{"id":"5e424ea4-6ff5-4108-bd28-d688ba8fdb7f","arxiv_id":"2501.07765","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"PINN-FEM enforces Dirichlet boundary conditions in PINNs by blending neural network fields with finite element shape functions in a boundary layer, but the 2D extension is ambiguous and the experimental comparisons are confounded.","lead":"This paper proposes PINN-FEM, a hybrid method that uses finite element shape functions near boundaries to exactly enforce Dirichlet conditions in physics-informed neural networks. The approach is tested on six 2D linear elasticity problems and claims to beat soft and distance-function baselines, but the paper's formulation has inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported superiority is confounded: all PINN-FEM runs use LBFGS while all baselines use Adam, so Table 2 cannot establish that the FEM-layer boundary enforcement, rather than the optimizer, drives the accuracy gains.","rationale":"The reader's verdict of REJECT is well-supported, but my primary concern differs from the reader's weakest_assumption. The reader focuses on the 2D ansatz in Eq. (33) being discontinuous/underspecified; that is a real defect in the write-up and would prevent independent verification. I instead put the weight on the optimizer confound, because the paper's headline is an empirical superiority claim and the experiments systematically vary two variables at once (method and optimizer). A concrete matched-optimizer replication can settle whether the accuracy gain is due to the FEM-layer boundary enforcement or merely to LBFGS. If the gain survives matched optimizers, the method might still be sound despite the presentation issues; if it does not, the central claim collapses regardless of Eq. (33). Since the reader already rejected for related reasons, my read does not change the verdict; it sharpens the primary evidence for rejection.","tokens_in":20319,"tokens_out":9181,"duration_ms":98132,"concrete_test":"Retrain all models on Experiment 1 (square plate with constant stress) under a matched 2x2 design: PINN-FEM and Soft, ADF, and DF baselines, each with both LBFGS and Adam, using identical network architecture, collocation mesh, random seeds, and energy-based loss (plus the same boundary penalty for Soft). Then compare Table 2's relative errors across optimizer-matched columns. If the order-of-magnitude advantages disappear or reverse, the reported superiority is an optimizer artifact and the central claim is unsupported; if PINN-FEM retains the advantage under both optimizers, the confound is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PINN-FEM outperforms standard PINN baselines. To support that claim, the experiments must isolate the boundary-treatment mechanism from other training choices. Section 4.3.1 states that the PINN-FEM model is trained with LBFGS, whereas all three baselines (Soft, ADF, DF) are trained with Adam because Adam 'is found to perform best for these models'; the same assignment is repeated in every experiment. Table 2 therefore varies the optimizer and the BC-enforcement strategy simultaneously. Large reported gaps (e.g., e(uy)=0.0003 for PINN-FEM vs 0.56 for Soft on the cantilever; e(ux)=0.05 vs 0.22 on the crack plate) may reflect the well-known superiority of LBFGS on small over-parameterized energy problems rather than the proposed FEM layer. A secondary unclarity is Eq. (33): the field ND and the role of g in Omega_FE are not specified so that one can verify ND=0 on Gamma_g, ND=1 on Gamma_I, and continuity with N_theta; as written the 2D trial space is not a well-defined H1 subset. But the optimizer confound alone is sufficient to invalidate the empirical support for the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PINN-FEM, a hybrid method that enforces Dirichlet boundary conditions exactly by decomposing the computational domain into a finite element layer adjacent to the boundary and a neural network interior, and trains the network by minimizing a potential-energy loss. The method is described for 1D, 2D, and 3D, and evaluated on six 2D linear elasticity problems against soft-boundary (Soft), approximate-distance-function (ADF), and distance-function (DF) PINN baselines. The authors report that PINN-FEM consistently outperforms these baselines in relative ℓ1 displacement error, and they highlight robustness on problems with discontinuous, point, and crack-like boundary conditions.","tokens_in":20602,"tokens_out":6849,"duration_ms":63184,"significance":"If substantiated, PINN-FEM would provide a practical mechanism for strong Dirichlet enforcement in energy-based