{"id":"4b4d843c-3a50-463a-9533-bf159535953a","arxiv_id":"2506.14036","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A three-network PINN with decoupled strain prediction and boundary-force calibration recovers absolute-scale heterogeneous Young's modulus and Poisson's ratio from noisy synthetic displacement data.","lead":"This paper introduces IE-PINN, a three-network physics-informed neural network that estimates maps of Young's modulus and Poisson's ratio from noisy displacement measurements. The method adds a dedicated strain network and a two-phase calibration using the known applied force to recover absolute-scale elasticity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Absolute-scale calibration rests on an unconstrained boundary stress prediction; the loaded-edge stress is outside the support of the physics losses, so boundary extrapolation error directly scales the recovered modulus field.","rationale":"The reader's weakest assumption identifies the calibration's dependence on the known applied force F and the proportionality between predicted boundary stress and F. My concern sharpens the same assumption: the predicted boundary stress is not merely uncertain in practice; it is structurally unconstrained during training, because all physics and strain losses are computed on interior grids (Equations 14, 15, 17) that exclude the loaded edge. The calibration therefore relies on an extrapolation of the strain and elasticity networks. This is a load-bearing condition for the central absolute-scale claim, and it is not independently verified anywhere in the paper. It is not a reason to reject the method, since the synthetic results suggest the extrapolation works for the tested smooth fields, but it reinforces the conditional verdict: the absolute-scale claim is demonstrated only under the favorable condition that the boundary stress extrapolation is accurate. The reader's verdict of CONDITIONAL remains appropriate, so no adjustment is needed. The concrete test would directly quantify whether this extrapolation error is small enough to support the claim or large enough to require additional boundary regularization or calibration terms.","tokens_in":20254,"tokens_out":7091,"duration_ms":80524,"concrete_test":"Retrain IE-PINN on the dragon/dog dataset at SNR 1000 and compare the Phase-1 predicted boundary stress σ̂_xx(x_b, y_i) against the true FEM stress at the loaded edge. Compute the relative error of the integrated boundary force B = Σ_i σ̂_xx(x_b, y_i) h versus the true F used in Eq. (26). If |B/F − 1| is comparable to or exceeds the reported relative MAE of absolute E (MAE divided by mean E), then the calibration scale is biased by an uncontrolled boundary extrapolation. As a complementary check, perturb F by ±10% and ±1% in Eq. (26) and measure the resulting MAE of the absolute E field; the response should be a clean linear scaling (slope 1.0 in relative terms) if the calibration is simply a scale factor, and any deviation indicates that boundary stress bias is contaminating the scale.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim asserts accurate absolute-scale Young's modulus recovery from noisy displacement. That claim hinges on the Phase 2 calibration in Eq. (26), which multiplies the relative modulus field by c = F / Σ σ̂_xx(x_b, y_i) h. The numerator F is supplied as known, but the denominator is the Phase-1 predicted stress at the loaded edge, computed from the strain and elasticity networks evaluated at boundary coordinates. Critically, no training loss constrains the strain network or the equilibrium residual on that boundary: L_ε (Eq. 14) is evaluated on the (N_x−1)×(N_y−1) interior, L_r (Eq. 15) on the (N_x−3)×(N_y−3) interior, and L_E (Eq. 17) on the interior. The boundary stress is therefore an extrapolation outside the region where the physics losses are enforced, and any systematic error in it enters c linearly, scaling the entire absolute E field via Eq. (27). The reported low MAE of absolute E implies the extrapolation happened to be accurate on the tested dragon/dog dataset, but the method provides no mechanism guaranteeing it, especially at SNR 100 or for less smooth boundary strain fields. Moreover, the synthetic F is computed from the true model (Supplementary Note S2), so the calibration is exact by construction only if the network's boundary stress matches truth; the paper does not report the boundary stress error or a sensitivity analysis with perturbed F. This is a concrete gap between the demonstrated synthetic performance and the claimed practical robustness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes IE-PINN, a physics-informed neural network framework for estimating spatially heterogeneous Young's modulus and Poisson's ratio from noisy displacement data. The method uses three separate neural networks for displacement, strain, and elasticity, and trains them by minimizing a weighted sum of displacement fitting, strain discrepancy, equilibrium residual, and a mean-modulus constraint. A second phase calibrates the absolute scale of Young's modulus by comparing the predicted boundary stress to the applied force. Experiments on synthetic 2D plane-stress datasets at SNR 1000, 500, and 100 show lower error than Elastnet and than an ablation without the strain network; additional ablations support the choices of sinusoidal activation, positional encoding, and pretraining.","tokens_in":20570,"tokens_out":7844,"duration_ms":75407,"significance":"If the results are reproducible, the proposed architecture addresses a real limitation of existing inverse-elasticity PINNs: sensitivity to noise and reliance on a known mean Young's modulus. The two-phase calibration