{"id":"5e387380-d580-44ad-b2ef-e099cb83ad4f","arxiv_id":"2412.07126","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network trained with finite-element residual losses and a new stencil convolution can act as a surrogate for 2D linear elasticity problems on unseen block geometries.","lead":"This paper trains neural networks to solve elasticity problems by using the finite element method's own equations as the training signal. The new 'stencil convolution' lets the network work directly on finite element meshes, and the authors show that the networks can predict solutions for block shapes they were not trained on.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fig. 8's 'best performing' selection makes the systematic-decrease claim potentially an artifact of model selection; no per-combination or per-seed distributions are reported.","rationale":"The reader's weakest assumption concerns the handling of stencil points outside the FE domain. That is a real and important implementation gap, but it is less directly load-bearing for the paper's central claim than the selection-bias problem: even if out-of-domain stencil points were handled perfectly, the abstract's assertion that testing errors 'systematically decrease' with more training geometries would still rest on Fig. 8, which appears to select the best-performing model at each training-set size. The FE-PINN method itself may be sound, and the provided code could resolve the stencil-handling ambiguity, but the quantitative evidence for the generalization trend is not robust as reported. A conditional acceptance requiring a reanalysis with means/variances over all seeds and combinations, plus a fixed-composition ablation, is appropriate. The out-of-domain concern should also be addressed by specifying the fallback rule, but it is secondary to the validity of the central trend claim.","tokens_in":13273,"tokens_out":6522,"duration_ms":73539,"concrete_test":"Recompute Fig. 8 without best-of selection: for each n = 1, 2, 3 and for every geometry combination and every seed, record the testing loss, then plot mean +/- std (and the full distribution) as a function of n. If the mean over all C(7,n) combinations does not decrease monotonically, or if the n=3 best value lies within the n=2 distribution, the systematic-decrease claim is unsupported. A complementary controlled check: fix the best triple (e.g., 0, 20, 50 degrees for wedges), train on every subset of that triple, and verify that testing losses decrease as the subset grows from size 1 to size 3.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative evidence for the abstract's claim that testing errors 'systematically decrease' with more training geometries is Fig. 8, which plots losses from 'best performing FE-PINNs' versus the number of training geometries. The paper states in Sec. 2.3 that, for each training scenario, only the best of three random seeds is kept. For n=2 and n=3, Fig. 8 further selects the best among all geometry combinations: 21 and 35 combinations, respectively, for the wedged-block family (and analogous for holes). Because the test set is fixed, the minimum over a growing pool of training scenarios decreases even if the model gains no generalization ability: the n=1 minimum is over only 7 single-geometry models, while the n=3 minimum is over 35 triple-geometry models. Thus the monotone decrease in Fig. 8 may be a selection artifact rather than evidence that FE-PINNs 'learn how to generalize.' The paper never reports the mean, median, or spread of testing losses across seeds or across all combinations of a given size, so a reader cannot tell whether a typical FE-PINN improves or only the lucky best one does. This directly undercuts the headline generalization claim, independent of the separate out-of-domain stencil-handling gap.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces FE-PINNs, a physics-informed neural network framework in which the finite-element weak-form residual is used as the training loss and a custom 'stencil convolution' operator performs convolutional operations directly on unstructured FE meshes. The method is demonstrated on two families of 2D linear-elastic boundary-value problems (wedged blocks and blocks with holes), training on one, two, or three geometries and testing on unseen geometries. The authors report that testing errors decrease as the number of training geometries grows, and they position the method as a step toward general-purpose surrogate models that leverage existing FE codes.","tokens_in":13496,"tokens_out":4465,"duration_ms":44542,"significance":"If the central claims hold, the stencil-convolution idea is a meaningful contribution: it offers a mesh-adaptive CNN-style operator that works on arbitrary FE discretizations, avoids labeled data by using the FE residual as the physics loss, and enforces boundary conditions through the weak Galerkin form. The paper also ships code and trained networks in a public repository, which strengthens reproducibility. However, the quantitative evidence for the headline generalization claim is weakened by