PINNs, including for boundary conditions that distance-function methods cannot represent. The paper draws on established blending ideas from mesh-free methods and on the deep energy method, so the conceptual novelty is moderate; the main value would be a demonstrated, reproducible robustness advantage across several boundary geometries. However, the current experimental design changes the optimizer together with the boundary-enforcement method, and the 2D trial-space definition is incomplete, so the reported superiority is not yet established. The method's energy-based loss and exact boundary handling are attractive, but the evidence as presented is insufficient for a strong claim.","major_comments":[{"comment":"Section 4.3.1 states that PINN-FEM is trained with LBFGS while all baselines (Soft, ADF, DF) are trained with Adam, and this same optimizer assignment is reused in every subsequent experiment (Sections 4.3.2–4.3.6). Table 2 therefore varies two factors simultaneously: the boundary-condition strategy and the optimizer. Because LBFGS is typically much more effective than Adam on small over-parameterized energy-minimization problems, the large reported gaps (for example, e(uy)=0.0003 for PINN-FEM versus 0.56 for Soft in the cantilever experiment) cannot be attributed to the FEM-layer boundary enforcement. A controlled comparison with identical optimizers, or at least a cross of optimizer and method, is required to support the paper's central claim.","section":"§4.3.1 / Table 2"},{"comment":"The discretized 1D energy in Eqs. (16)–(20) omits the body-force term −∫ f u dx that is present in the continuous energy of Eq. (13), and the traction work is written as hNθ,x|0 instead of hNθ(0). As a result, minimizing Eq. (21) does not correspond to minimizing the potential energy of the boundary value problem (10)–(12). The same body-force omission appears in the both-ends version, Eqs. (27)–(32). These equations need to be corrected and the subsequent energy balance re-derived; as written, the 1D formulation is not a valid variational statement of the model problem.","section":"§3.2.1, Eqs. (16)–(20)"},{"comment":"The 2D trial field u = Nθ(x)ND(x) + g in ΩFE is not a well-defined finite element interpolation as written. If ND is a scalar blending function equal to 1 on the interface ΓI and 0 on Γg, then on ΓI the field becomes Nθ + g, which is discontinuous with the interior field Nθ unless g = 0. If ND instead denotes a vector of element shape functions, the product form does not match standard nodal interpolation, and the extension of the prescribed boundary data g into the interior of ΩFE is not specified. The paper must define ND precisely (for example, as a piecewise-linear function with ND=1 on ΓI and ND=0 on Γg, or by replacing the ansatz with u = Σ_i Nθ(x_i)N_i(x) + Σ_{j∈Γg} g_jN_j(x)) and demonstrate that the resulting trial space is a subset of H^1. Without this, the 2D energy minimization is performed over an ill-defined trial space.","section":"§3.3, Eq. (33)"},{"comment":"The energy in the FEM region is evaluated by one-point quadrature at a single centroid over the whole ΩFE region. Since ΩFE contains many triangular elements and the strain field varies between elements, this is not the standard per-element constant-strain quadrature used in linear FEM. The approximation error introduced by collapsing the entire FE layer to one centroid should be quantified, or the energy should be assembled element-wise; otherwise the reported accuracy of the 2D method may reflect this additional quadrature approximation rather than the proposed hybrid construction.","section":"§3.3, Eq. (37)"}],"minor_comments":[{"comment":"Section 4.1 says the collocation points are the centroids of the triangles, while Sections 4.3.1 and 4.3.4 say the collocation points are the nodal points of the mesh; please clarify which is actually used.","section":"§4.1 / §4.3.1"},{"comment":"Eq. (17) contains the typo 'gR2 NR2,x' and Eq. (28) contains 'gL2' and 'gR2' where the prescribed boundary values g, gL, and gR should appear; these likely are transcription errors from the shape-function indices.","section":"§3.2.1, Eq. (17); §3.2.2, Eq. (28)"},{"comment":"The soft-boundary weight β is introduced but its value and tuning procedure are never reported; please state the value used for all soft baseline runs.","section":"§4.2, Eq. (47)"},{"comment":"Several rows in Table 2 are missing entries for ADF and DF baselines; the text explains the reasons, but a footnote in the table itself would make the comparison clearer at a glance.","section":"Table 2"},{"comment":"Reference [46] should be 'Timoshenko and Goodier' rather than 'Goodyear'; please correct.