{"id":"e17d563d-633b-42b9-95bd-af2396687b12","arxiv_id":"2607.04288","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"Mass-weighted Fourier preconditioning drives energy-minimizing PINNs for adhesive line contact to machine-zero loss in ~400 Adam steps and matches GFMD displacement and stress fields.","lead":"A spectral reweighting trick lets physics-informed neural networks solve sticky elastic contact problems that previously stalled in training. It matches classical molecular-dynamics benchmarks on smooth and rough surfaces without needing Green's-function integrals.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the paper's own stated preconditioner assumption.","rationale":"The paper cleanly diagnoses a spectral stiffness pathology (elastic kernel ∝q) that stalls energy-minimizing Fourier PINNs and supplies a practical Fourier-space mass-weighting fix. Within the 1-D Morse setting the evidence is strong: loss reaches machine zero only with MW (Fig. 3), high-q mode is suppressed (Fig. 4), stress becomes physical (Fig. 5), and NRMSE vs GFMD stays ≤~10^{-3} across six cases (Table 1). The reader's weakest_assumption is precisely the paper's own limitation statement; it correctly motivates CONDITIONAL rather than full ACCEPT (missing code URL, untuned hyperparameters, untested 2-D/other potentials). That limitation does not falsify the strongest claim as written. No deeper load-bearing flaw (e.g., incorrect energy assembly, non-reproducible GFMD comparison, or circular validation) appears. Verdict therefore stays CONDITIONAL; no adjustment is warranted.","tokens_in":15358,"tokens_out":555,"duration_ms":6861,"concrete_test":"Re-train the exact Hertz case of Fig. 3 (Rc=1, p0/E*=-1e-3, γ0=1e-3, ρ/L=2e-3) with α varied over {0.25,0.5,0.75,1.0} and qc varied by ±50% around the interaction-range setting; if machine-zero residual and NRMSE≤10^{-3} vs GFMD are retained for all four α values, the fixed-form preconditioner is robust inside the claimed regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption correctly flags the paper's own caveat (end of §2.2): the spectral preconditioner G(q)=w(q)f(q) is designed under the premise that the gradient spectrum is dominated by the elastic stiffness contribution qE*/2. The central claim, however, is scoped to the Morse-adhesive line-contact benchmarks actually run (Hertz + fractal H=0.5, six pressures, nx=2048). Within that scope the ablations (Figs. 3–5) and NRMSE table (Table 1, all entries ≤1.02×10^{-3}) directly support the strongest claim. The abstract's 'extension is direct' language is aspirational and already caveated; it does not underwrite the quantitative 1-D results. No internal inconsistency or hidden failure of the reported experiments is evident.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces a spectral mass-weighting (MW) preconditioner for energy-minimizing Fourier PINNs applied to one-dimensional adhesive elastic contact. The elastic kernel qE*/2 grows linearly with wavenumber and causes high-q modes to dominate Adam updates, stalling macroscopic convergence. Before back-propagation the displacement gradient is Fourier-transformed, reweighted by G(q)=w(q)f(q) (Eqs. 12–14) that amplifies low-q modes and applies a low-pass filter, then inverse-transformed. On Hertz and fractal (H=0.5) Morse-adhesive line contacts the MW-PINN reaches machine-zero residual within ~400 Adam steps (vs. stalling three orders higher without MW) and yields displacement/stress fields whose NRMSE relative to independent GFMD is at most ~10^{-3} across six pressure cases (Table 1).","tokens_in":15603,"tokens_out":948,"duration_ms":9029,"significance":"If the reported ablations and GFMD agreement hold, the work supplies a practical, energy-only training device that removes a concrete spectral-stiffness obstacle for Fourier PINNs in contact mechanics. Strengths that should be credited are the direct MW-on/off loss and fastest-mode diagnostics (Figs. 3–5), the quantitative RMSE/NRMSE table against an external GFMD reference that does not depend on the PINN or the MW parameters (Table 1), and the clean energy formulation that needs no PDE residual or boundary-penalty balancing. The method is restricted to 1-D line contact with Morse adhesion, but within that scope the evidence is concrete and the free parameters of G(q) are stated.","major_comments":[{"comment":"End of §2.2 and the free-parameter list: the central claim that MW is “necessary” for stable convergence rests on a fixed G(q) with α=0.5, ξ, qc, β and clamps that are never varied. A short sensitivity study (or at least a statement of the ranges that still reach machine-zero loss) is needed to show that the acceleration is not an artifact of one hand-tuned setting; without it the necessity claim is only partially supported.","section":null},{"comment":"Abstract and §4: the assertion that “extension to two-dimensional