{"id":"99154f74-a5ac-4028-acde-c223eb1ff91b","arxiv_id":"2507.03087","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A framework that uses neural implicit geometry representations to feed shifted-boundary finite element simulations, removing explicit surface meshing for linear elasticity on complex shapes.","lead":"This paper connects neural networks that store 3D shapes as smooth distance fields directly to a solid-mechanics solver, so simulations no longer need a cleaned-up surface mesh. It demonstrates the pipeline on linear-elasticity tests of the Stanford Bunny, an Eiffel Tower model, and a gyroid.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not yet supported: no convergence study runs the INR in the loop, and Section 5.1 shows distance-vector errors concentrated at sharp curvature, so the link between INR accuracy and SBM's O(h^2) consistency is unverified.","rationale":"The paper's intended contribution is a workflow, not a new PDE method: it couples a trained INR to SBM for linear elasticity. For that workflow to deliver 'high-fidelity' simulation, the INR must supply distance vectors accurate enough not to break SBM's convergence theory. The manuscript provides strong standalone evidence for INR quality (NMSE ~1e-7–1e-6, mean cosine similarities ≥0.98) and strong standalone evidence for SBM (ring convergence slope ~2), but the composition of the two is tested only once, at a single refinement, on an icosphere. That single test does not reveal whether the INR error is asymptotically smaller than the discretization error. The reader's weakest_assumption identified exactly this gap, and the sign-convention conflict between Eq. 9 and Algorithm 4 reinforces it: if the implementation follows Algorithm 4's sign convention (negative inside), then Eq. 9 as written supplies the wrong vector direction, so the reported 2.99e-4 error cannot be traced to the paper's stated formula. Because the concern is about missing validation and internal consistency rather than an inevitable flaw, the CONDITIONAL verdict is appropriate; the requested convergence and sign check are inexpensive and would settle the issue.","tokens_in":17630,"tokens_out":5382,"duration_ms":64048,"concrete_test":"Re-run the Section 5.2.2 icosphere manufactured-solution case with the same trained INR at octree base levels 4, 5, and 6 (boundary levels 8, 9, 10), and plot the L2 displacement error against element size h; if the slope departs from the SBM-only value near 2 or flattens, the INR distance-vector error is the limiting factor. In the same run, flip the sign in Algorithm 4 to match Eq. 9 and record the displacement; a material change would indicate the sign inconsistency affects the solution, while no change would indicate the inconsistency is confined to notation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's accuracy rests on the distance vector d from Eq. 9 being a valid closest-point projection to the true boundary, because SBM's shifted Taylor condition (Eq. 20) and the consistency/adjoint-consistency terms (Eq. 24) require d=x−x̃ with ∥d∥=O(h). For a learned INR this holds only if f_θ is an exact SDF; Section 5.1 shows this is not the case: cosine-similarity errors reach 1−0.980=0.02 on average and SD 0.13 for the turbine, and misalignment concentrates at sharp features. The paper validates SBM alone on a ring (Section 5.2.1) and reports one fixed-refinement icosphere comparison (Section 5.2.2); there is no INR-in-the-loop mesh-convergence study. If the INR's distance-vector error is O(1) in a thin boundary layer rather than O(h), the Taylor expansion in Eq. 20 has an error that does not vanish under refinement, and the method may not converge at the expected rate or at all. A separate sign issue compounds this: Eq. 8 defines f>0 inside and Eq. 9 gives d=−f∇f, while Algorithm 3 treats f<0 as interior and Algorithm 4 computes d=(∇f/∥∇f∥)f without the minus sign. Under either convention, one of the two formulas points the distance vector the wrong way. The manuscript's own Section 5.1 error plots are consistent with this being a real accuracy bottleneck rather than a cosmetic inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a computational pipeline that couples implicit neural representations (INRs) of geometry with the shifted boundary method (SBM) for linear elasticity. The authors train an MLP-based signed distance field from a triangle soup using a hybrid sampling strategy and a loss that combines a clamped distance term with eikonal and normal-similarity regularization. The trained INR is then used to classify octree cells, construct an incomplete octree, identify the surrogate boundary, and compute distance vectors for SBM's shifted Taylor boundary condition. The