{"id":"6f5b846c-d56e-4b66-a4ad-49d725eb105a","arxiv_id":"2505.20361","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Sequentially separating and retraining around the strongest discontinuity, then stitching subdomains, lets physics-informed neural networks resolve multiple shocks more accurately.","lead":"This paper introduces a training strategy called ST-PINNs that solves fluid equations with multiple shock waves by handling one strong discontinuity at a time. It reports large error reductions on one- and two-dimensional shock-interface problems compared to existing physics-informed neural network methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'always improve' claim rests on unvalidated extraction of shock trajectories and boundary states; all tests use known linear fits and averaged front values, so the method's generality is not yet established.","rationale":"The reader's weakest-assumption analysis already identifies the same load-bearing point: the method's improvement is contingent on reliable extraction of the primary discontinuity and accurate boundary values from the base model. I agree that this is the place where the central claim is least secure. The paper's evidence is strong within its tested cases: error reductions are consistent across multiple base models and hyperparameter settings, and the 2D shock-refraction test is a genuine extension. However, all tested discontinuities are straight, planar fronts whose trajectories are fitted to known linear forms, and the boundary conditions are single averaged states. This means the numerical experiments do not exercise the step where the algorithm must infer a general, possibly curved, discontinuity trajectory and use locally varying boundary data. Appendix B varies only base-model weights; it does not vary the extraction or boundary-condition inputs, so it cannot support the word 'always' in Section V. A second, related gap is the absence of an ablation separating the contribution of transfer learning from plain subdomain retraining. That gap affects attribution of the mechanism but is secondary to the boundary-data dependence. The proposed perturbation test would directly settle whether the claimed improvement survives realistic extraction errors. Because the concern is addressable and does not invalidate the reported results, the existing CONDITIONAL verdict should stand.","tokens_in":15456,"tokens_out":4390,"duration_ms":59434,"concrete_test":"Re-run the Section IVA one-dimensional test with Algorithm 1 unchanged except for deliberately perturbed extracted inputs: shift the fitted shock trajectory x_shock(t) by ±0.01 in x and scale the averaged boundary states (rho, u, p) by 0.95 and 1.05, keeping all other hyperparameters fixed. If the final L2 errors remain at or below the baseline PINNs-WE errors reported in Table II, then the claim is robust to extraction error; if the errors degrade noticeably, the 'always improve' assertion fails unless accurate trajectory and boundary extraction is guaranteed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section V—that once a base model captures one discontinuity with sufficient precision, ST-PINNs 'can always improve'—depends on two quantities extracted from the base model: the fitted primary-discontinuity trajectory and the adjacent states VNN,± used as subdomain boundary conditions (Algorithm 1, lines 4–8). In every numerical test these quantities are supplied in a highly favorable form: the fitted trajectories are straight lines with known planar structure (x_shock = 1.606 t + 0.5, x1 = 2.228 t + 0.20, x2 = 2.232 t + 0.70, x_contact = 1.05 t + 0.2), and the boundary values are single averaged states (e.g., rho = 1.590, u = 0.603, p = 1.945 in Section IVA). Section IVC additionally uses the incident-shock front and its intersection point as 'known prior information.' If the extracted position is biased or the boundary state is inaccurate, every downstream subdomain inherits that error, and the improvement over the baseline is not guaranteed. Appendix B sweeps base-model hyperparameters but never perturbs the fitted trajectory or the boundary values, so it cannot establish the claimed 'always.' The method is internally coherent, but its general applicability to curved or nonlinear discontinuities—where the straight-line fits used here are invalid—is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ST-PINNs, a staged training procedure for PINNs solving Euler equations with multiple discontinuities. The method first trains a base PINN variant over the full domain until the strongest discontinuity is captured, extracts the discontinuity locus and adjacent state values from that model, divides the domain along the locus, and retrains the subdomain containing weaker discontinuities via transfer learning, using the base model's near-front predictions as boundary conditions. Sequential subdivision is repeated until all discontinuities are resolved, and the subdomain networks are combined. Validation is reported for a 1D shock-interface Riemann problem, a quasi-1D planar shock-interface interaction, and a 2D unsteady planar shock refraction