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early preprint. The optimizer confound is elementary and should have been avoided; a revision will require substantially new experiments with controlled optimizers. The novelty over the cited blending literature is incremental, and the authors should more sharply distinguish their construction from [31,32,36]. The 1D energy errors and the ambiguity of Eq. (33) are correctable in principle, but the empirical support for the headline claim is currently invalid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my read. The paper's headline—PINN-FEM outperforms standard PINNs—is not supported by the experiments as written, because every PINN-FEM run uses LBFGS while every baseline uses Adam (Sec. 4.3.1). That confound is enough to invalidate Table 2 as evidence for the boundary-enforcement mechanism. The method itself is a practical special case of Sukumar & Srivastava (2022): the scalar distance function φ is replaced by an FE shape-function layer, which the paper acknowledges in Sec. 3.1. So the novelty is real but modest—an engineering trick for complex or discontinuous boundaries, not a new idea.\n\nWhat the paper does well: the 1D derivation in Sec. 3.2 works out interface continuity explicitly, the six experiments span a good range of 2D elasticity problems, and the energy-based loss is appropriate. The FE-layer approach has genuine appeal for point/discontinuous BCs, where global distance functions are hard to define.\n\nThe soft spots are substantial. First, the optimizer confound: with LBFGS versus Adam, the large reported gains (e.g., 0.0003 vs 0.56 on the cantilever) could easily be optimizer effects. Second, the 1D energy loses the body-force term between Eq. (13) and Eq. (16), and the traction term in Eq. (20) is written as h Nθ,x(0) instead of h Nθ(0). The experiments have zero body force, so the numbers may not be affected, but the formulation as stated is wrong. Third, the 2D ansatz in Eq. (33), u = Nθ(x) ND(x) + g, is not well-defined: ND is never given explicitly, and as written the field jumps by g at the interface if ND=1 there. It needs to be defined as a proper FE interpolation with ND=0 on Γ_g and 1 on Γ_I, or as a sum of nodal shape functions. Fourth, no code, seeds, or error bars, so the results are not reproducible.\n\nVerdict: the paper deserves a serious referee because the idea is practical and the flaws are fixable, but in current form the central empirical claim is unproven. I would send it to review with a request for major revision: rerun baselines with the same optimizer, fix the energy terms, define the 2D ansatz precisely, and release code.","headline":"Confounded optimizer comparison and ambiguous 2D ansatz undercut an otherwise practical FE-layer twist on distance-function BC enforcement.","tokens_in":21106,"tokens_out":7110,"would_cite":false,"duration_ms":66677,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","68T07","74S05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Hybrid PINN-FEM enforces Dirichlet boundary conditions exactly with a finite-element boundary layer and a neural-network interior.","keywords":["physics-informed neural networks","finite element method","Dirichlet boundary conditions","domain decomposition","energy-based loss","linear elasticity","boundary condition enforcement"],"falsifier":"Compute the value of the hybrid trial field on both sides of the interface $\\Gamma_I$ for a problem with nonzero prescribed displacement $g$ and compare: if $u|_{\\Omega_{FE}} - u|_{\\Omega_{NN}} \\neq 0$ at the interface, the trial space is discontinuous and the reported errors do not correspond to a solution of the original boundary value problem. A direct check would be a two-dimensional plate with constant nonzero $g$ and a single-element FEM strip, evaluating the field at $\\Gamma_I$ and at $\\Gamma_g$.","tokens_in":20097,"feed_emoji":"🧮","tokens_out":7092,"duration_ms":66574,"temperature":0.7,"pith_summary":"The paper proposes a hybrid method, PINN-FEM, for solving boundary value problems in solid mechanics: a finite element strip is placed along Dirichlet boundaries and a neural network approximates the solution in the interior, with the two fields meeting at an interface. The trial field is trained by minimizing total potential energy, so the Dirichlet data enter as exact constraints through FEM shape functions instead of as penalty terms. The authors claim this gives exact boundary condition enforcement without the convergence problems of soft PINN losses, and they support it with