{"id":"8cb8596f-fda4-44f8-808e-4357bf1f278c","arxiv_id":"2607.22215","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"By mapping PDE residuals and boundary conditions to a fixed latent geometry through deformation gradients, latent PDE mapping improves out-of-distribution geometric generalization in PINNs and PI-DONs for the Aliev-Panfilov cardiac model, with 4–6× lower L2 error on rotation-dominated families.","lead":"This paper introduces a training trick for physics-informed neural networks that need to solve the same equations on many different shapes: pull every geometry back to one fixed reference shape using deformation gradients, so the network learns from a small set of shapes and transfers to new ones. It works especially well on rotated geometries, cutting error 4–6× in those cases with almost no extra cost at inference.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline '4–6x improvement' rests on a single training run per model; without multi-seed replication, the paired Wilcoxon does not quantify stochastic variability and the central generalization claim is not yet supported.","rationale":"The reader's weakest assumption—that freezing F and J drops the deformation-map sensitivity terms—is real, but it is not the most load-bearing threat to the paper's empirical claim. For the rotation families that drive the headline results, the mapped residual is actually independent of F and J because the deformation and fiber orientation rotate together; moreover, F and J do not depend on network parameters, so the omitted terms do not alter the weight gradients used in training. The more decisive gap is the lack of multi-seed replication. The entire quantitative case for '4–6x improvement' comes from one training run per model, and the reported significance tests do not account for stochastic training variability. This is explicitly admitted in the limitations. If re-running with different seeds reverses or substantially shrinks the gaps, the central claim of robust geometric generalization is unsupported. If the gaps persist, the conditional acceptance remains justified, but the paper's mechanistic narrative should still be corrected. I therefore agree with the CONDITIONAL verdict, while shifting the technical emphasis from the frozen-F/J issue to the single-seed statistical fragility.","tokens_in":22270,"tokens_out":12753,"duration_ms":145280,"concrete_test":"Run at least 10 independent training runs with different random seeds/initializations for LPM-PINN and LG-PINN on G*_rot, and for LPM-PINN vs LG-PINN on Hy*_rot and Hz*_rot, using identical data and hyperparameters. Report the distribution of mean relative L2 errors over seeds and a paired comparison across seeds. If the 4.3x/5.1x/3.9x gaps persist with non-overlapping intervals, the empirical claim is robust; if not, the central claim fails. As a secondary check, instrument the frozen-F/J LPM on G_nonlin/G_shear to compute the full shape-gradient terms in Eq. (33) and report their magnitude relative to the solution-sensitivity terms, to determine whether the omitted deformation-map sensitivity is actually negligible.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim—factor ~4–6 reduction in mean relative L2 on rotation families—is generated from a single training run per model (Tables 3 and 4). The reported standard deviations are across test geometries, not across random seeds, and the paired Wilcoxon test treats each geometry as an independent observation from one fixed run. This cannot establish that LPM's advantage is stable under reinitialization, which is necessary for the claim of 'generalizable models from limited data.' The authors explicitly acknowledge this in Section 4.1: 'reported experiments were conducted without a systematic evaluation across multiple random seeds.' A related but secondary concern is that the advertised shape-gradient mechanism is not actually exercised—F and J are frozen (Appendix A.1, Eq. 34), and since s is an input rather than an optimized variable, the omitted ∂F/∂s and ∂J/∂s terms do not affect the network-weight updates. The mechanism story is therefore unsupported, but the residual pullback itself remains a valid algorithmic contribution. If seeding changes the ranking, the headline quantitative claim collapses; if not, the concern lands only on framing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes latent PDE mapping (LPM), a method for training physics-informed neural networks (PINNs and PI-DONs) on families of geometries. The PDE residual and boundary terms are pulled back from physical domains Ω(s) to a fixed latent domain Ω0 using the deformation gradient F and Jacobian J, and the solution is represented as a function of latent coordinates and a shape descriptor s. The method is instantiated on the anisotropic Aliev–Panfilov cardiac electrophysiology model in 2D and 3D, with training on 10 geometries per family and evaluation on held-out internal and external parameter ranges. The reported mean relative L2 errors show consistent but modest gains on expansion/shear families and