problem, using PINNs-WE, GA-PINNs, simplified variants, and their ST-PINNs counterparts. Across these cases the reported L2 relative errors are consistently lower after ST-PINNs refinement than for the base models.","tokens_in":15711,"tokens_out":4965,"duration_ms":52561,"significance":"If the central claim holds, ST-PINNs offers a conceptually simple and potentially general strategy for turning any PINN variant that can capture one discontinuity into a solver for multi-discontinuity problems. The paper's main strengths are the consistency of the reported improvements across six base-model configurations (Tables II, IV, VI, VII, and VIII), the use of PDE and initial-condition information only during training, and the first reported PINN treatment of the 2D unsteady planar shock refraction problem. The method is algorithmic and does not rely on synthetic training data from classical solvers during ST-PINNs training. However, the strength of the evidence is tempered by the incomplete definition of the error metric, the absence of repeated-run statistics, and the fact that the discontinuity extraction and boundary-value construction are tested only in favorable straight-front configurations.","major_comments":[{"comment":"The L2 relative error is written as sqrt((fpred - ftrue)^2 / ftrue^2), with no summation over sample points and no normalization over the field. As written this defines an error for a single point, not an L2 error, and it cannot reproduce the reported percentages. Since every quantitative claim in Tables II, IV, VI, VII, and VIII rests on this metric, the definition must be corrected (e.g., sum over sample points and normalize by the sum of ftrue^2) and the concrete evaluation formula should be stated explicitly.","section":"§IV, Eq. (14)"},{"comment":"The conclusion that ST-PINNs 'can always improve' is not supported by the reported experiments because the two inputs on which the method depends—the fitted discontinuity locus and the boundary values VNN,±—are supplied in favorable form in every test. The loci are straight-line fits with known planar structure (x_shock = 1.606t + 0.5 in §IVA; x1 = 2.228t + 0.20 and x2 = 2.232t + 0.70 in §IVB; x_shock = 2.370t + 0.2 and x_contact = 1.05t + 0.2 in §IVC), and the boundary values are single averaged states. In §IVC the incident-shock front and its intersection point are additionally used as known prior information. Appendix B perturbs only base-model hyperparameters, never the extracted trajectory or the base-model boundary values, so error propagation from an imperfect extraction is not characterized. The claim should either be restricted to the tested straight-front cases or supported by experiments with perturbed or biased front locations and with non-planar fronts.","section":"§V; Algorithm 1, lines 4–8"},{"comment":"All numerical results are single-run values. Tables II, IV, VI, VII, and VIII report one error per configuration, with no seed information, no standard deviation, and no repeated training trials. Because neural-network training is stochastic and the method sequentially reuses a fitted trajectory and boundary values, the claimed consistency of improvement requires at least a few independent runs per configuration to rule out that the reported gains are within run-to-run variability.","section":"§IV, Tables II/IV/VI and Appendix B"}],"minor_comments":[{"comment":"The name of the method alternates between 'Separation-Transfer' in the title and abstract and 'Separate-Transfer' in Section I; this should be made consistent.","section":"§I and title"},{"comment":"The label 'ST-PINNS' appears with an extra 'S' in Tables IV and VI, while the text uses 'ST-PINNs'; please correct the typo.","section":"Tables IV and VI"},{"comment":"The statement that 'the same network architecture is employed across all stages, with only the loss function weights adjusted accordingly' is contradicted by Tables I, III, and V, where sampling-point counts, learning-rate schedules, and initial-condition weights also change between models; the sentence should be clarified to refer to architecture alone.","section":"§III, end of Methodology"},{"comment":"The entries 'Boundary Points NAN' and 'ωBC NAN' should be replaced by a conventional notation such as 'not used' or '—', since 'NAN' is a numeric sentinel with a different meaning.","section":"Table I"},{"comment":"The notation |·| is used both for element-wise absolute value and for the mean of element-wise absolute values without a clear distinction; this makes Eq. (9) ambiguous and should be clarified.","section":"§IIB, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is coherent and the reported improvements are consistent, but the load-bearing 'always improve' claim in Section V needs to be either narrowed or backed by perturbation and repeated-run experiments. I would ask for those experiments before final acceptance, since the current evidence cannot distinguish a generally robust algorithmic strategy from a method that works well when the discontinuity extraction is already nearly perfect."