{"id":"39e7718a-3b0d-46c9-b7af-5f657637a73d","arxiv_id":"2411.18014","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Using a conformal map to a reference annulus makes a latent neural operator for the 2D Laplace equation about ten times more accurate and far more data-efficient than using LDDMM or optimal transport maps.","lead":"This paper compares three ways of mapping PDE solutions from different shapes onto a single reference shape before training a neural operator. It finds that a conformal map, which preserves the structure of the Laplace equation, needs far fewer training examples than smooth or nonsmooth alternatives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neumann boundary data are not transported correctly under conformal maps: the pulled-back solution's normal derivative acquires the conformal scale factor c, so the latent operator is not geometry-free as claimed.","rationale":"The reader's weakest assumption exactly identifies this issue: the conformal factor c in the transported Neumann boundary condition. My analysis confirms the mathematics: for a conformal map with derivative J = c R, the normal derivative transforms as n0·∇u0 = c (nα·∇uα), so the pulled-back solution satisfies a Neumann condition with data c·b0, not b0. This directly undermines the paper's assertion that conformal maps make the latent representation geometry-independent and eliminate the need for a geometry branch. The central claim about data efficiency may still be true, but the explanation given is at best incomplete and at worst hides a shape-dependent rescaling that the network would need to learn or receive. The proposed test settles whether the empirical result is robust to the correct boundary transport. I keep the verdict CONDITIONAL because the issue is addressable and does not necessarily invalidate the observed qualitative ordering of conformal, LDDMM, and OT mappings; it does, however, require a substantive revision of the theoretical interpretation and possibly of the experimental setup.","tokens_in":11479,"tokens_out":6781,"duration_ms":64319,"concrete_test":"Compute the conformal scale factor c_α(x) = |φ'(x)| on the outer boundary for the Schwarz-Christoffel maps used in Section 4, using the derivative of the SC map from the reference annulus to each target domain. Then: (1) check whether c_α is constant along ∂ΩO0 and across a sample of α; (2) for one test case, compare the pullback solution u0 with the harmonic function on the annulus whose Neumann data are c_α·b0 and with the harmonic function whose Neumann data are b0. If u0 matches the former, Eq. (19) is incorrect as a transport of the boundary condition. (3) Retrain the conformal latent operator with the corrected input c_α·b0 while still omitting the geometry branch; if the 0.26% relative L2 error at 80 samples is preserved, the empirical claim survives, otherwise the 'no geometry branch' result is an artifact of ignoring c_α.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism is that a conformal map makes the latent operator F0 independent of geometry because the Laplace operator is conformally invariant. That would be true for the PDE operator alone, but the numerical experiment uses a Neumann boundary value problem (Eqs. 1-3). For a conformal map φ: Ω0 → Ωα, if u0 = uα ∘ φ, then u0 is harmonic, but its Neumann data on ∂Ω0 is not simply (bα ∘ φ). A direct calculation gives ∂u0/∂n0 = c_α(x) (bα ∘ φ)(x), where c_α(x) = |φ'(x)| is the local scale factor. Thus the true pullback of the solution is not the solution of the reference Laplace problem with the input boundary data b0 = bα ∘ φ; it is the solution with c_α·b0. Equation (19) ignores this factor. Consequently, the claim in Results that conformally mapped solutions 'match the true Laplace solution with the same original boundary conditions' is unsupported, and the conclusion that 'geometry is completely factored out' and 'no geometry branch was required' does not follow: unless c_α is constant on the boundary or is supplied to the network, F0 still depends on α through c_α. The empirical gain may still hold, but the stated explanation of why conformal maps achieve an order-of-magnitude improvement is invalid as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a latent neural operator framework in which PDE solutions on a family of domains are mapped to a fixed reference domain, and a DeepONet-style operator is trained on the pulled-back solutions. The authors compare three mapping strategies for the 2D Laplace equation on doubly connected domains: conformal Schwarz--Christoffel maps, LDDMM diffeomorphisms, and discrete optimal transport. They report that conformal maps, which preserve the Laplace operator, yield the lowest relative L2 error and the best data efficiency, and they argue that operator-preserving mappings should be preferred when constructing latent neural operators.","tokens_in":11794,"tokens_out":4276,"duration_ms":42721,"significance":"The qualitative claim--that preserving the differential operator when mapping to a latent reference domain improves data efficiency--is well motivated and potentially useful for applications with scarce simulation data. The experimental setup cleanly isolates three mapping families on a PDE with a known conformal invariance, and the visual contrast between the mapped solution fields is informative. However, the paper's central mechanistic explanation has a load-bearing gap: the Neumann boundary data are not transported by simple composition under conformal maps. In addition, the quantitative