large gains on rotation families, with a ~4–6× reduction claimed on external rotations. Ablations include LG-PINN/DON (latent geometry but no residual pullback), PA-PINN, and Basic-PINN, with either deformation parameters or PCA scores as shape descriptors.","tokens_in":22472,"tokens_out":10257,"duration_ms":106205,"significance":"The underlying pullback construction is mathematically standard and clearly presented; if the empirical claims hold, LPM would be a useful, architecture-agnostic ingredient for geometry-generalizable physics-informed models. The paper deserves credit for a detailed derivation, public synthetic data, and a broader comparison than many PINN papers. However, the headline quantitative claims hinge on experimental details that are not yet established: the external rotation families may not actually be outside the training distribution under 180° symmetry, all results are single-seed, and the implemented loss does not exercise the full advertised shape gradient. The significance of the contribution therefore remains conditional.","major_comments":[{"comment":"The claimed out-of-distribution gains on rotation families are undermined by a symmetry degeneracy. Under the natural centered-coordinate interpretation of the latent square/cube, rotation by θ+180° maps Ω0 to itself, and since the diffusion tensor D is invariant under R180 (D = R180 D R180^T), the physical problem for θ+180° is identical to that for θ. Because G*_rot / H*_rot are defined only by θ ∉ [-90,90], every external angle is equivalent to an internal one (e.g., 120° ≡ -60°). Thus the ~4.3–5.1× improvements on G*_rot, Hy*_rot, Hz*_rot in Tables 3–4 do not demonstrate generalization to unseen geometries. Please either use a latent geometry without 180° symmetry or choose an external range within one symmetry period (e.g., internal [-45,45], external [45,135] with endpoints excluded).","section":"Section 2.2 / Table 1 / Tables 3–4"},{"comment":"No multi-seed evaluation is reported. The ± values in Tables 3 and 4 are standard deviations across test geometries for a single training run; the paired Wilcoxon tests treat each geometry as an independent observation from that one run and cannot quantify stochastic variability. As the authors acknowledge in Section 4.1, the influence of random initialization is unquantified. The central claim of 'generalizable models from limited data' requires at least 3–5 seeds per configuration with the resulting error distributions reported, especially for the headline rotation-family comparisons.","section":"Section 4.1 / Tables 3–4"},{"comment":"The implemented shape gradient drops the deformation-map sensitivity terms. The text claims Eq. (12) computes the shape gradient directly and that 'the integrand does not vary with s', but R_LPM in Eq. (6) depends on s through F and J; the full chain rule in Eq. (33) includes ∂R/∂F ∂F/∂s and ∂R/∂J ∂J/∂s. In training these terms are absent because F and J are precomputed and treated as constant coefficient fields (Appendix A.1). Consequently, the advertised mechanism—accurate shape-gradient calculation—is not what is exercised in the experiments; only the solution-sensitivity part (34) remains. Please either implement the full gradient or revise the motivation/claims to state that the benefit of LPM in these experiments comes from the residual pullback itself, not from deformation-map gradients.","section":"Section 2.1.2 / Eq. (12) / Appendix A.1, Eq. (34)"},{"comment":"The PCA descriptors are constructed using all 85 geometries, including the 35 external test geometries ('The sampled geometries covered the full parameter ranges of the internal and external geometric families... n=85'). The test points therefore influence the PCA basis and the scores used as network inputs. This is a form of test leakage for the PCA experiments in Tables 3 and 4 and weakens the claim that LPM works with low-dimensional descriptors on unseen geometries. PCA should be fit on the training geometries only, or the experiment should be repeated in a fully nested fashion.","section":"Section 2.2"}],"minor_comments":[{"comment":"The boundary-condition integral should be over ∂Ω0 with dS, not over Ω0 with dΩ0; the current equation is dimensionally inconsistent.","section":"Appendix A, Eq. (30)"},{"comment":"The sentence 'the integrand does not vary with s' is inaccurate: R_LPM depends on s through F, J, and u. Consider rephrasing to 'the integration domain does not vary with s'.","section":"Section 2.1.2"},{"comment":"The text refers to 'best and second-best performing models'; please clarify explicitly that comparisons are within each architecture (PINN vs PINN, DON vs DON), as a reader may otherwise attempt cross-architecture comparisons that the paper says are out of scope.","section":"Tables 3–4"}],"recommendation":"major_revision","confidential_remarks":"The symmetry issue in the rotation families is the most serious: if confirmed, the headline 4–6× improvement is not an out-of-distribution result. The paper can be repaired by rerunning rotations with a non-symmetric latent