rough surfaces is direct” is stronger than the evidence. The paper itself notes that effectiveness for other potentials or 3-D geometries “remains to be investigated.” Soften the language to a conjecture or supply a minimal 2-D demonstration; otherwise the claim over-reaches the 1-D Morse benchmarks that actually support the quantitative results.","section":null}],"minor_comments":[{"comment":"Abstract and Fig. 3 caption: “machine-zero residual loss” should be quantified (e.g., floating-point floor or a concrete threshold) so that the three-order improvement is unambiguous.","section":null},{"comment":"Eq. (12): the low-q floor k0=ξ(E*qmax/2) and the clamps [wmin,wmax] are introduced without numerical values used in the runs; list them for reproducibility.","section":null},{"comment":"Fig. 3 y-label and caption contain the typo “Resccaled”; correct to “Rescaled.”","section":null},{"comment":"Data-availability statement: replace the placeholder “can be found here” with an actual repository link or DOI before publication.","section":null},{"comment":"Several sentences lack spaces after commas or periods (e.g., abstract “imbalance,that is”); a light copy-edit pass would improve readability.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid, self-contained methods contribution for the contact-mechanics / soft-matter community. The two major points are easily fixable and do not undermine the 1-D results. I see no citation or novelty issues that would require editorial intervention beyond ordinary revision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful thing here is a practical spectral preconditioner that fixes a real training pathology: the elastic kernel ~q makes high-q modes dominate Adam updates, so the macroscopic deformation never settles. With mass-weighting G(q)=w(q)f(q) (plus warm-up and a low-pass), their Fourier PINN hits machine-zero residual in ~400 steps on the Hertz benchmark while the unweighted run stalls three orders higher; the stress field goes from noisy garbage to clean contact patches. That ablation (loss curves, fastest-mode amplitude, stress profiles) is direct and convincing.\n\nWhat is new is the packaging: energy-minimizing PINN + Fourier features + this particular MW form (borrowed from their earlier GFMD work) + explicit filter, applied to Morse-adhesive line contact and checked quantitatively against independent GFMD. Table 1 NRMSE values stay ≤~10^{-3} across six cases (Hertz tension-to-compression and fractal H=0.5 rough surfaces). The energy functional itself is classical; the MW is only a training device, so there is no circularity. Citations look appropriate.\n\nSoft spots are real but proportionate. The free parameters (α=0.5, ξ, qc, β, clamps, warm-up length) are hand-set with no sensitivity study. Code/data link is promised but empty in the manuscript. The abstract’s “extension to 2-D is direct” is aspirational; the paper itself correctly flags that the preconditioner assumes elastic stiffness dominates the gradient spectrum and that other potentials or 3-D remain untested. Within the stated 1-D Morse scope the claims hold.\n\nThis is for people who already run or want to run energy-based PINNs on elastic contact. It does not open a new scientific question, but it removes a concrete obstacle. I would send it to referees; once code is released and a short parameter-sensitivity note is added it is accept-shaped. Worth a look if you care about neural solvers for rough adhesive contact.","headline":"Clean, well-ablated spectral fix that makes energy-based Fourier PINNs actually converge on adhesive half-space contact; 1-D evidence is solid, free parameters and missing code are the real soft spots.","tokens_in":16226,"tokens_out":515,"would_cite":true,"duration_ms":4851,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A mass-weighting spectral preconditioner lets energy-minimizing Fourier PINNs reach machine-zero residual on adhesive contact problems and match GFMD displacement and stress fields.","keywords":["contact mechanics","adhesion","PINN","mass weighting","spectral preconditioning","Fourier PINN","rough surfaces","energy minimization"],"falsifier":"Train the identical network and energy functional on the same Hertz or rough-surface cases with mass weighting disabled (or with a deliberately mismatched G(q)); if residual loss still reaches machine zero and the stress field remains oscillation-free and GFMD-accurate, the claimed necessity of spectral rebalancing is false.","tokens_in":16227,"feed_emoji":"⚙️","tokens_out":911,"duration_ms":7260,"temperature":0.7,"pith_summary":"Physics-informed neural networks that minimize the total potential energy of an elastic contact problem stall because the elastic stiffness grows linearly with wave number: short-wavelength modes dominate the gradients and the macroscopic shape never settles. This paper shows that reweighting those