idea is simple and potentially practical when the applied load is known. The paper contains useful ablations (strain network, activation function, positional encoding, pretraining) and reports results on a public-style synthetic benchmark. However, the validation is narrow: one 2D synthetic problem family, a single baseline method, and the critical calibration step is tested with a force value derived from the ground-truth model, so the absolute-scale claim is not yet convincingly demonstrated in realistic conditions. The significance is moderate; the paper is likely to interest the PINN and elastography communities if the calibration is properly stress-tested.","major_comments":[{"comment":"The absolute-scale calibration is not validated independently: the applied force F in the synthetic benchmark is computed from the true Young's modulus and boundary strain (Supplementary Note S2), so the calibration is exactly consistent with the ground truth by construction. The paper presents F as 'experimentally measured' but no experiment with an independently measured force or perturbed F is reported. Since Eq. (27) makes the entire recovered modulus field proportional to F, the authors should report a sensitivity analysis with perturbed F and, if possible, test on data where F is measured separately from the displacement data.","section":"Section 4.3, Eq. (26), Supplementary Note S2"},{"comment":"The boundary stress used in the calibration equation (26) is evaluated at x = x_b, but the strain discrepancy loss (Eq. 14) is evaluated on the (Nx-1)x(Ny-1) interior, the equilibrium residual (Eq. 15) on the (Nx-3)x(Ny-3) interior, and the mean-modulus loss (Eq. 17) on the interior. Therefore the boundary stress is an extrapolation outside the support of all physics losses, and any systematic error in that extrapolation enters the scaling factor c_hat in Eq. (26) linearly, scaling the entire field in Eq. (27). The paper should quantify the boundary stress error (e.g., report relative error of sigma_xx at the loaded edge) or incorporate boundary-adjacent constraints so the calibration is not solely reliant on an unconstrained extrapolation.","section":"Section 4.2, Eqs. (14),(15),(17)"},{"comment":"The central robustness claim is supported only by a single synthetic 2D plane-stress setup with Gaussian noise; no error bars or repeated-seed statistics are reported in Figures 6 and 7, and the comparison is limited to one baseline method (Elastnet). To substantiate the claims of robustness and state-of-the-art performance, the authors should report mean and variance over multiple independent runs (different noise realizations and network initializations) and compare with at least one additional inverse-elasticity method, particularly one that incorporates a learned denoising step.","section":"Sections 2.3 and Figures 6-7"}],"minor_comments":[{"comment":"The text states that training was done 'with pretraining described in Section 10,' but the manuscript has no Section 10; the reference should be to the appropriate part of the Experimental Section.","section":"Section 2.6"},{"comment":"References [70] and [73] are identical (Sitzmann et al. 2020); one should be removed or replaced with a distinct relevant reference.","section":"References"},{"comment":"The finite-difference kernels in Eqs. (10)-(11) appear to implement central differences with a factor of 2 that is not accounted for by the division by h_t; the authors should clarify the exact convolution convention, including the role of the factor 1/2 if the kernels are meant to be averaged over adjacent rows or columns.","section":"Section 4.1, Eq. (9)"},{"comment":"The figure reports 'MRE across 50 independent datasets' but does not specify whether the plotted quantity is a mean or median, and no measure of dispersion is shown; please provide this information and add error bars or box plots.","section":"Figure 4"},{"comment":"The table reports MAE values without standard deviations, and the text does not state how many random initializations and noise realizations were used; please provide this information for reproducibility.","section":"Table S1"},{"comment":"The mean-modulus loss L_E is defined as the sum of absolute deviations of each predicted E(i,j) from E_c, which is not exactly a constraint on the spatial mean; the text says it constrains the mean, so please clarify the relationship (e.g., by noting that it is a softened mean constraint or changing the loss to penalize the difference of means).","section":"Section 4.2, Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the absolute-scale calibration: the synthetic benchmark uses F derived from ground truth, and the boundary stress is an unconstrained extrapolation. If the authors supply a sensitivity analysis, boundary stress error quantification, and repeated-run statistics, the paper could become acceptable. The current evidence does not yet support the strong practical claim of robust absolute-scale estimation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nIE-PINN is a solid, well-ablated incremental advance: the decoupled strain network is a genuinely good idea for dampening noise amplification, and the paper shows each component earns its keep. What it does not fully prove is the headline claim of robust absolute-scale calibration, because the calibration constant is a ratio of a known force to a boundary stress that is extrapolated outside the physics-loss support. That is the part to look at carefully.