evidence is thin, with single runs, unequal training budgets, and no error bars. If the boundary-condition issue is corrected and the comparison is made fair, the finding would be a useful empirical contribution; in its current form, the stated interpretation of the results is not supported.","major_comments":[{"comment":"Eq. (19) defines the pullback of boundary data as v0_alpha = v_alpha ∘ phi_alpha^{-1}, omitting the conformal scale factor. For the Neumann problem (1)--(3), if u0 = u_alpha ∘ phi_alpha^{-1} with conformal phi_alpha, then a direct calculation gives ∂u0/∂n0 = c_alpha (b_alpha ∘ phi_alpha^{-1}) on the outer boundary, where c_alpha is the local scale factor of the conformal map. Unless c_alpha is constant on the boundary (which it is not for the Joukowski-type inner boundaries used here) or is supplied to the network, the latent operator F0 still depends on alpha through c_alpha. Therefore the statements that conformal maps 'completely factor out' geometry and that 'no geometry branch was required' do not follow from the stated mechanism. The empirical advantage of conformal maps may still hold, but the explanation given in the paper is incorrect as written.","section":"Eq. (19), Results and Discussion"},{"comment":"The numerical comparison conflates mapping quality with training budget and model capacity. Conformal maps use 1,000 epochs and zero PCA modes; LDDMM uses 10,000 epochs and 10 PCA modes; discrete OT uses 50,000 epochs and 10 PCA modes. With only one run per method and no standard deviations, the reported relative L2 errors (0.26%, 2.56%, 22.4%) cannot be attributed to the choice of mapping alone. The data-efficiency claim requires matched training budgets, multiple random seeds, and ideally an ablation that gives the conformal case a geometry branch or that trains all methods with the same number of epochs.","section":"Table 1"},{"comment":"The sample-size comparison is incomplete. The text says that 80 conformally mapped training samples outperform LDDMM with 400 samples, but no full sample-size curves, error bars, or repeated experiments are provided. It is also unclear whether the LDDMM run at 400 samples uses the same architecture, optimizer, and training schedule as the conformal run at 80 samples; without this information, the comparison does not isolate the effect of the mapping on data efficiency.","section":"Figure 3 and Results"}],"minor_comments":[{"comment":"The direction of phi_alpha is inconsistent: it is first defined as a C^2 diffeomorphism from Ω0 to Ωα, but in the numerical experiment the pullback uses phi_alpha^{-1} and the text refers to phi_alpha as mapping Ωα to Ω0. Please standardize the notation throughout.","section":"Section 'Diffeomorphic Mapping Operator Learning' and Eq. (18)-(19)"},{"comment":"The citations for DeepONet and FNO appear to be swapped: the text attributes DeepONet to Li et al. (2020) and FNO to Lu et al. (2021), whereas the standard references are Lu et al. for DeepONet and Li et al. for FNO.","section":"References"},{"comment":"Several hyperparameters and data-generation details are missing, including the reference annulus radii, mesh resolution, train/test split size, number of test domains, learning rate, optimizer, and network width/depth. These details are needed to reproduce or interpret the reported errors.","section":"Table 1 and Experimental Setup"},{"comment":"The compatibility condition is written with a missing dA in the first integral, and the claim that the second equality follows from the divergence theorem is unclear because it also uses ∇²u = 0; please rewrite the derivation more carefully.","section":"Eq. (5)"},{"comment":"The caption does not clearly describe the layout: it refers to 'first three columns' and a 'rightmost panel,' but the figure appears to contain multiple grouped columns per mapping method. Please clarify the arrangement of rows and columns.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is plausible and the experimental design is simple and readable, but the boundary-condition error in the conformal argument is load-bearing, and the quantitative comparison is not yet rigorous enough for a journal publication. I do not see evidence of a deeper circularity; the issues are technical and can in principle be fixed within the scope of the manuscript. The authors should also consider whether the contribution is sufficiently substantial relative to the existing DIMON framework (Yin et al. 2024), since the present work appears to be an empirical study of mapping choices within that framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's contribution is a side-by-side comparison of three ways to pull PDE solutions back to a reference domain for neural operator training: conformal (Schwarz-Christoffel), LDDMM, and discrete optimal transport. That comparison is new and useful, and the headline result that conformal maps make training dramatically more data-efficient is plausible. The authors deserve credit for stating the mechanism clearly: the closer the mapped solutions are to solutions of the same PDE on the reference domain, the easier the latent operator is to learn.\n\nBut the mechanism as stated is wrong for their actual problem. They solve a Neumann boundary value problem, Eqs. (1)-(3). Under a conformal map with local scale factor c, the pulled-back solution is harmonic, but its Neumann data is c times the pulled-back boundary function, not the pulled-back function alone. Equation (19) omits this factor. Unless c is constant on the boundary (and they never report the modulus or boundary conformal factors), the latent problem still depends on the target geometry through c. That undermines the claim that the conformal representation is geometry-free and that no geometry branch is needed. The empirical gain might still hold, but the explanation is not valid as written.