paper validates the SBM solver on a ring benchmark with a known analytical solution, compares an INR-based icosphere against its triangle-mesh counterpart in one configuration, and presents simulation results for the Stanford Bunny, the Eiffel Tower, and a gyroid.","tokens_in":17971,"tokens_out":6043,"duration_ms":71219,"significance":"If substantiated, the framework would remove the explicit surface triangulation from the simulation pipeline and make boundary-related geometric queries independent of the surface triangle count, a practically useful integration of INR geometry with embedded finite element analysis. The paper has several strengths: the SBM ring convergence test is a sound check of the solver; the proposed loss and sampling strategy are clearly specified; and the wall-time comparison in Figure 11 directly demonstrates the expected asymptotic advantage of neural queries over triangle-soup traversal. However, the central claim that INR geometry can replace the meshing step for accurate SBM analysis is not yet fully supported: the only INR-in-the-loop accuracy test is a single fixed-refinement icosphere case, the distance-vector sign convention is internally inconsistent, and Section 5.1 documents substantial distance-vector misalignment near sharp features. These are load-bearing gaps because SBM consistency relies on the distance vector being an accurate closest-point vector of size O(h).","major_comments":[{"comment":"The sign convention for the signed distance field is internally inconsistent. Section 2.3, Eq. (8) defines f > 0 inside the domain and f < 0 outside, and Eq. (9) gives d = -f_theta grad f_theta. In contrast, Algorithm 3 classifies a Gauss point as interior when f_theta(gp) < 0, and Algorithm 4 computes d_gp = (grad f_theta / ||grad f_theta||) f_theta(Q). Under the Eq. (8) convention, Eq. (9) points from an interior point toward the interior, opposite to the closest-point projection to the true boundary; under the Algorithm 3 convention, Algorithm 4 points inward rather than outward. The discrepancy is not cosmetic because the shifted Taylor condition in Eq. (20) requires d = x - x-tilde to point from the surrogate boundary to the true boundary. Please correct the sign convention consistently across Eq. (8), Eq. (9), Algorithm 3, and Algorithm 4, and state the chosen convention explicitly.","section":"Section 2.3, Eq. (8)-(9); Algorithm 3; Algorithm 4"},{"comment":"The convergence study in Section 5.2.1 uses an analytic ring geometry and therefore validates the SBM formulation, not the INR-in-the-loop pipeline. The only direct INR comparison, Section 5.2.2, is a single icosphere at base level 4 and boundary level 8, reporting one surface L2 error of 2.99e-4 with no mesh refinement sequence, no repeated training runs, and no error bars. To support the claim that the INR-SBM combination achieves the expected second-order convergence, please report L2 displacement error versus mesh size for the INR-based icosphere (or another INR geometry) over several refinement levels, including at least three independently trained INRs to quantify training variability.","section":"Sections 5.2.1-5.2.2"},{"comment":"Table 5 reports a mean cosine similarity of the distance vector of 0.980 with standard deviation 0.130 for the turbine, and Figure 7 shows that misalignment concentrates at sharp edges. Because the consistency and adjoint-consistency terms in Eq. (24) rely on d = x - x-tilde with ||d|| = O(h), an O(1) angular error in a boundary-layer region can introduce an error in the Taylor remainder of Eq. (20) that does not vanish under mesh refinement. Please quantify the effect of the measured distance-vector errors on the SBM consistency, for example by comparing ring solutions with exact and INR-perturbed distance vectors, or by reporting the convergence of the INR distance-vector error with respect to h at the Gauss points used by the solver.","section":"Section 5.1, Table 5"},{"comment":"The abstract states that the framework 'effectively eliminates the meshing step,' but the pipeline still requires generating and adaptively refining an octree mesh; what is eliminated is the explicit surface triangulation as the geometric input. This distinction is important for assessing the claimed computational advantages, and the wording should be adjusted so that the claim is not overstated.","section":"Abstract and Section 4.3"}],"minor_comments":[{"comment":"The heading contains a typo: 'Dirihclet' should be 'Dirichlet.'","section":"Section 4.1"},{"comment":"The text contains a typo: 'exapnsion' should be 'expansion.'","section":"Section 4.2"},{"comment":"The loss function in Eq. (10) uses the ground-truth signed distance s and normal n-hat, but s is not defined