gradients in Fourier space with a mass-weighting function (amplifying low-q modes and low-pass filtering high-q noise) removes the imbalance. On adhesive line-contact benchmarks the preconditioned network reaches machine-zero residual within a few hundred Adam steps and produces displacement and contact-stress fields that agree with Green’s-function molecular dynamics to within a fraction of a percent, for both smooth Hertz indenters and fractal rough surfaces spanning tension to compression. The method works on a plain uniform grid, needs no Green’s-function assembly or quadrature, and is written entirely as scalar energy minimization, so the same spectral idea transfers directly to two-dimensional rough surfaces.","feed_headline":"Mass weighting lets PINNs solve adhesive contact to machine zero","feed_subtitle":"Spectral rebalancing of Fourier gradients matches GFMD stress fields on smooth and rough surfaces","key_machinery":"The mass-weighting spectral preconditioner G(q)=w(q)f(q), where w(q) reweights Fourier modes of the displacement gradient to amplify low-wavenumber contributions and f(q) is a built-in low-pass filter that suppresses sub-grid noise; the reweighted gradient is inverse-transformed and used for ordinary back-propagation.","core_discovery":"Spectral stiffness imbalance is the dominant obstacle to training energy-minimizing Fourier PINNs for elastic contact; a mass-weighting preconditioner applied to the displacement gradient in Fourier space before back-propagation eliminates that imbalance, driving residual loss to machine zero within roughly 400 Adam iterations and recovering displacement and stress fields that match GFMD reference solutions quantitatively for both smooth and multi-scale rough adhesive contacts.","pith_inferences":["The same spectral-reweighting idea is likely to help any energy-based PINN whose linear operator has a strongly growing Fourier symbol, not only contact elasticity.","If the low-pass cutoff is tied to the interaction range, the method may automatically adapt to potentials of different interaction widths without retuning.","Failure modes on three-dimensional or finite-body geometries would most probably appear first as residual high-q oscillations in the stress field rather than as slow macroscopic convergence."],"forward_implications":["Energy-minimizing PINNs become a practical mesh-free alternative to GFMD or BEM for one-dimensional adhesive line contact once spectral preconditioning is used.","The same Fourier elastic energy and wave-number-only preconditioner transfer without structural change to two-dimensional rough surfaces.","Smooth real-space interaction potentials other than Morse (e.g., Lennard-Jones adhesion or rate-dependent friction) can be substituted without redesigning the spectral reweighting.","Training dynamics no longer require manual balancing of multi-term loss weights; a single scalar energy suffices once gradients are spectrally rebalanced."],"fun_headline_variants":["Mass weighting drives Fourier PINNs to machine-zero adhesive contact loss","Spectral preconditioner eliminates PINN stiffness imbalance in contact","Mass-weighted PINN matches GFMD for smooth and rough adhesive contacts","Fourier gradient reweighting converges energy PINNs in 400 Adam steps","Mass weighting fixes spectral imbalance for elastic contact PINNs"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The preconditioner is assumed to work because elastic stiffness dominates the gradient spectrum; if short-range adhesive forces instead dominate the high-wavenumber gradients, the fixed reweighting may fail to rebalance the modes.","fun_headline_variants_meta":{"raw":{"variants":["Mass weighting drives Fourier PINNs to machine-zero adhesive contact loss","Spectral preconditioner eliminates PINN stiffness imbalance in contact","Mass-weighted PINN matches GFMD for smooth and rough adhesive contacts","Fourier gradient reweighting converges energy PINNs in 400 Adam steps","Mass weighting fixes spectral imbalance for elastic contact PINNs"]},"model":"grok-4.5","effort":"low","cost_usd":0.004704,"raw_usage":{"total_tokens":1362,"prompt_tokens":771,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":47040000,"prompt_tokens_details":{"text_tokens":771,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":521,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":771,"tokens_out":70,"duration_ms":4849,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T20:20:18.279498+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the identical network and energy functional on the same Hertz or rough-surface cases with mass weighting disabled (or with a deliberately mismatched G(q)); if residual loss still reaches machine zero and the stress field remains oscillation-free and GFMD-accurate, the claimed necessity of spectral rebalancing is false.","supporting_citations":[],"review_version":1}