\n\nThe experimental support is also thin. Table 1 gives one run per method, no standard deviations, and the methods are trained for different numbers of epochs (1,000 vs 10,000 vs 50,000) with different PCA modes. That conflates mapping quality with training budget. There is no comparison to standard baselines (e.g., DeepONet trained directly on the original domains) and no code or data released. These are fixable, but they should be fixed before the claim is accepted.\n\nThe core idea—that operator-preserving maps improve data efficiency in latent neural operators—is sound and worth pursuing. The paper is a reasonable motivating study, but it reads as a preliminary report, not a definitive demonstration. I'd send it to peer review, with the expectation of substantial revision: correct the boundary condition pullback, add error bars and matched training budgets, and include at least one standard baseline.\n\nIf you're working on geometry-generalization for neural operators, this is worth a skim for the comparison table alone. But I wouldn't cite it yet.","headline":"A useful empirical comparison of mapping choices for latent neural operators, but the conformal invariance argument mishandles Neumann boundary data and the experimental evidence is under-powered.","tokens_in":12321,"tokens_out":2439,"would_cite":false,"duration_ms":22226,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","35J05","30C30","65N30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that mapping all PDE domains to one reference shape with a map that preserves the differential operator, rather than merely increasing training data, is what makes a latent neural operator accurate and data-efficient…","keywords":["neural operator","Laplace equation","conformal mapping","Schwarz-Christoffel","LDDMM","latent operator","data-efficient learning","geometric generalization"],"falsifier":"Train the conformal latent operator and test it on target shapes that share the same physical outer-boundary condition but differ strongly in the conformal scale factor along that boundary; if the representation is truly geometry-free, prediction errors should not track that factor, whereas a systematic correlation would expose the residual shape dependence in the pulled-back boundary data.","tokens_in":11299,"feed_emoji":"📐","tokens_out":10321,"duration_ms":82519,"temperature":0.7,"pith_summary":"The paper argues that a latent neural operator trained on solutions pulled back to a fixed reference domain becomes dramatically more data-efficient when the pullback map preserves the differential operator of the PDE, not merely when the map is smooth and invertible. Working with the 2D Laplace equation on planar domains with one inner hole, it compares conformal maps, LDDMM diffeomorphisms, and discrete optimal transport as transporters of solutions to a fixed ring-shaped reference domain. Because conformal maps leave the Laplacian invariant, the mapped solutions remain true Laplace solutions on the reference domain, so the network needs no geometry information and learns only the boundary-condition-to-solution relationship. This yields 0.26% relative L2 error from 80 conformally mapped training samples, versus 2.56% for LDDMM with 400 samples and 22.4% for optimal transport, showing that mapping design can substitute for large datasets.","feed_headline":"Conformal map gives 0.26% PDE error on 80 samples","feed_subtitle":"Preserving the PDE operator, not smoothness, is what buys data efficiency.","key_machinery":"The load-bearing identity is the conformal invariance of the Laplacian: a harmonic function composed with a conformal map is again harmonic, so a conformal pullback converts the family of Laplace problems on varying domains into one Laplace problem on a fixed reference annulus. The concrete construction composes the inversion h(z)=1/z, which fixes the unit-circle inner boundary, with a Schwarz–Christoffel exterior map g_alpha that sends the unit circle to the target inner boundary, giving $phi_alpha^{{-1}}$=g_alpha∘h; the outer boundary is carried along passively, which keeps all generated domains conformally equivalent. The latent operator F0 on the annulus is then approximated by a neural operator whose inputs are encoded in a physical-condition branch and, except in the conformal case, a geometry branch. In the conformal case no geometry encoding is needed: the mapping has already quotiented out the geometry, so the learning problem reduces to the boundary-condition-to-solution map.","core_discovery":"The central claim is that the data efficiency of the latent operator F0 is governed by how faithfully the map preserves the differential operator. For the Laplace equation on doubly connected domains, the paper constructs F0 on a fixed annulus and evaluates three transport maps. The Schwarz–Christoffel conformal map sends each domain to a conformally equivalent annulus; because the Laplacian is conformally invariant, the pulled-back solution is exactly the Laplace solution on the reference domain with the original boundary data pulled back, so the geometry branch of the network can be dropped entirely. LDDMM maps are smooth and invertible but do not preserve angles, so the mapped fields deviate slightly from