the exclusive use of best-of-three and best-of-combination selections without reporting any distribution, and the stencil-convolution construction is incompletely specified for stencil points outside the domain. Both issues are fixable with additional reporting and targeted experiments, so the work is worth considering after major revision.","major_comments":[{"comment":"The stencil-convolution construction is incomplete for stencil points that lie outside the FE domain. For any non-zero stencil spacing Delta, stencil points associated with nodes near the boundary will fall outside the union of elements; the element search via the GeoPandas sjoin described in Sec. 2.3 then returns no element, and the manuscript states no fallback rule. Since the stencil point values are formed by evaluating shape functions at the containing element, an undefined or inconsistent choice here would corrupt the convolutional features at boundary nodes and could explain the difficulty reported for the 0-degree wedge block in Sec. 4. Please specify the out-of-domain handling (e.g., zero-padding, clamping to the nearest element, or discarding such stencil points) and confirm that all reported results use the same rule.","section":"Sec. 2.2.2, Eqs. (14)-(17)"},{"comment":"The claim in the abstract and Sec. 4 that testing errors systematically decrease with the number of training geometries is based on losses from 'best performing FE-PINNs'. For the wedged-block family, the n=1, n=2, and n=3 minima are selected from 7, 21, and 35 training-set combinations, respectively, with an additional best-of-three-seeds selection per combination. The minimum over a growing pool of models can decrease even if no typical model improves, and the manuscript reports no mean, median, or spread across seeds or across combinations. Please report the distribution of testing losses (e.g., all combinations, or at least median and range) or otherwise demonstrate that the systematic decrease holds beyond the best-case selection.","section":"Sec. 3.2 and Fig. 8"}],"minor_comments":[{"comment":"The word 'implemeneted' should be 'implemented', and 'ansjoin' should be 'an sjoin' (or 'a spatial join').","section":"Sec. 2.3"},{"comment":"The phrase 'the training geometery' should be 'the training geometry'.","section":"Sec. 2.4"},{"comment":"The bias term is written as b_r \\bar\\delta_{mm}, which is not a standard scalar bias; it should presumably be simply b_r. Please correct the notation.","section":"Eq. (15)"},{"comment":"The notation \\bar\\delta_{kk} = \\delta_{kk}/2 is unusual and appears to be a typo; if a Kronecker delta is intended, use the standard symbol and define the coordinate shift accordingly.","section":"Eq. (14)"},{"comment":"The linear regression fits are drawn without reporting slopes, confidence intervals, or p-values; adding these (or replacing the fits with median trends) would make the claimed systematic decrease more quantitative.","section":"Fig. 8"},{"comment":"The phrase 'lowest average loss' is not defined; specify whether the average is over the test geometries and over the three seeds.","section":"Sec. 3.2.1, Fig. 4(b)"},{"comment":"The phrase 'in principal' should be 'in principle'.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"The two main concerns raised by the stress-test are valid and load-bearing: the out-of-domain stencil handling is undocumented, and the Fig. 8 claim is vulnerable to selection effects because only best-performing models are shown. Both are addressable within the manuscript's scope by specifying the fallback rule and by reporting distributions across seeds and geometry combinations. The code and trained-network repository is a genuine asset and should be acknowledged during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The stencil convolution is a real new operator: it lets a CNN run directly on finite-element meshes by evaluating fields at stencil points via the inverse isoparametric map, and the tensor formulation is clean and efficient. That is a contribution worth taking seriously. The second thing is that the headline claim—that testing errors systematically decrease as more training geometries are added—is probably true, but the figure meant to prove it, Fig. 8, is weaker than it looks because it only plots best-performing models.\n\nThe paper does a lot right. The physics loss is the FE residual, so BCs enter through the Galerkin form rather than a penalty term, which is a principled choice. The writing is clear, the method is generalizable in principle, and code plus trained weights are provided. The full results in Figs. 4–7 also support the generalization trend: most three-geometry combinations beat most single-geometry models, and the displacement fields look sensible.\n\nNow the soft spots, in proportion. First, the stress-test note is correct about Fig. 8. The paper keeps only the best of three seeds, and for n=2 and n=3 it appears to select the best among all geometry combinations as well. The minimum over a larger pool of models tends to decrease even if no model actually generalizes better. The paper never reports the mean, median, or spread of losses across seeds or combinations, so the monotone line in Fig. 8 is not by itself evidence of systematic improvement. The broader figures rescue the trend, but the reporting should be fixed.