geometry and by adding multi-seed evaluation. I would not accept as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper deserves a real referee, but the abstract oversells the mechanism and the central quantitative claim is built on single training runs. The core idea—pull back geometry-specific PDE residuals and boundary terms to a reference geometry via the deformation gradient—is not new in continuum mechanics, but integrating it into PINN and PI-DON training for a nonlinear, time-dependent PDE with sharp gradients, using only 15 training geometries, is a legitimate and useful contribution. The experiments are broad for a method paper: 2D and 3D, four deformation families, two descriptors (deformation parameters and PCA), two architectures, and out-of-distribution test sets. The 4–6x error reductions on the rotation families are large and the visualizations support them. The FEM data are public, which is a plus.\n\nNow the soft spots, in proportion. First, the single-seed issue is real and the authors admit it in Section 4.1. The paired Wilcoxon tests compare per-geometry errors within one fixed run; they say nothing about stability across reinitializations. With effect sizes that large, I suspect the ranking would survive reseeding, but that is a guess, not a demonstrated result. A multi-seed study with error bars across seeds is needed before the headline claim is solid.\n\nSecond, the shape-gradient story. Section 2.1.2 promises that LPM enables accurate geometry-consistent shape gradients, and the abstract repeats this. But Appendix A.1 shows that F and J are precomputed and frozen, so the deformation-map sensitivity terms in Eq. (33) are dropped. What remains is only the solution-sensitivity part. The authors do disclose this in the discussion and appendix, which is good, but it means the mechanism advertised as the reason for the gains was not actually exercised. The residual pullback itself is still a valid algorithmic change—it changes the loss landscape in a way that plainly helps—so the paper is not wrong, just overframed.\n\nThird, no direct comparison to the closest prior work (Burbulla's transformed-geometry PINN, JacobiNet). The paper cites them but never runs them. That makes the incremental contribution harder to calibrate. Also, no code is released, which limits reproducibility beyond the published numbers and data.\n\nBottom line: this is a promising method paper with honest limitations and a genuinely useful empirical contribution for the cardiac/digital-twin crowd. Send it to peer review, but require multi-seed runs, a code release, and either a full implementation of the shape gradient or an explicit statement that the contribution is the residual pullback, not the full gradient story.","headline":"The pullback idea is standard, but this is a careful, honestly reported test of it on a hard nonlinear PDE with limited data; the big wins on rotations are real but rest on single runs, and the advertised shape-gradient mechanism is not what was actually implemented.","tokens_in":22998,"tokens_out":2236,"would_cite":true,"duration_ms":26369,"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":"Training physics-informed networks on a shared latent geometry, with PDE residuals pulled back through the deformation gradient, yields large accuracy gains on unseen rotated geometries when training data is scarce.","keywords":["latent PDE mapping","physics-informed neural networks","geometric generalization","Aliev-Panfilov","deformation gradient","limited data","cardiac electrophysiology","shape gradients"],"falsifier":"Compute the two deformation-map sensitivity terms ∂R/∂F · ∂F/∂s and ∂R/∂J · ∂J/∂s for the rotation families and check whether their magnitudes are negligible against the solution-sensitivity terms; alternatively, run latent PDE mapping on geometries whose deformation gradient is approximated by noisy or estimated maps and see whether the 4-6x advantage over baselines disappears.","tokens_in":22094,"feed_emoji":"⚡","tokens_out":3231,"duration_ms":34357,"temperature":0.7,"pith_summary":"This paper claims that training physics-informed networks on a predefined latent geometry—pulling PDE residuals and boundary conditions back through the deformation gradient—recovers shape-gradient information that ordinary physics-informed losses discard. On the nonlinear Aliev-Panfilov cardiac PDE with only fifteen training geometries, the mapped models achieve roughly four to six times lower mean relative L2 error than latent-geometry-only baselines on rotation-dominated shape families, in both 2D and 3D, and with either explicit deformation parameters or PCA descriptors. Gains are modest for affine and shear families, indicating the technique helps most when geometric variation substantially changes the solution.","feed_headline":"Pulling PDEs to one latent geometry cuts error 4-6x","feed_subtitle":"On unseen rotated cardiac shapes, mapped physics losses beat plain PINNs trained on just 15 geometries.","key_machinery":"The deformation gradient F(X,s)=∇_X Φ^{-1}(X,s) and its determinant J, used