six linear elasticity experiments of increasing complexity (smooth, holed, discontinuous, point, cracked, and cantilever domains). Across those experiments the relative $\\ell^1$ displacement errors are consistently lower than standard PINN baselines, including PINNs with approximate distance functions.","feed_headline":"FEM layer at the boundary gives PINNs exact Dirichlet enforcement","feed_subtitle":"Six elasticity tests show lower displacement error than soft or distance-function PINN baselines.","key_machinery":"The central object is the decomposed trial space: the domain $\\Omega$ is split into an FEM region $\\Omega_{FE}$ adjacent to the Dirichlet boundary and a neural network region $\\Omega_{NN}$, with interface $\\Gamma_I$. In the FEM strip, linear shape functions carry the prescribed boundary value $g$ into the trial field, so the Dirichlet data are enforced strongly; in the interior, the neural network $N_\\theta(x)$ is free. The carrying identity is the energy functional $E(u)=\\int_\\Omega (\\tfrac12 \\varepsilon : C : \\varepsilon - f\\cdot u)\\,d\\Omega - \\int_{\\Gamma_h} h\\cdot u\\,d\\Gamma$, minimized over the hybrid field, which automatically absorbs Neumann conditions and makes the loss a scalar potential energy rather than a residual sum.","core_discovery":"On the paper's own terms, the central discovery is that domain decomposition with finite elements at the boundary converts the Dirichlet condition from a soft loss term into a strong constraint, while the interior stays mesh-free and flexible. The hybrid displacement field is $u=N_\\theta(x)N_D(x)+g$ in the FEM strip and $u=N_\\theta(x)$ in the neural network region, with an energy-based loss derived from the principle of minimum potential energy. This is claimed to be exact at the Dirichlet boundary, automatically consistent with Neumann conditions through the variational formulation, and general enough to handle discontinuous and point boundary conditions where global distance-function constructions fail. The reported numbers show improvements such as a reduction in relative error for the crack problem from $0.07$ (ADF) to $0.05$ for $u_x$, and for the cantilever beam from $0.56$ (soft) to $0.0009$ for $u_x$.","pith_inferences":["A natural extension the authors leave implicit: the strip construction could be applied to time-dependent boundary data by making the FE layer track a moving Dirichlet boundary or by using space-time shape functions.","The same energy-loss argument suggests the method should extend to finite-deformation hyperelasticity, where a potential energy is still well defined, as long as the interface continuity condition is built into the ansatz.","One testable extension: using full quadrature in the FEM strip rather than a single Gauss point should improve accuracy where strain varies strongly inside the strip.","The hybrid ansatz could also serve as a mesh-conforming initializer for standard FEM solvers, since it supplies a field that already matches the boundary-layer geometry."],"forward_implications":["The Dirichlet condition becomes a strong constraint built into the trial field, so no boundary-loss weight $\\beta$ needs tuning in the energy loss.","Neumann conditions are absorbed by the variational formulation, so the loss function needs no separate traction residual.","Discontinuous and point Dirichlet data, which break global distance-function constructions, still fit inside a conforming local FEM mesh.","On smooth problems the method can reach near-analytical accuracy, as in the cantilever case with relative errors around $10^{-4}$."],"supporting_citations":[{"why":"Supplies the deep energy method whose potential-energy loss the paper adopts for training.","marker":"[27]"},{"why":"Provides the benchmark elastic plate problem and the energy-based comparison used in the first experiment.","marker":"[25]"},{"why":"First coupled finite-element/mesh-free formulation that motivates placing a finite element layer next to a mesh-free interior.","marker":"[31]"},{"why":"Presents a continuous blending formulation for exact essential boundary condition imposition in coupled methods.","marker":"[32]"},{"why":"Shows enforcement of essential boundary conditions in meshless approximations using finite elements, a direct precedent for the strip construction.","marker":"[36]"},{"why":"Prior hybrid FEM-NN model combining finite elements with neural networks, which the paper builds on.","marker":"[35]"},{"why":"Provides the approximate distance function method used as the ADF baseline and the contrast case for discontinuous boundaries.","marker":"[44]"},{"why":"Mesh generator