{"id":"53da5881-d801-4c90-9892-33f45ab8d474","arxiv_id":"2505.12389","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Physics-informed neural networks reproduce Saint-Venant torsion solutions for 2D cross-sections and 1D sharp-transition problems, with reported relative errors as low as 0.1% for simple shapes.","lead":"This paper applies physics-informed neural networks to the Saint-Venant torsion problem, computing torsional constants and twist angles for bars with various cross-sections. It also tests a variable-scaling variant for sharp diameter transitions and a parametric variant that predicts solutions across a parameter range without retraining.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's analytical torsional constants disagree with standard formulas for the stated 0.2 m geometries, so the 0.1% accuracy claim is not yet supported.","rationale":"The most vulnerable link in the paper is the 2D validation in Section 5.2 and Table 1. The headline torsional-constant accuracy rests entirely on the four reference values, but the analytical formulas for those references are never stated. Applying standard Saint-Venant formulas to the geometries as described in the paper gives values that differ from Table 1 by factors of roughly 29, 23, and 49 for the circle, square, and triangle, respectively. The shape-dependence of these factors rules out a simple unit or missing-factor explanation, meaning the benchmark itself appears to be wrong rather than merely unstated. Since the same reference values are used to compute both PINN and ANSYS relative errors, the reported 0.1% and 3.0% accuracy numbers cannot be interpreted. The 1D VS-PINN and parametric PINN experiments are more self-contained and less affected by this issue, but the 2D torsional-constant claim is central to the paper's stated contribution. The reader's weakest assumption identifies exactly this issue, so my stress-test agrees rather than introduces a new concern. The conditional verdict remains appropriate: the paper should be accepted only after the analytical references are corrected and justified, or after an independent reference solution is provided. No adversarial interpretation of the authors' intent is needed; the concern is purely about the verifiability and internal consistency of the quantitative claim.","tokens_in":12267,"tokens_out":9666,"duration_ms":93016,"concrete_test":"Recompute the 'Analytical Solution' column of Table 1 from the stated dimensions and standard torsion formulas: J_circle = πR^4/2 with R = 0.1 m; J_square ≈ 0.1406 a^4 with a = 0.2 m; J_triangle = √3 a^4/80 with a = 0.2 m. If these do not reproduce 4.63202E-03, 5.11785E-03, and 1.71155E-03, request the authors' formula sheet or run a high-resolution FEM solve of Eq. (15) on the same geometries as an independent reference. If the corrected reference values change the relative errors by more than about one percentage point, the paper's 2D validation and headline accuracy claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 in Section 5.2 is the only quantitative support for the paper's central 2D claim: that PINN computes Saint-Venant torsional constants with 0.1–3.0% relative error. The paper never states the formulas behind the 'Analytical Solution' column, and standard formulas for the geometries as described do not reproduce those numbers. For the stated 0.2 m circle, J = πR^4/2 = 1.571E-4 m^4, but Table 1 lists 4.63202E-03 m^4 (29.5× larger). For a 0.2 m square, J ≈ 0.1406 a^4 = 2.25E-4 m^4, versus 5.11785E-03 m^4 (22.7× larger). For a 0.2 m equilateral triangle, J = √3 a^4/80 = 3.46E-5 m^4, versus 1.71155E-03 m^4 (49.4× larger). The mismatch is shape-dependent, so it cannot be a single missing factor or unit conversion. Because these same values define the reported PINN and ANSYS errors, the 0.1%/3.0% accuracy claim is uninterpretable until the analytical references are supplied and reconciled with Eq. (15). No code, data, random-seed statistics, or checkpoint-selection protocol is provided, but the inconsistent benchmark is the decisive problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops and tests three physics-informed neural network (PINN) approaches for Saint-Venant torsion problems: a standard PINN that solves the 2D Poisson equation for the Prandtl stress function on circular, square, triangular, and irregular cross-sections; a variable-scaling PINN (VS-PINN) for a 1D torsion problem with a sharp transition in the torsional stiffness J(x); and a parametric PINN that takes the parameters of a Gaussian body force as additional inputs. The authors report relative errors as low as 0.1% for the 2D circular and square cases, a reduction in relative L2 error from 9.7E-3 to 1.1E-3 for VS-PINN compared with the baseline PINN, and a relative L2 error of 1.07E-2 for the parametric PINN. They conclude that these methods provide accurate, mesh-free alternatives to FEM for torsional analysis.","tokens_in":12612,"tokens_out":9501,"duration_ms":84169,"significance":"If the reported accuracies were properly validated, this would be a useful engineering demonstration that PINN-based solvers can compute torsional constants and