to pull back the strong-form PDE, boundary conditions, and data from each physical geometry to a preselected latent domain. In the implementation, F and J are precomputed per geometry and treated as constant coefficient fields during training, so the shape gradient actually computed is the solution sensitivity ∂V/∂s and ∂W/∂s, not the deformation-map sensitivity ∂F/∂s and ∂J/∂s.","core_discovery":"Latent PDE mapping turns a geometry-dependent training problem into a parametric one: the PDE residual and boundary conditions are rewritten on a fixed reference domain with the deformation gradient F and Jacobian J encoding each geometry. The authors show that conventional physics losses implicitly drop a boundary-motion term when taking shape derivatives, whereas the mapped formulation lets the network's geometry sensitivity flow through the solution on a fixed domain. On the Aliev-Panfilov model, this yields error reductions of 4.3x in 2D rotation families and 3.9-5.1x in 3D rotations about the y- and z-axes, where rotations alter the fiber orientation and therefore the solution most.","pith_inferences":["The theoretical motivation advertises the full shape gradient including ∂F/∂s and ∂J/∂s, but the experiments freeze F and J; if those deformation-map sensitivity terms are non-negligible, the empirical gains may come less from 'accurate shape gradients' than from simply evaluating residuals on a fixed reference domain.","A testable screening rule follows from the paper's own measurements: latent PDE mapping should help most when the boundary-motion term ∂x/∂s·n is large and the residual on the boundary is non-negligible, which is consistent with the small gains on expansion and shear families.","For real-world anatomies without closed-form deformation maps, the method would need numerically estimated F and J; performance likely degrades with map-error, and the paper leaves this unquantified.","Static deformations are assumed throughout; time-dependent or moving-boundary problems would require reinstating the dropped ∂F/∂s and ∂J/∂s terms, so the practical extension to morphing geometries is an open question."],"forward_implications":["In geometric families where boundary motion is large and the PDE solution changes substantially, latent PDE mapping gives roughly 4-6x lower mean relative L2 error than latent-geometry-only baselines, across two network architectures and two geometric descriptors.","Because the PDE is evaluated on a fixed latent domain, collocation points can be shared across geometries, and the added cost is modest during training and negligible at inference since F and J are computed once during meshing.","The framework is architecture-agnostic and formulated for diffeomorphic geometries, so the same pull-back construction can be attached to graph networks, neural operators, or other physics-informed backbones.","The paper's numerical quantification shows that the boundary-motion term neglected by conventional physics losses is larger than the retained shape gradient in every 2D family tested, giving a concrete diagnostic for when latent PDE mapping should help.","Effectiveness persists when geometries are described by PCA scores rather than explicit deformation parameters, which is the more realistic setting for anatomical populations."],"fun_headline_variants":["Latent PDE mapping yields 4-6x lower error on unseen rotated hearts","From 15 training geometries to accurate physics models via mapping","Fixed latent geometry makes physics-informed learning geometry-aware","Mapped losses help PINNs see new shapes with limited data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes an exact deformation gradient from each physical geometry to the latent one, and the implemented training treats F and J as frozen constants, so if real geometries lack known maps or the map's own sensitivity matters, the advertised shape-gradient mechanism and practical applicability are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Latent PDE mapping yields 4-6x lower error on unseen rotated hearts","From 15 training geometries to accurate physics models via mapping","Fixed latent geometry makes physics-informed learning geometry-aware","Mapped losses help PINNs see new shapes with limited data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000432,"raw_usage":{"total_tokens":2052,"prompt_tokens":768,"completion_tokens":1284,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":1213}},"tokens_in":512,"tokens_out":1284,"duration_ms":13737,"temperature":1.0,"reasoning_tokens":1213,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:26:14.599702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the two deformation-map sensitivity terms ∂R/∂F · ∂F/∂s and ∂R/∂J · ∂J/∂s for the rotation families and check whether their magnitudes are negligible against the solution-sensitivity terms; alternatively, run latent PDE mapping on geometries whose deformation gradient is approximated by noisy or estimated maps and see whether the 4-6x advantage over baselines disappears.","supporting_citations":[],"review_version":1}