immediately before Eq. (10); please state that s denotes the target signed distance from the triangle soup and n-hat denotes the true surface normal.","section":"Section 3, Eq. (10)"},{"comment":"The color bars in Figures 6 and 7 are labeled 'Order of Error' with values from about 4.5 to 8, while the text says the error is on the order of 10^-4; the figures appear to plot -log10(error) or a similar transformed quantity. Please relabel the color bars and clarify the plotted quantity.","section":"Section 5.1, Figures 6 and 7"},{"comment":"The wall-time comparison in Figure 11 does not specify whether the octree refinement levels are kept fixed as the number of surface triangles is increased; without this detail, the comparison may conflate surface-mesh refinement with volume-mesh refinement. Please state the fixed setup or explain how the triangle count is varied.","section":"Section 5.2.2, Figure 11"},{"comment":"The statement that the gradient is computed numerically 'by using two stencils on each axis using the central difference method' is vague; please specify the stencil size and how it relates to the mesh size h.","section":"Algorithm 4"}],"recommendation":"major_revision","confidential_remarks":"The paper combines two established ingredients, INR-based signed distance fields and the shifted boundary method, in a straightforward way that is nevertheless useful and within the journal's scope. The missing INR-in-the-loop convergence study and the sign inconsistency are fixable with additional work, so I would not reject on novelty or scope grounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The core contribution is real: the paper couples INR-based SDF geometry with the Shifted Boundary Method for linear elasticity, and it designs a training loss (clamped distance plus eikonal plus normal similarity) and a hybrid sampler specifically to make the INR's distance vectors reliable in the narrow band SBM needs. That integration is new in the literature as far as I can tell. The octree-based surrogate boundary identification and the wall-time scaling comparison (Figure 11) are also useful, and the authors are honest about using triangle soups as ground truth when other modalities are unreliable.\n\nThe problem is that the central validation is incomplete. The ring convergence study (Section 5.2.1) exercises SBM alone, not the INR-in-the-loop pipeline. The only full-loop quantitative test is one icosphere at a fixed refinement level, with no convergence study and no error bars. Table 5 shows the distance-vector cosine similarity drops to 0.980 with SD 0.13 for the turbine, and misalignment concentrates at sharp curvature. Since SBM's shifted Taylor condition (Eq. 20) requires d to be the exact closest-point vector with ||d||=O(h), an INR that is off by more than that in a boundary layer could spoil convergence. I don't think the paper has yet shown otherwise.\n\nMore concretely, there's a sign inconsistency. Eq. 8 defines f>0 inside; Eq. 9 then gives d = -f ∇f, which is correct under that convention. But Algorithm 3 treats f<0 as interior, and Algorithm 4 computes d = (∇f/||∇f||) f without the minus sign. Under either convention, one of these formulas points the distance vector the wrong way. That's load-bearing, not cosmetic, because the direction of d enters the shifted boundary condition directly. It might be a typo, but as written the paper is internally contradictory.\n\nAlso, the abstract says 'effectively eliminating the meshing step.' What the method eliminates is surface-mesh preprocessing, not meshing: an adaptive octree background grid is still constructed. That overstatement should be toned down.\n\nThe complex-geometry demonstrations (bunny, Eiffel Tower, gyroid) are qualitative. They show the pipeline runs, not that it runs accurately.\n\nWho's this for: people working on INR-based simulation, immersed boundary methods, and mesh-free workflows. The integration idea is worth their attention. I'd send it to peer review because the contribution is potentially significant, but I would expect a major revision that fixes the sign, adds an INR-in-the-loop convergence study, and softens the mesh-elimination claim.","headline":"Promising INR–SBM integration that lacks a full-loop convergence study and carries a load-bearing sign inconsistency in the distance-vector formula.","tokens_in":18546,"tokens_out":4918,"would_cite":false,"duration_ms":55164,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Finite element analysis can run directly on neural signed-distance-field geometry, using the Shifted Boundary Method to skip the explicit mesh and keep preprocessing time independent of triangle count.","keywords":["Implicit Neural Representations","Signed