ error by about one percentage point, so studies of interfaces with fine structure should report the paper's feature-restricted, sign-based classification measures."],"supporting_citations":[{"why":"The two-dimensional predecessor that supplies the architecture, training protocol, and the eikonal-weight ordering whose transfer to three dimensions this paper tests.","marker":"[10]"},{"why":"Supplies the reversed-vortex velocity field used as the stretching benchmark, whose exact solution is constructed by backward characteristic integration.","marker":"[21]"},{"why":"The fifth-order WENO reconstruction that forms the classical solver against which the PINN is benchmarked on identical grids and error measures.","marker":"[17]"},{"why":"The third-order TVD Runge-Kutta time stepping used to advance the WENO comparison scheme.","marker":"[18]"},{"why":"Defines the physics-informed neural network framework whose composite residual loss the eikonal term augments.","marker":"[4]"},{"why":"Documents how convection-dominated problems make PINN training fail even with adequate capacity, motivating the loss-weighting focus of this study.","marker":"[5]"},{"why":"The causal weighting scheme applied to the residual loss on the adaptive-sampling benchmarks (slotted sphere and reversed vortex).","marker":"[20]"}],"fun_headline_variants":["Eikonal weight ordering holds in 3D, values shift","Eikonal tuning: 3D ordering transfers, values don't","Eikonal weight tracks distance property, not dimension","Eikonal regulariser buys reproducibility along with accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inherited eight-layer, 256-unit tanh architecture and the fixed training protocol (Adam schedule, sample counts, adaptive-sampling and causality settings) remain adequate in three dimensions, so that the measured eikonal-weight ordering reflects the regulariser's effect rather than capacity or convergence failure; the paper itself concedes that transfer to substantially different networks is untested.","fun_headline_variants_meta":{"raw":{"variants":["Eikonal weight ordering holds in 3D, values shift","Eikonal tuning: 3D ordering transfers, values don't","Eikonal weight tracks distance property, not dimension","Eikonal regulariser buys reproducibility along with accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3473,"prompt_tokens":1210,"completion_tokens":2263,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":826,"completion_tokens_details":{"reasoning_tokens":2193}},"tokens_in":826,"tokens_out":2263,"duration_ms":17385,"temperature":1.0,"reasoning_tokens":2193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:10:12.627104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same protocol and six-weight sweep on a rigidly transported smooth ellipsoid, whose exact solution is an exact signed distance function at every time, and check whether any weight at or below $10^{-3}$ wins; if so, the claimed ordering fails. Conversely, a stretching benchmark that selects a weight near $10^{-1}$ would falsify it from the other side. The paper itself names exactly this test in its limitations section.","supporting_citations":[{"cited_title":"A Hybrid Parti- cle Level Set Method for Improved Interface Capturing.Journal of Computational Physics, 183(1):83–116, 2002","cited_arxiv_id":null,"evidence_quote":"Supplies the reversed-vortex velocity field used as the stretching benchmark, whose exact solution is constructed by backward characteristic integration."},{"cited_title":"Krishnapriyan, Amir Gholami, Shandian Zhe, Robert M","cited_arxiv_id":null,"evidence_quote":"Documents how convection-dominated problems make PINN training fail even with adequate capacity, motivating the loss-weighting focus of this study."}],"review_version":1}