twist angles without meshes, and that VS-PINN improves convergence for stiff coefficient transitions. The paper's positive features include the use of analytic and FEM references for the main comparisons, the clear formulation of the 1D and parametric test problems, and the explicit statement that no ground-truth data were used during training. However, the central 2D accuracy claim is currently unsupported because the analytical benchmark values are not stated and appear inconsistent with standard formulas for the stated geometries. In addition, the reference solutions for the 1D and parametric error measurements are not defined, and all results come from single training runs. These issues prevent the reader from verifying the paper's quantitative claims.","major_comments":[{"comment":"The 'Analytical Solution' column is not derived or cited, and its values contradict standard Saint-Venant torsion constants for the geometries described in Section 4.1. For the stated circle of diameter 0.2 m, the standard value is J = πR^4/2 ≈ 1.57×10^-4 m^4, whereas Table 1 lists 4.63202×10^-3 m^4 (about 29.5 times larger); for the square, 0.1406 a^4 ≈ 2.25×10^-4 m^4 versus 5.11785×10^-3 m^4; and for the equilateral triangle, √3 a^4/80 ≈ 3.46×10^-5 m^4 versus 1.71155×10^-3 m^4. The ratios are shape-dependent, so they cannot be explained by a single missing factor or unit conversion. Because the reported relative errors for both PINN and ANSYS are computed with respect to this column, the central claim of 0.1–3.0% accuracy for the 2D solver is unverifiable until the analytical formulas are supplied and reconciled with Eq. (15).","section":"Section 5.2, Table 1"},{"comment":"The reference solutions used to compute the relative L2 errors for the VS-PINN experiment (9.7E-3 vs. 1.1E-3) and for the parametric PINN experiment (1.07E-2) are never defined. For Eq. (19), the coefficient J(x) contains a sigmoid transition, so no closed-form solution is given; for Eq. (21), an exact solution in terms of error functions exists but is not stated. Without a reproducible reference, these error values cannot be checked, which undermines the quantitative claims for the two advanced methods.","section":"Sections 5.3 and 5.4"},{"comment":"All neural-network results are reported for a single training run, with no seed statistics, no variance bars, and no description of how the reported errors were selected from the training history (e.g., 'highest accuracy'). Given that PINN training is sensitive to initialization, collocation-point sampling, and loss weights, the paper's conclusion of 'accuracy and robustness' is not supported by the presented evidence. The authors should report at least the mean and standard deviation over multiple seeds, or explicitly justify why a single run is representative.","section":"Sections 5.2–5.4"}],"minor_comments":[{"comment":"The text says 'λr, λd > 0' but the total loss is defined with λb; this appears to be a typo.","section":"Section 3.3 after Eq. (14)"},{"comment":"The units of the tabulated values (presumably m^4) should be stated in the table or in the caption.","section":"Table 1 caption"},{"comment":"The term 'true solutions' should be defined, and the caption should state how these exact solutions were computed.","section":"Section 5.4, Figure 11"},{"comment":"The problem statement does not explain how the boundary condition ϕ'(1) = 32/(πJ(1)) is derived after the normalization, and the notation for the twist angle is inconsistent with the earlier definition θ = αz in Eq. (2).","section":"Section 4.2, Eq. (19)"},{"comment":"References [KP25a] and [KP25b] are the same paper cited twice; the duplicate entry should be removed.","section":"References"},{"comment":"The phrase 'the highest accuracy achieved was the 9.7E-3 relative L2-error' should be rephrased as 'the lowest error achieved' or 'the best relative L2-error'.","section":"Section 5.3, paragraph 2"},{"comment":"The phrase 'temperature fields' is a typo; the figure shows the Prandtl stress function field.","section":"Section 5.2, Figure 7 description"},{"comment":"The contribution list states 'We propose the VS-PINN formulation' and 'we introduce a parametric PINN framework', but both methods are cited to prior work [KP25a, CJL+24]; the contributions should be rephrased as applications or extensions rather than new method proposals.","section":"Section 1, contributions"}],"recommendation":"major_revision","confidential_remarks":"The Table 1 inconsistency is likely a genuine error rather than a simple omission: the analytical values imply effective geometric sizes that differ across shapes and are inconsistent with the stated 0.2 m dimensions. This should be the first item the authors address. The novelty of the work is mainly in the engineering application, since the VS-PINN and parametric PINN methods are taken from prior literature; the authors should adjust their contribution claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a readable engineering study that takes three known PINN formulations and points them at Saint-Venant torsion. The one genuinely useful piece is the 1D sharp-transition problem, where VS-PINN beats vanilla PINN by roughly an order of magnitude. The 2D torsional constant results, which the abstract sells hardest, are not verifiable as reported.