used to create the finite element mesh and the collocation points for the experiments.","marker":"[45]"},{"why":"Supplies the analytical cantilever stress and displacement solution used as ground truth for the final experiment.","marker":"[46]"}],"fun_headline_variants":["PINN-FEM hybrid enforces Dirichlet exactly, not softly","FEM at boundary gives PINNs strong Dirichlet control","Hybrid PINN-FEM: exact Dirichlet, mesh-free interior","Domain decomposition makes PINN boundary conditions exact","PINN-FEM: finite elements lock in Dirichlet conditions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The two-dimensional trial field $u=N_\\theta N_D+g$ presupposes a shape-function field $N_D$ that is $0$ on the Dirichlet boundary and $1$ at the interface with the neural network region; the paper does not define such a field for a multi-element strip or verify continuity of the hybrid field at the interface, so if this construction fails the energy is minimized over discontinuous trial functions.","fun_headline_variants_meta":{"raw":{"variants":["PINN-FEM hybrid enforces Dirichlet exactly, not softly","FEM at boundary gives PINNs strong Dirichlet control","Hybrid PINN-FEM: exact Dirichlet, mesh-free interior","Domain decomposition makes PINN boundary conditions exact","PINN-FEM: finite elements lock in Dirichlet conditions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1294,"prompt_tokens":888,"completion_tokens":406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":504,"tokens_out":406,"duration_ms":4187,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:36:19.245072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the value of the hybrid trial field on both sides of the interface $\\Gamma_I$ for a problem with nonzero prescribed displacement $g$ and compare: if $u|_{\\Omega_{FE}} - u|_{\\Omega_{NN}} \\neq 0$ at the interface, the trial space is discontinuous and the reported errors do not correspond to a solution of the original boundary value problem. A direct check would be a two-dimensional plate with constant nonzero $g$ and a single-element FEM strip, evaluating the field at $\\Gamma_I$ and at $\\Gamma_g$.","supporting_citations":[{"cited_title":"Anenergyapproachtothesolutionofpartialdifferentialequations incomputationalmechanicsviamachinelearning: Concepts,implementationandapplications","cited_arxiv_id":null,"evidence_quote":"Supplies the deep energy method whose potential-energy loss the paper adopts for training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the benchmark elastic plate problem and the energy-based comparison used in the first experiment."},{"cited_title":"A coupled finite element-element-free galerkin method","cited_arxiv_id":null,"evidence_quote":"First coupled finite-element/mesh-free formulation that motivates placing a finite element layer next to a mesh-free interior."},{"cited_title":"Enrichment and coupling of the finite element and meshless methods","cited_arxiv_id":null,"evidence_quote":"Presents a continuous blending formulation for exact essential boundary condition imposition in coupled methods."},{"cited_title":"Enforcement of essential boundary conditions in meshless approximations using finite elements.Computer Methods in Applied Mechanics and Engineering, 131(1-2):133–145, 1996","cited_arxiv_id":null,"evidence_quote":"Shows enforcement of essential boundary conditions in meshless approximations using finite elements, a direct precedent for the strip construction."},{"cited_title":"Hybrid fem-nn models: Combining artificial neural networks with the finite element method.Journal of Computational Physics, 446:110651, 2021","cited_arxiv_id":null,"evidence_quote":"Prior hybrid FEM-NN model combining finite elements with neural networks, which the paper builds on."},{"cited_title":"Exact imposition of boundary conditions with distance functions in physics- informed deep neural networks.Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the approximate distance function method used as the ADF baseline and the contrast case for discontinuous boundaries."},{"cited_title":"A three-dimensional finite element mesh generator with built-in pre-and post-processing facilities.International Journal for Numerical Methods in Engineering, 11:79, 2020","cited_arxiv_id":null,"evidence_quote":"Mesh generator used to create the finite element mesh and the collocation points for the experiments."},{"cited_title":"Elasticity theory.M: Science, 1975","cited_arxiv_id":null,"evidence_quote":"Supplies the analytical cantilever stress and displacement solution used as ground truth for the final experiment."}],"review_version":1}