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and worth taking seriously: let the base PINN capture the strongest shock, split the domain at that front, use the base model's nearby predictions as boundary conditions, and transfer-learn the weaker discontinuities one at a time. That strong-to-weak sequential decomposition paired with transfer learning is not in the cited CPINN/XPINN/PINNs-WE/GA-PINN line, and the 2D unsteady planar shock refraction test is a genuinely new PINN application. I also credit them for training only against the PDE and initial conditions, no simulation data in the loss.\n\nThe evidence is better than the abstract suggests. Error reductions are consistent across six base-model/hyperparameter combinations (Appendix B, Tables VII and VIII) and all three problem classes, typically cutting errors by half or more. That consistency gives the mechanism credibility, and the hyperparameter sweep shows the improvement is not tied to one tuning.\n\nThe soft spots are real but mostly fixable. Eq. (14) is not an L2 error as written; it is a pointwise relative error with no summation or normalization. All results are single runs, so we have no idea about seed variance. There is no ablation separating transfer learning from simply retraining each subdomain from scratch, and no comparison against CPINNs/XPINNs as domain-decomposition baselines. The bigger limitation, which the stress-test note correctly identifies, is that every fitted trajectory is a straight line with known planar structure, and the subdomain boundary states are single averaged values. That means the method's generality to curved or folding discontinuities is untested. So the Section V claim that ST-PINNs 'can always improve' once the base captures one discontinuity is stronger than the evidence supports; it should be scoped to the tested cases.\n\nThis paper is for researchers actively fighting PINN gradient pathology on compressible flow problems. It deserves a serious referee, but the referee should push for the ablation, multi-seed runs, a corrected error metric, and a humility pass on the generality claims. As is, I would send it to peer review expecting major revision, not a desk reject.","headline":"A plausible sequential domain-decomposition trick that consistently helps PINNs on multiple discontinuities, but the 'always improves' claim outruns the straight-line test cases.","tokens_in":16300,"tokens_out":2141,"would_cite":true,"duration_ms":23133,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A separation-transfer training loop lets physics-informed neural networks resolve multiple shock discontinuities sequentially, starting from the strongest.","keywords":["physics-informed neural networks","Euler equations","multiple discontinuities","shock-interface interaction","shock refraction","domain decomposition","transfer learning","gradient pathology"],"falsifier":"Run the same three experiments with the fitted shock trajectories deliberately offset (for example, use $x_{\\text{shock}} = 1.606\\, t + 0.5 \\pm 0.05$ in the one-dimensional case) and check the final L2 errors; if the reported improvements over the base model disappear for any offset, the claim that ST-PINNs always improves once the primary front is captured is false.","tokens_in":15203,"feed_emoji":"💥","tokens_out":7992,"duration_ms":77606,"temperature":0.7,"pith_summary":"Physics-informed neural networks (PINNs) solve PDEs by turning residual, initial, and boundary conditions into a loss, but when several shock fronts coexist, the strongest front dominates the training and the weaker fronts stay unresolved. ST-PINNs breaks this deadlock by resolving the fronts one at a time: train any PINN variant until the strongest discontinuity is captured, locate that front, split the domain along it, and transfer-learn on the subdomain containing the remaining discontinuities with predicted values at the cut as boundary conditions. The paper shows the cycle captures every front in order of decreasing strength, and reports improved L2 accuracy over the PINNs-WE and GA-PINNs base models on three Euler problems, including the two-dimensional unsteady planar shock refraction problem.","feed_headline":"Split-and-transfer PINNs capture multiple shock fronts one by one","feed_subtitle":"Starting from the strongest jump, each retrained subdomain resolves the next discontinuity until all fronts are captured.","key_machinery":"The load-bearing object is the discontinuity intensity $I = |V_+ - V_-|/\\Delta x$, computed from a chosen physical quantity on a discretized grid, which ranks the jumps and identifies the primary discontinuity as the point of maximum $I$. Around that front, the method fits a trajectory (in all experiments a straight line in $x$--$t$), reads off averaged values of the base model just ahead of the front, and divides the domain into subdomains. Transfer learning, initialized from the previous model, retrains only the subdomain that still contains poorly resolved fronts, using the near-front values as boundary conditions; with the primary front removed, the next-strongest front becomes dominant in the loss and can be learned.","core_discovery":"The paper's thesis