\n\nSecond, there is a reproducibility gap in the stencil definition. Sec. 2.2.2 says stencil points are evaluated using the element containing each point, but it never says what happens when a stencil point falls outside the mesh. The GeoPandas sjoin would find no element; there must be a fallback (probably zero-padding) but it is undocumented. That matters for boundary nodes and for anyone trying to reimplement the method.\n\nThird, there are no baselines. Comparing against PhyGeoNet, a graph-convolution PINN, or even a standard uniformly-gridded CNN on the same test problems would make the practical advantage of stencil convolution concrete. Only two families of 2D linear problems are tested, so the scope is honestly small.\n\nNot a circularity problem: the test metric is the FE residual on held-out geometries, and no parameters are fit to the test set. The missing error bars and selection details are the real issues, not the evaluation protocol itself.\n\nWho is this for? Researchers building mesh-adaptive PINN surrogates for solid mechanics and related fields. It deserves a serious referee; the core operator is novel and the paper is honest about limitations, but it needs a revision that reports distributions, documents the out-of-domain stencil handling, and adds at least one baseline comparison.","headline":"A genuinely new stencil-convolution operator for FE-mesh CNNs, with a plausible generalization story that Fig. 8 weakens by only plotting selected best models.","tokens_in":14069,"tokens_out":2618,"would_cite":true,"duration_ms":30468,"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":"Stencil convolution lets physics-informed neural networks trained on finite-element meshes generalize to unseen geometries, with testing error decreasing as training geometries grow.","keywords":["physics-informed neural networks","finite element method","surrogate modeling","stencil convolution","inverse isoparametric map","convolutional neural networks","unstructured meshes","linear elasticity"],"falsifier":"Inspect the FE-PINN code's element lookup at stencil points: if out-of-domain stencil points are silently skipped or zero-filled, then a test mesh with many nodes near a concave boundary should show a sharp jump in prediction error at those boundary nodes, contradicting the claimed mesh-insensitive generalization.","tokens_in":13018,"feed_emoji":"🧠","tokens_out":8758,"duration_ms":78432,"temperature":0.7,"pith_summary":"This paper introduces FE-PINNs, a way to train physics-informed neural networks as surrogate models for boundary-value problems. Rather than requiring a uniform grid or precomputed solution data, the network is trained on the residual of the finite-element weak form, and a new 'stencil convolution' operation samples fields at stencil points through the inverse isoparametric map of the mesh. The authors show that networks trained on a small set of finite-element geometries predict displacement fields for similar but unseen geometries with reasonable accuracy, and that testing error systematically decreases as the number of training geometries grows. If this holds, FE-PINNs provide a route to surrogate models that inherit the flexibility of existing finite-element codes and need no labeled dataset.","feed_headline":"Finite-element meshes train neural nets to solve unseen shapes","feed_subtitle":"A custom convolution on FE meshes trains surrogates with no labeled data; more training geometries cut testing error.","key_machinery":"The load-bearing object is the stencil convolution tensor $S_{klmn}$, which stores the shape-function values that map a node's field values to the stencil points around it: $S_{klmn} = N^{n,e}(\\xi_{ikl}(x_{ikl}))$ if node $n$ belongs to the element containing stencil point $(k,l)$, and zero otherwise. It is built once per mesh by locating the element for each stencil point via a spatial join and inverting the isoparametric map analytically, then reused as a sparse tensor contraction during training. This operation is what lets the CNN run on irregular meshes: the field is always stored on nodes, and the stencil weights read it through piecewise polynomial shape functions, so the network sees a grid-like receptive field without requiring a grid.","core_discovery":"The central claim is that stencil convolution—defined by a stencil tensor of points around each node, with field values obtained by finite-element shape functions via the inverse isoparametric map—makes convolutional PINNs practical on arbitrary, unstructured finite-element meshes. With this operator, a CNN can be trained by minimizing the norm of the finite-element residual, so boundary conditions enter through the external force vector and need no soft-constraint penalty. The paper reports that networks trained on several geometries generalize to untrained geometries that are