true Laplace solutions on the annulus, and that small deviation is enough to raise relative L2 error from 0.26% to 2.56% with five times more data. Discrete optimal transport, lacking smoothness, produces noisy mapped solutions and 22.4% error. The paper concludes that even small deviations from PDE preservation degrade neural operator training substantially.","pith_inferences":["Going beyond the paper: for any PDE with a known invariance group, the reference map should be taken from that group, and for PDEs without exact invariance the map should be chosen to minimize a pullback-operator penalty rather than geometric distortion alone.","The paper leaves unmeasured the conformal modulus and the boundary scale factors of the generated domains; a testable consequence is that the conformal latent operator's error should not correlate with those quantities, and any correlation would indicate that the pulled-back boundary data carry a shape-dependent factor.","This suggests that in data-scarce applications such as patient-specific modeling, the highest-value investment is in constructing operator-preserving maps for each new geometry rather than collecting more simulation samples across geometries."],"forward_implications":["Choosing a mapping that preserves the differential operator can substitute for training data: 80 conformally mapped samples beat 400 LDDMM-mapped samples by an order of magnitude in relative L2 error.","When the mapping preserves the operator, the geometry encoding is no longer needed (0 PCA modes), reducing the latent learning problem to the boundary-condition-to-solution map on the reference domain.","Small distortions in the mapped solution fields, even ones that are barely visible in plotted comparisons, translate into large performance gaps, so mapping fidelity is a first-order factor in neural operator training.","For harmonic problems on doubly connected domains, the conformal construction makes the latent operator essentially independent of the target geometry, so a single geometry-free model covers the whole family of domains."],"supporting_citations":[{"why":"Supplies the diffeomorphic mapping operator learning framework, including the latent operator F0 and the pullback construction that this paper adapts.","marker":"Yin et al. 2024"},{"why":"Provides the Schwarz–Christoffel toolbox implementation used to compute the conformal maps numerically.","marker":"Driscoll 1996"},{"why":"Establishes the conformal invariance of the Laplacian, the property on which the main data-efficiency claim rests.","marker":"Parker and Rosenberg 1987"},{"why":"Defines the LDDMM framework used as the smooth-but-non-conformal comparison mapping.","marker":"Beg et al. 2005"},{"why":"Supplies the discrete optimal transport formulation used as the non-diffeomorphic benchmark.","marker":"Peyré, Cuturi et al. 2019"},{"why":"Provides the multiple-input operator network architecture adapted to approximate the latent operator F0.","marker":"Jin, Meng, and Lu 2022"}],"fun_headline_variants":["Conformal maps give 10x PDE accuracy with 5x less data","Preserve the PDE operator, not smoothness, for data-efficient learning","Conformal invariance gives 0.26% PDE error with 80 samples","Operator-preserving maps key to low-data PDE surrogates","Keeping the Laplacian intact boosts neural operator efficiency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument carries if pulling the boundary data back by simple composition with the inverse conformal map is the full and correct way to transport the PDE data, so that the mapped problem is genuinely identical on every target shape; if the map's local stretching along the outer boundary leaves a shape-dependent factor in the boundary term, the geometry has not actually been fully factored out.","fun_headline_variants_meta":{"raw":{"variants":["Conformal maps give 10x PDE accuracy with 5x less data","Preserve the PDE operator, not smoothness, for data-efficient learning","Conformal invariance gives 0.26% PDE error with 80 samples","Operator-preserving maps key to low-data PDE surrogates","Keeping the Laplacian intact boosts neural operator efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000814,"raw_usage":{"total_tokens":3588,"prompt_tokens":986,"completion_tokens":2602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2510}},"tokens_in":602,"tokens_out":2602,"duration_ms":15856,"temperature":1.0,"reasoning_tokens":2510,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:38:09.776491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the conformal latent operator and test it on target shapes that share the same physical outer-boundary condition but differ strongly in the conformal scale factor along that boundary; if the representation is truly geometry-free, prediction errors should not track that factor, whereas a systematic correlation would expose the residual shape dependence in the pulled-back boundary data.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Schwarz–Christoffel toolbox implementation used to compute the conformal maps numerically."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the conformal invariance of the Laplacian, the property on which the main data-efficiency claim rests."},{"cited_title":"F.; Miller, M","cited_arxiv_id":null,"evidence_quote":"Defines the LDDMM framework used as the smooth-but-non-conformal comparison mapping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multiple-input operator network architecture adapted to approximate the latent operator F0."}],"review_version":1}