is that a multi-discontinuity Euler problem becomes tractable when the discontinuities are solved sequentially from strongest to weakest. Given any PINN variant that can capture at least one discontinuity with reasonable precision, ST-PINNs determines the primary front through a discontinuity intensity measure, fits its trajectory, cuts the domain there, and uses the base model's near-front predictions as boundary conditions for a transfer-learned subdomain; the next unresolved front then becomes the primary one and is captured in the following iteration. The authors claim this loop always improves the base prediction once the first front is well captured, and they validate it on one-dimensional shock-interface interaction, quasi-one-dimensional planar shock-interface interaction, and two-dimensional unsteady planar shock refraction, the last being the first PINN-based treatment of that problem.","pith_inferences":["A natural stress test is to apply the same loop to a curved or kinked shock front, since the paper's fitted trajectories are straight lines; a curved front would require a local front-fitting step or a piecewise fit, and the method's success would depend on that step.","The separation logic transfers beyond Euler equations: any conservation law with traveling discontinuities, such as shallow-water bores or elastoplastic waves, could in principle be solved by the same strong-to-weak split-and-transfer recipe.","Because each transfer-learning stage reuses the previous model, training cost should scale roughly with the number of discontinuities; the paper does not report wall-clock timings, so the practical cost trade-off against a single high-resolution network remains open.","The improvement depends on the base model's near-front values being accurate; pairing ST-PINNs with an adaptive sampling scheme at the cut could reduce the inherited error and make the method more robust."],"forward_implications":["Any PINN variant that captures one strong discontinuity can be wrapped in the separation-transfer loop to handle multiple discontinuities without changing its architecture.","The multi-front problem is reduced to a sequence of single-front subproblems, so the gradient-pathology failure caused by one dominant front is bypassed.","The training pipeline needs no simulation or experimental data: the experiments use only the PDE residual and initial conditions, with the reference solution used solely for measuring error.","The approach extends, in principle, to any number of discontinuities, since each iteration promotes the next-strongest unresolved front to primary."],"supporting_citations":[{"why":"Introduces the physics-informed neural network formulation (residual, initial, and boundary loss terms) that ST-PINNs builds on.","marker":"1"},{"why":"Analyzes gradient flow pathologies in PINNs and provides the adaptive loss-weighting scheme used to balance loss terms.","marker":"17"},{"why":"Supplies the PINNs-WE discontinuity-weighting method used as the base model in the one-dimensional and quasi-one-dimensional experiments.","marker":"23"},{"why":"Supplies the GA-PINNs gradient-annihilated weighting method used as the complementary base model in all three experiments.","marker":"24"},{"why":"Provides the high-resolution finite-volume reference solutions against which every reported L2 error is measured.","marker":"26"}],"fun_headline_variants":["ST-PINN: capture shock fronts from strongest to weakest","Sequential PINNs tame multiple discontinuities in Euler flows","Transfer learning helps PINNs resolve shock-interface interactions","One-by-one shock capture: ST-PINN splits and transfers","From strong to weak: ST-PINN tackles multi-front Euler problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the strongest jump's location can be reliably read off the first model's output and fitted as a simple line, and that the model's values near that jump are accurate enough to serve as boundary conditions for the retrained subdomains.","fun_headline_variants_meta":{"raw":{"variants":["ST-PINN: capture shock fronts from strongest to weakest","Sequential PINNs tame multiple discontinuities in Euler flows","Transfer learning helps PINNs resolve shock-interface interactions","One-by-one shock capture: ST-PINN splits and transfers","From strong to weak: ST-PINN tackles multi-front Euler problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2556,"prompt_tokens":837,"completion_tokens":1719,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":1633}},"tokens_in":453,"tokens_out":1719,"duration_ms":103908,"temperature":1.0,"reasoning_tokens":1633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:08:33.183284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three experiments with the fitted shock trajectories deliberately offset (for example, use $x_{\\text{shock}} = 1.606\\, t + 0.5 \\pm 0.05$ in the one-dimensional case) and check the final L2 errors; if the reported improvements over the base model disappear for any offset, the claim that ST-PINNs always improves once the primary front is captured is false.","supporting_citations":[],"review_version":1}