\n\nWhat's new: the three-network split (displacement, strain, elasticity) with a strain-discrepancy loss and a sequential pretraining protocol, plus the two-phase calibration of Young's modulus scale from the resultant force. The ablations are honest: removing the strain network hurts accuracy, especially at SNR 100; positional encoding and SIREN help; pretraining gives a big reduction in error. The comparison with Elastnet makes the point clearly — direct finite differencing of noisy displacement is brittle.\n\nWhere it's soft: the benchmark is one 2D plane-stress FEM problem with Gaussian noise, one baseline, and no error bars or code. More importantly, the Phase-2 calibration divides the known force F by the sum of predicted boundary stress σ_xx at x_b. All the physics losses — strain discrepancy, equilibrium residual, mean modulus constraint — are evaluated on interior points, so the boundary stress is an extrapolation. Any systematic error there enters the absolute scale linearly and scales the whole field. The synthetic F is computed from the true model (Supplementary Note S2), so the benchmark cannot expose a biased boundary stress. The paper should report boundary stress error and a sensitivity analysis with perturbed F, and ideally test on a case with a different boundary condition or a rough boundary strain field.\n\nThese are addressable gaps, not fatal flaws. The central claim is plausible for smooth fields and modest noise. For a subfield that struggles with noise, the decoupled strain network is a real contribution. I'd send this to peer review; it needs a revision that tightens the calibration analysis and broadens the evaluation, but the core idea deserves referee time.","headline":"Solid, well-ablated incremental advance in PINN-based inverse elasticity; the decoupled strain network is genuinely useful, but the absolute-scale calibration claim is shakier than the paper admits because the calibration factor inherits any boundary extrapolation error and the synthetic force is derived from the true model.","tokens_in":21092,"tokens_out":2298,"would_cite":true,"duration_ms":23005,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65M32","74B05"],"pacs":[],"model":"deepseek-v4-flash","headline":"IE-PINN recovers absolute-scale heterogeneous elasticity maps from noisy displacement data, where direct finite-difference inversion collapses.","keywords":["inverse elasticity","physics-informed neural networks","elastography","Young's modulus estimation","Poisson's ratio estimation","noise robustness","absolute scale calibration","heterogeneous materials"],"falsifier":"A decisive test would be to run the method on a phantom with a known stiffness inclusion while deliberately misspecifying the applied force by 10 percent; if the recovered absolute modulus does not shift by roughly 10 percent, the calibration equation fails, and if the method cannot handle a different heterogeneity pattern at SNR 100, the robustness claim is narrower than stated.","tokens_in":20053,"feed_emoji":"📐","tokens_out":8795,"duration_ms":83454,"temperature":0.7,"pith_summary":"The paper is trying to establish that inverse elasticity—recovering the spatial distribution of material stiffness from measured deformation—can be made stable and noise-tolerant, despite being classically ill-posed. Its proposed IE-PINN separates the job across three neural networks: one fits the noisy displacement field, one predicts strain directly, and one predicts the elasticity parameters. That separation keeps noise from being amplified through repeated differentiation, and a two-phase procedure recovers the absolute Young's modulus scale by matching the predicted boundary stress to the known applied loading force. The claim is supported on synthetic plane-stress data with Gaussian noise at signal-to-noise ratios from 1000 down to 100, where the method reports low errors while a direct finite-difference baseline fails. If the claim holds, elastography-style measurements could produce absolute stiffness maps without priors on mean modulus or internal stress.","feed_headline":"Three-network PINN recovers elasticity maps from noisy data","feed_subtitle":"Applied force alone sets the stiffness scale, replacing the usual assumption of a known mean modulus.","key_machinery":"The mechanism that carries the argument is the three-network decomposition. The displacement network is fitted to the noisy observations and acts as a smoother; the strain network predicts strain directly and is penalized for disagreeing with strain computed from the displacement network, so the PDE residual is built from strain-network outputs rather than from doubly differentiated noisy displacements; the elasticity network outputs $\\hat{E}$ and $\\hat{\\nu}$. The elasticity prediction is constrained only to an arbitrary mean during training, which keeps the phase-one problem well-conditioned. The second load-bearing mechanism is the calibration step: the traction boundary condition gives $F = \\int \\hat{c}\\,\\hat{\\sigma}_{xx}^{(b)}\\,\\mathrm{d}y$, so the scalar $\\hat{c}$ is fixed by numerical integration of predicted boundary stress, and the final modulus is $\\hat{c}\\hat{E}(x,y)$. All coordinates enter through positional encoding, all hidden layers use sine activations, and the three networks are pretrained sequentially.","core_discovery":"The central discovery is that noise sensitivity in inverse elasticity is not an unavoidable feature of the problem, but a consequence of how derivatives are taken. IE-PINN therefore replaces second derivatives of noisy data with a dedicated strain network that is trained to agree with the displacement-derived strain, so that the equilibrium equations are enforced on a smooth strain representation. The absolute-scale problem is handled separately by the traction boundary condition: after phase one yields a relative modulus map $\\hat{E}(x,y)$ and a relative