Distance Fields","Shifted Boundary Method","Linear Elasticity","Mesh-free simulation","Octree meshes","Finite element analysis","Complex geometry"],"falsifier":"Take a high-curvature geometry such as the tetrakis or bunny, compute exact closest-point distance vectors at the integration points on the surrogate boundary, and compare against the network's gradient-derived vectors; if the cosine-similarity error grows beyond a few percent at sharp edges, the shifted boundary condition will misplace the displacement data there. A complementary test: run a mesh-convergence study with the INR in the loop—if the $L^2$ error stops decreasing at the level of the INR's distance-vector error (around $10^{-4}$), then the neural representation, not the SBM discretization, is the accuracy bottleneck.","tokens_in":17393,"feed_emoji":"📐","tokens_out":18217,"duration_ms":170747,"temperature":0.7,"pith_summary":"Finite element analysis usually needs an explicit surface mesh, but this paper proposes to run linear elasticity simulations directly on a neural network that encodes the geometry as a signed distance field. The key move is to pair the neural field with the Shifted Boundary Method (SBM), which applies boundary conditions on a grid-aligned surrogate boundary instead of the true surface. The network supplies the surrogate boundary, the inside/outside classification, and the distance vectors SBM requires, so no triangle soup or boundary-fitted mesh is needed. The authors demonstrate the workflow on a 2D ring benchmark, a 3D icosphere, and complex shapes from point clouds and triangle soups, reporting that meshing and assembly time no longer grow with the number of surface triangles.","feed_headline":"Neural geometry removes the mesh from elastic simulation","feed_subtitle":"Combining neural signed-distance fields with the shifted boundary method makes meshing and assembly time independent of triangle count.","key_machinery":"The central object is the neural signed distance field $f_\\theta$, trained so that its gradient has unit norm near the boundary. SBM then uses the distance vector $d$ in a first-order shifted boundary condition on a surrogate octree boundary, enforced weakly through a consistency plus penalty term. The octree is built by querying the network for inside/outside classification of integration points, and distance vectors at the surrogate boundary are obtained by numerical gradient of the network, with a cache so each query is evaluated once. The same network therefore replaces both the mesh traversal and the closest-point projection that triangle-soup SBM would otherwise perform.","core_discovery":"The central claim is that a neural signed distance field, trained with hybrid sampling near the boundary and a clamped, eikonal-regularized loss, can act as the geometric oracle for SBM end to end. For any point on the surrogate boundary, the network gives the signed distance, the unit gradient, and hence the distance vector, which is inserted into the shifted boundary condition that moves the displacement data from the true boundary to the surrogate boundary. The distance vector is taken as $d = -f_\\theta(x)\\nabla_x f_\\theta(x)$, relying on the eikonal condition $\\|\\nabla_x f_\\theta(x)\\|=1$ near the surface. On the 2D ring with an analytical solution, the method recovers second-order convergence; on an icosphere, the surface-integrated displacement error against the triangle-soup ground truth is about $2.99\\times 10^{-4}$. Wall-time experiments show constant meshing and assembly time for the INR as the triangle count grows, whereas the mesh-based pipeline scales with the number of triangles.","pith_inferences":["An unstated consequence of the wall-time plot is that the INR advantage depends on the network size and hardware: for very small triangle counts, the fixed cost of the neural network forward pass could exceed the cost of traversing a small triangle soup.","The paper's own error maps show the largest distance-vector errors at sharp curvature, which suggests that CAD-style geometry with edges and corners will need curvature-aware training or adaptive sampling before the method reaches engineering accuracy there.","The distance-vector definition in Eq. (9) and the sign used in Algorithm 4 are opposite; an implementation must settle the sign convention consistently, otherwise the shifted boundary condition would push the displacement data in the wrong direction.","A direct testable extension is to repeat the ring convergence study with the INR supplying the boundary: the convergence slope should flatten at the mesh size where the neural distance-vector error starts to dominate, giving a practical accuracy ceiling for the approach."],"forward_implications":["Meshing