\n\nThe derivation of the governing Poisson equation is standard and mostly correct (minor typo: one γxz has −αx instead of −αy). The 1D experiment is clean: it defines a clear stiff coefficient problem, normalizes load and modulus, and compares N=1 vs N=2,4. The parametric PINN demo is simple but fine—it shows a surrogate over a small parameter box, nothing groundbreaking. None of this is methodologically new; the VS-PINN is the authors' own earlier paper and the parametric PINN follows Cho et al. That's legitimate, but the abstract's 'novel' framing overstates it.\n\nThe problem is Table 1. The analytical column is supposed to anchor the 0.1–3% error claims, but the formulas behind it are never given, and the numbers don't match standard results for the stated geometries. For a 0.2 m diameter circle, J = πR⁴/2 ≈ 1.57e−4 m⁴, but the table lists 4.63e−3. For a 0.2 m square, the usual formula gives ≈ 2.25e−4, versus 5.12e−3. The ratios differ by shape, so it isn't a missing factor or unit error. Without a corrected analytical benchmark, the comparison to PINN and ANSYS is meaningless. Also, the paper never states the shear modulus G used in the 2D loss, so even the setup isn't reproducible. On top of that, the training appears to be a single seed with the best checkpoint reported, and no code or data is provided.\n\nThe 1D and parametric sections are not implicated in the table problem, so they likely survive a rewrite. But as it stands, the paper's central quantitative claim rests on a benchmark that I cannot reproduce.\n\nWho is this for? Practitioners in computational mechanics who want a recipe for applying PINN variants to torsion problems. It's not a methods paper. I'd send it to a competent referee rather than desk-reject, but the referee should ask the authors to reconcile the analytical constants, state G, release data/seeds, and report error bars. If the table is fixed, the paper becomes a solid application note.\n\nRegards.","headline":"A clean application of existing PINN variants to a classical problem, but the 2D validation table rests on analytical constants that don't match textbook formulas, so the headline accuracy claim is not yet supported.","tokens_in":13124,"tokens_out":7416,"would_cite":false,"duration_ms":64076,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Physics-informed neural networks can compute Saint-Venant torsion constants and twist angles without meshes, with errors as low as 0.1 percent on simple cross-sections.","keywords":["Saint-Venant torsion","physics-informed neural networks","Prandtl stress function","torsional constant","variable-scaling PINN","parametric PINN","mesh-free PDE solvers","Poisson equation"],"falsifier":"Recompute the torsional constant for a 0.2 m diameter circular cross-section with the classical formula $J = \\pi R^4/2$, which gives about $1.57\\times 10^{-4}$ m$^4$ for $R = 0.1$ m. The Table 1 analytical value is $4.63202\\times 10^{-3}$; if the true analytical value is the former, then the claimed 0.1% error is measured against an unexplained reference, and the same check applied to the square and triangle entries would settle whether the benchmark is consistent.","tokens_in":12079,"feed_emoji":"🌀","tokens_out":9797,"duration_ms":87454,"temperature":0.7,"pith_summary":"This paper sets out to show that physics-informed neural networks (PINNs) can serve as practical solvers for Saint-Venant torsion problems, replacing the mesh generation and numerical integration of conventional finite-element workflows. It builds three solver variants: a standard PINN for the 2D Poisson equation of the Prandtl stress function, a variable-scaling PINN (VS-PINN) for shafts with sharp diameter transitions, and a parametric PINN that takes load parameters as extra inputs for real-time prediction. The reported results put the circle and square torsional constants within 0.1 percent of the paper's analytical references, the triangle within 3.0 percent, and the parametric surrogate at a $1.07\\times 10^{-2}$ relative $L^2$ error over a range of forcing parameters. If those results hold, the practical payoff is a mesh-free path to torsional constants and twist angles that can be evaluated instantly for many design configurations.","feed_headline":"PINNs hit 0.1% error on torsion constants, no mesh needed","feed_subtitle":"Torsional-constant errors reach 0.1%, and a variable-scaling variant tames sharp shaft transitions.","key_machinery":"The central object is the scaled Prandtl stress function $\\phi'$, which satisfies the Poisson equation $\\nabla^2 \\phi' = -2G$ with a constant (taken as zero) boundary value; the torsional constant is its integral, so the whole method rests on representing one smooth scalar field with a network and automatic differentiation. The VS-PINN mechanism is the spatial rescaling $x \\mapsto x/N$, which maps a sharply varying solution profile onto a smoother one over an enlarged domain, making the residual easier for the optimizer to fit. The parametric PINN mechanism is to concatenate