boundary stress $\\hat{\\sigma}_{xx}^{(b)}$, phase two computes the multiplier $\\hat{c} = F / \\sum_i \\hat{\\sigma}_{xx}^{(b)}(x_b,y_i) h$ from the known applied force $F$, and reports $E_{\\mathrm{absolute}}(x,y) = \\hat{c}\\hat{E}(x,y)$. With arbitrary mean-modulus constraints, the calibration still produces consistent absolute errors, and Poisson's ratio is recovered simultaneously without an incompressibility assumption. The author's claim, in short, is that decoupling displacement, strain, and elasticity networks plus boundary-force calibration turns a noisy ill-posed inversion into a tractable one.","pith_inferences":["Beyond the paper's synthetic benchmarks, the calibration logic predicts that any measurement error in the applied force $F$ enters the absolute modulus scale linearly, so real-world accuracy reports should include force calibration uncertainty.","The strain-discrepancy trick is not specific to elasticity; other inverse PDE problems that differentiate noisy data, such as thermal conductivity imaging or hydraulic tomography, could borrow the decoupling idea.","The paper does not test three-dimensional or low-resolution clinical data, but the same two-phase scheme would need reworked finite-difference kernels and boundary integration before the claim could extend there.","A direct test on experimentally collected digital image correlation data with a known stiffness inclusion would show whether the synthetic-noise robustness transfers to real noise structure."],"forward_implications":["Clinical elastography could estimate absolute tissue stiffness maps from noisy displacement data using only the applied load, without knowing the mean modulus beforehand.","Compressible materials with spatially varying Poisson's ratio can be handled directly, removing the common incompressibility assumption that earlier methods relied on.","Displacement data at SNR down to 100 could be used without aggressive pre-denosing that might blur stiffness boundaries.","The two-phase calibration removes the need for prior knowledge of internal or boundary stress distributions, which are usually unavailable in practice."],"supporting_citations":[{"why":"Provides the finite-element displacement dataset and the direct finite-difference baseline whose failure under noise motivates the proposed architecture.","marker":"[64]"},{"why":"Supplies the convolution-kernel finite differentiation scheme and an earlier relative-modulus-only formulation that the two-phase calibration extends.","marker":"[46]"},{"why":"Supports the periodic sine activation used in all networks, which the paper finds yields the lowest estimation errors.","marker":"[73]"},{"why":"Supports the positional encoding applied to input coordinates, which the paper shows improves accuracy.","marker":"[72]"},{"why":"Documents the ill-conditioning that arises when traction boundary conditions are folded into PINN losses, motivating the separate calibration phase.","marker":"[66]"},{"why":"Provides the physics-informed neural network formulation that the three-network inverse scheme builds on.","marker":"[52]"},{"why":"Gives the general basis for treating neural network fitting as a smoothing operation over noisy observations.","marker":"[65]"},{"why":"Demonstrates that local modulus and force can be measured experimentally, supporting the assumption that the applied loading force is available for calibration.","marker":"[69]"}],"fun_headline_variants":["Force sets stiffness scale in PINN inverse elasticity","Boundary force pins down Young's modulus in noisy inverse problem","Strain net beats noise: IE-PINN maps stiffness and Poisson ratio","No mean-modulus guess needed: PINN calibrates from applied force","Three networks, one clean strain field, true modulus from force"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Phase-two calibration assumes the total applied force on the loaded boundary is known and that the predicted relative boundary stress integrates to a value proportional to that force; in the synthetic benchmarks the force is computed from the true model, so any real measurement error in force magnitude, boundary geometry, or relative stress bias would enter the absolute modulus scale directly.","fun_headline_variants_meta":{"raw":{"variants":["Force sets stiffness scale in PINN inverse elasticity","Boundary force pins down Young's modulus in noisy inverse problem","Strain net beats noise: IE-PINN maps stiffness and Poisson ratio","No mean-modulus guess needed: PINN calibrates from applied force","Three networks, one clean strain field, true modulus from force"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3222,"prompt_tokens":1006,"completion_tokens":2216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2138}},"tokens_in":622,"tokens_out":2216,"duration_ms":14240,"temperature":1.0,"reasoning_tokens":2138,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:55:17.880514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to run the method on a phantom with a known stiffness inclusion while deliberately misspecifying the applied force by 10 percent; if the recovered absolute modulus does not shift by roughly 10 percent, the calibration equation fails, and if the method cannot handle a different heterogeneity pattern at SNR 100, the robustness claim is narrower than stated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the general basis for treating neural network fitting as a smoothing operation over noisy observations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates that local modulus and force can be measured experimentally, supporting the assumption that the applied loading force is available for calibration."}],"review_version":1}