similar, and that increasing the number of training geometries systematically reduces testing loss while training loss stays flat. This is taken as evidence that the network learns the underlying operator rather than memorizing isolated solutions.","pith_inferences":["The paper leaves unspecified how stencil points that fall outside the domain are treated; if the element lookup returns nothing, boundary-node features could be distorted, which may explain why the 0-degree wedge block resists prediction from nearby geometries.","A testable extension would be to change element size after training; the authors note discretization invariance is plausible because inputs and outputs are functions, but it has not been demonstrated.","Non-uniform or multi-length-scale stencil arrangements, which the paper explicitly floats as possibilities, could outperform the single uniform spacing used here.","If the observed scaling with training-set size continues, training on hundreds of finite-element models should yield a high-fidelity operator surrogate; the authors mention this as a large-scale target but do not test it."],"forward_implications":["A FE-PINN trained on a handful of geometries can replace a full finite-element solve for new, similar geometries, giving near-instant predictions with physics-based accuracy.","No labeled solution data is required: the physics loss is the finite-element residual, and boundary conditions enter automatically through the external force vector.","Testing error decreases systematically as the training set grows, indicating that the network learns a generalizable operator rather than memorizing training cases.","The same machinery is expected to extend to nonlinear and three-dimensional problems, and to variable boundary conditions, material properties, and body forces by adding input channels."],"supporting_citations":[{"why":"Introduces physics-informed neural networks and the residual-loss training paradigm that FE-PINN builds on.","marker":"[4]"},{"why":"PhyGeoNet, the geometry-adaptive CNN that motivates FE-PINN's architecture and whose uniform-reference-domain mapping FE-PINN avoids.","marker":"[18]"},{"why":"Physics-informed graph neural Galerkin networks, the graph-convolution alternative that stencil convolution is designed to improve upon.","marker":"[21]"},{"why":"Neural operators, which provides the function-space view that FE-PINN inputs and outputs are piecewise functions.","marker":"[2]"},{"why":"Hughes's finite element text, which supplies the weak form, shape functions, and isoparametric mapping used to define stencil convolution.","marker":"[22]"},{"why":"Analytical inverse transformation for quadrilateral isoparametric elements, used to evaluate stencil point coordinates.","marker":"[24]"}],"fun_headline_variants":["Stencil convolution makes PINNs work on any mesh","FE residual loss trains surrogates without labels","Neural nets learn FE physics on arbitrary meshes","CNN surrogate: train on meshes, test on new shapes","Stencil conv nets: FE training, zero labels, new shapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that stencil points that fall outside the finite-element domain near boundaries can be handled without corrupting the convolutional features, yet the paper never specifies what happens when the element search returns no element for such a point.","fun_headline_variants_meta":{"raw":{"variants":["Stencil convolution makes PINNs work on any mesh","FE residual loss trains surrogates without labels","Neural nets learn FE physics on arbitrary meshes","CNN surrogate: train on meshes, test on new shapes","Stencil conv nets: FE training, zero labels, new shapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2618,"prompt_tokens":821,"completion_tokens":1797,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":1717}},"tokens_in":437,"tokens_out":1797,"duration_ms":13042,"temperature":1.0,"reasoning_tokens":1717,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:05:27.456428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the FE-PINN code's element lookup at stencil points: if out-of-domain stencil points are silently skipped or zero-filled, then a test mesh with many nodes near a concave boundary should show a sharp jump in prediction error at those boundary nodes, contradicting the claimed mesh-insensitive generalization.","supporting_citations":[{"cited_title":"Kovachki, Z","cited_arxiv_id":null,"evidence_quote":"Neural operators, which provides the function-space view that FE-PINN inputs and outputs are piecewise functions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Hughes's finite element text, which supplies the weak form, shape functions, and isoparametric mapping used to define stencil convolution."},{"cited_title":"Hua, An inverse transformation for quadrilateral isoparametric elements: Analysis and application, Finite Elements in Analysis and Design 7 (2) (1990) 159–166","cited_arxiv_id":null,"evidence_quote":"Analytical inverse transformation for quadrilateral isoparametric elements, used to evaluate stencil point coordinates."}],"review_version":1}