the parameters $(T,m,\\sigma)$ with the spatial coordinate as network inputs and average the physics loss over the parameter space, so a single trained network maps any parameter set in range to its solution.","core_discovery":"The central claim is that the torsional constant $J$ of a prismatic bar can be obtained by training a neural network to solve $\\nabla^2 \\phi' = -2G$ for the scaled Prandtl stress function $\\phi'$ with a zero boundary condition, then evaluating $J = \\frac{2}{G}\\int_A \\phi'\\,dA$; all three solvers are exercises in learning this single scalar field. For the 1D shaft with a sharp transition, the paper claims that rescaling the spatial coordinate by a factor $N$ before forming the loss function lowers the relative $L^2$ error from $9.7\\times 10^{-3}$ to $1.1\\times 10^{-3}$ at the same epoch count and no extra per-epoch cost. For parametric studies, treating $(T,m,\\sigma)$ as extra network inputs yields a surrogate with a $1.07\\times 10^{-2}$ relative $L^2$ error that predicts twist solutions for unseen parameter combinations without retraining.","pith_inferences":["The parametric PINN could be coupled with an optimizer to solve inverse torsion design problems, such as finding a torque profile that produces a target twist distribution.","Feeding material properties or cross-section shape descriptors as additional parameter inputs would extend the same surrogate idea from one geometry to a family of geometries.","The variable-scaling idea is demonstrated only on the 1D shaft equation; applying it to 2D torsion with sharp re-entrant corners is a testable extension that should show the same gradient-smoothing benefit."],"forward_implications":["A trained PINN outputs $\\phi'$ continuously on the whole cross-section, so torsional constants for arbitrary shapes can be evaluated by sampling the network instead of building a mesh and assembling stiffness matrices.","The VS-PINN result makes stepped shafts and other stiff-transition geometries solvable at a $1.1\\times 10^{-3}$ relative $L^2$ error with the same per-epoch cost as a standard PINN, removing the main accuracy bottleneck for such cases.","The parametric PINN turns a torsion solver into a surrogate that can be queried in real time for any torque parameters in the training range, which is what design optimization and digital-twin workflows require.","Because the paper fixes the PINN collocation points to match the finite-element grid, the two approaches can be compared resolution-for-resolution rather than accuracy-for-accuracy.","The three variants are presented as complementary tools: the standard PINN for accurate constants, VS-PINN for stiff sections, and the parametric PINN for parameter sweeps."],"supporting_citations":[{"why":"Supplies the physics-informed loss construction, residual plus boundary terms, that all three solvers minimize.","marker":"[RPK19a]"},{"why":"Provides the variable-scaling transformation that is the core of the VS-PINN accuracy improvement on stiff transitions.","marker":"[KP25b]"},{"why":"Introduces the parameter-as-input formulation used for the parametric PINN surrogate.","marker":"[CJL+24]"},{"why":"Derives the Prandtl stress-function and Poisson-equation formulation on which the torsional-constant integral rests.","marker":"[EB10]"},{"why":"Supplies the finite-element method for Saint-Venant torsion that serves as the conventional numerical baseline for comparison.","marker":"[JM14]"}],"fun_headline_variants":["Neural solvers achieve 0.1% error on torsion constants without meshing","Variable-scaling PINN sharpens torsion accuracy for abrupt geometry changes","Parametric PINN one-shot predicts twists for unseen shaft shapes and loads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analytical benchmark values in Table 1 are never derived or cited, and the listed circle value does not match the classical torsion-constant formula for the stated 0.2 m geometry, so the reported error percentages rest on an unverified reference solution.","fun_headline_variants_meta":{"raw":{"variants":["Neural solvers achieve 0.1% error on torsion constants without meshing","Variable-scaling PINN sharpens torsion accuracy for abrupt geometry changes","Parametric PINN one-shot predicts twists for unseen shaft shapes and loads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1317,"prompt_tokens":995,"completion_tokens":322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":258}},"tokens_in":611,"tokens_out":322,"duration_ms":3929,"temperature":1.0,"reasoning_tokens":258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:36:18.087400+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the torsional constant for a 0.2 m diameter circular cross-section with the classical formula $J = \\pi R^4/2$, which gives about $1.57\\times 10^{-4}$ m$^4$ for $R = 0.1$ m. The Table 1 analytical value is $4.63202\\times 10^{-3}$; if the true analytical value is the former, then the claimed 0.1% error is measured against an unexplained reference, and the same check applied to the square and triangle entries would settle whether the benchmark is consistent.","supporting_citations":[],"review_version":1}