{"id":"5418bba5-b444-4039-8a31-242280d3b0c0","arxiv_id":"2504.12144","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Configured PINNs with Fourier features, periodic mappings, and causal training match or exceed traditional finite-difference solvers on 1D and 2D unsteady Maxwell test cases, though NTK analysis shows convergence effort does not track spatial error.","lead":"This paper tests physics-informed neural networks on time-dependent Maxwell's equations and compares their accuracy with FDTD and compact Pade solvers. With the right enhancements, PINNs can be competitive on simple 1D and 2D benchmarks, but the study's own diagnostics show they do not concentrate learning where errors are largest.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1D 'sharp discontinuity' claim is unquantified and the tanh network cannot represent a true shock, so the main 'outperforms FDTD/Pade' evidence is not established.","rationale":"The reader's weakest assumption identifies exactly the load-bearing weakness: Section 4.1 claims a sharp, oscillation-free resolution of a discontinuity while reporting no error norm, and the smooth tanh architecture cannot represent a true discontinuity. My stress-test agrees and sharpens it: the 1D case is the only place in the paper where PINNs are claimed to exceed, rather than merely match, traditional solvers, and that claim is supported only by a figure. The missing quantitative comparison is not a minor omission because the conclusion 'can match or exceed traditional solvers' is comparative by construction. The dielectric-case and NTK issues noted by the reader are real but secondary: they affect the robustness of the diagnostic and the generality of the claims, whereas the unquantified 1D result directly undermines the headline. The concern is addressable by adding a proper error table, so the appropriate verdict remains CONDITIONAL rather than REJECT; the reader already set CONDITIONAL, so no verdict adjustment is needed.","tokens_in":14749,"tokens_out":8705,"duration_ms":93369,"concrete_test":"Re-run the 1D benchmark exactly as in Section 4.1 and report, at t=0.8, the relative L2 and L_infinity errors of the PINN, FDTD, and Pade solutions against the same high-resolution filtered FDTD reference used in Figure 2, together with the 10%-90% transition width of the PINN near z=0.8 and the maximum overshoot/undershoot for each method. If the PINN's L_infinity error is not smaller than Pade's, or if its transition width spans more than a few collocation spacings, the claim that PINNs outperform traditional solvers on this case is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that 'PINNs, when properly configured, can match or exceed traditional solvers' (Section 6) rests heavily on the 1D Gaussian-pulse result in Section 4.1. The text asserts that the PINN 'accurately captures the field distribution and sharply resolves the discontinuity' and that it 'outperforms both FDTD and Pade methods near steep gradients'. But no error norm is reported for this case: no L2 or L_infinity error against the reference solution, no overshoot/undershoot measurement, and no measure of the transition width near z=0.8. The comparison is purely visual. This matters because the network uses tanh activations, so its output is a smooth function of (z,t); it cannot represent a true derivative discontinuity in E_x. Any apparent 'sharp resolution' must be a finite-width transition, and an oscillation-free appearance is not the same as small L_infinity error. In fact, FDTD and Pade visibly oscillate near the front, and those oscillations are heavily penalized in L_infinity but the PINN's smeared shock may be even more inaccurate in L_infinity while looking 'clean'. Since the first conclusion bullet explicitly claims superiority on this case, the absence of quantitative error comparison leaves the central comparative claim unsupported as written. No error table for FDTD or Pade appears anywhere in the paper, so the 'match or exceed' claim is not backed by a single direct accuracy comparison.","agreement_with_reader":"agree"},"referee_report":null,"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper earns its keep as an ablation study of convergence-enhancing tricks for PINNs on unsteady 2D Maxwell problems, plus a novel NTK-based diagnostic showing that training effort correlates with temporal error but not spatial. The broader claim that PINNs can match or beat FDTD and Pade is not backed by the numbers as written, but the core training-dynamics observations are solid enough to referee.\n\nThe genuinely new parts are the systematic ablations (RFF, strict periodicity, temporal causality) on the 2D periodic and dielectric cases, and the empirical NTK observation that local convergence rates track L2(t) error but not the spatial error map. The tables are informative: causality gives little in the periodic box, spatial periodicity actively hurts in the dielectric case, and RFF is the workhorse. That kind of ablation is useful to people configuring PINNs for wave propagation.\n\nSoft spots, in rough order of importance.\n\nFirst, the 1D 'sharply resolves the discontinuity and outperforms FDTD/Pade' claim rests on a visual comparison. No error norm is reported for any method on that case, and the reference solution itself is a filtered FDTD that shows oscillations. A tanh network cannot represent a true derivative jump; the oscillation-free appearance could come with a smeared front and a worse L_infinity error than the oscillatory solvers. The authors need an L2/Linf table for this case before claiming superiority.\n\nSecond, the text misreads its own Table 1: with hidden width 64, removing causality lowers the L2 error (0.07853 vs 0.08471), the opposite of 'moderate increase.' Also Figure 6's caption says t=1.5 while the text sets t in [0,0.7].\n\nThird, the dielectric case enforces spatial periodicity on a medium with a material interface and never states interface conditions. The residual involves 1/epsilon, which jumps at x=0.4; without continuity conditions, the PDE residual is not well-defined at the interface. The authors acknowledge the approximation, but the comparison to 'traditional solvers' needs this handled explicitly or justified with a sensitivity check.\n\nFourth, the 'match or exceed traditional solvers' conclusion lacks a single direct accuracy table for PINN vs FDTD/Pade on the 2D cases. Subfigure (f) gives a low-res Pade error map, but that is not a comparison of errors at matched resolution or runtime.\n\nThe NTK quantity is a mean eigenvalue of a residual NTK, not a measured convergence rate; the authors should validate it against actual per-region loss decay before calling it a convergence rate. The trainable temporal period Pt is also tuned on the test problems, a mild circularity worth acknowledging.\n\nBottom line: the ablation content deserves a serious referee, and the spatial-vs-temporal NTK observation is worth publishing even if the comparative claims get scaled back. I would send it out, but the revision needs to quantify the 1D and 2D comparisons, fix the internal inconsistencies, and clean up the dielectric interface treatment.","headline":"Useful ablation and NTK diagnostic for PINNs on unsteady Maxwell, but the headline accuracy claims overreach the reported numbers.","tokens_in":15629,"tokens_out":4689,"would_cite":false,"duration_ms":44544,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-16T12:37:02.065459+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}