and assembly wall times become independent of the number of surface triangles, so very large triangle soups can be analyzed at constant preprocessing cost.","Geometries from point clouds, multi-view images, or generative models can be simulated without converting them to a boundary-fitted mesh.","Adaptive octree refinement near boundaries can be driven directly by neural queries, automatically resolving fine geometric features.","The framework currently applies to linear elasticity and linear PDEs; the authors propose extending it to large-deformation and multi-physics problems."],"supporting_citations":[{"why":"Defines the shifted boundary method for embedded-domain boundary-value problems, providing the surrogate-boundary treatment that the paper adopts.","marker":"[7]"},{"why":"Extends SBM to advection-diffusion and incompressible flow, establishing the boundary-condition shifting machinery used here.","marker":"[8]"},{"why":"Develops the shifted boundary method for solid mechanics, supplying the linear-elasticity formulation and the traction boundary treatment.","marker":"[17]"},{"why":"Gives the optimal surrogate boundary selection and distance-vector computation on octree meshes that the neural field is meant to replace.","marker":"[18]"},{"why":"Introduces the implicit neural representation architecture used as the INR backbone for the signed distance fields.","marker":"[3]"},{"why":"Provides the geometric (eikonal) regularization loss that trains the neural field to behave like a signed distance function near the surface.","marker":"[4]"},{"why":"Supplies the scalable adaptive octree meshing strategy used to construct the surrogate grid around the INR.","marker":"[20]"},{"why":"Provides the ring benchmark with an analytical elasticity solution used for the convergence validation.","marker":"[32]"}],"fun_headline_variants":["Neural fields remove mesh from elasticity simulation","Meshless simulation via neural implicit representations","Complex geometry without meshing using neural nets","Neural signed distance fields cut out the mesh","No mesh needed for complex elasticity simulations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands or falls on whether the learned neural field behaves like a true signed distance function very close to the true boundary, so that its gradient—not just its value—points along the shortest path to the surface and gives the correct distance vector for the shifted boundary condition.","fun_headline_variants_meta":{"raw":{"variants":["Neural fields remove mesh from elasticity simulation","Meshless simulation via neural implicit representations","Complex geometry without meshing using neural nets","Neural signed distance fields cut out the mesh","No mesh needed for complex elasticity simulations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000894,"raw_usage":{"total_tokens":3831,"prompt_tokens":903,"completion_tokens":2928,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2863}},"tokens_in":519,"tokens_out":2928,"duration_ms":26280,"temperature":1.0,"reasoning_tokens":2863,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:18:33.952160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a high-curvature geometry such as the tetrakis or bunny, compute exact closest-point distance vectors at the integration points on the surrogate boundary, and compare against the network's gradient-derived vectors; if the cosine-similarity error grows beyond a few percent at sharp edges, the shifted boundary condition will misplace the displacement data there. A complementary test: run a mesh-convergence study with the INR in the loop—if the $L^2$ error stops decreasing at the level of the INR's distance-vector error (around $10^{-4}$), then the neural representation, not the SBM discretization, is the accuracy bottleneck.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends SBM to advection-diffusion and incompressible flow, establishing the boundary-condition shifting machinery used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Develops the shifted boundary method for solid mechanics, supplying the linear-elasticity formulation and the traction boundary treatment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the optimal surrogate boundary selection and distance-vector computation on octree meshes that the neural field is meant to replace."},{"cited_title":"Saurabh, M","cited_arxiv_id":null,"evidence_quote":"Supplies the scalable adaptive octree meshing strategy used to construct the surrogate grid around the INR."},{"cited_title":"Schillinger, M","cited_arxiv_id":null,"evidence_quote":"Provides the ring benchmark with an analytical elasticity solution used for the convergence validation."}],"review_version":1}