{"id":"e6f7003d-8c8e-4da3-94fe-55f968e80b95","arxiv_id":"1908.10407","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Using physical energy as the loss function lets a deep neural network solve a range of computational mechanics PDEs, from linear elasticity to fourth-order plate bending, without meshes or data.","lead":"This paper trains deep neural networks to solve mechanics problems by making the physical energy of the system the loss the network minimizes, instead of fitting data or matching equations point by point. It demonstrates the idea on elasticity, fracture, piezoelectricity, and plate bending, and compares it with standard equation-matching networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The printed loss functions are not consistently the mechanical energies the DEM claim rests on, so the reported benchmark errors may not be for the method as stated.","rationale":"The reader chose quadrature accuracy as the weakest assumption; I agree partially, but the more specific and more load-bearing problem is that the loss formulas printed in the manuscript are internally inconsistent. A quadrature error analysis would not fix Eq. (77), which is not the phase-field energy, or Eq. (28), whose boundary term is not the external work. Under the stress-test rule to flag explicit in-text limitations, Section 6.7.4 itself concedes sampling sensitivity, and the self-referential B parameter in Eq. (70) adds a further circularity in the phase-field initialization. None of this requires doubting the authors' intent; a preprint with unverified code and many typos can easily contain such discrepancies. But because the central claim is an empirical proof-of-concept, the equations defining the loss are load-bearing. The underlying idea, minimizing a physical energy with a DNN, is well motivated and the benchmarks are plausible, so I would not move the verdict to reject; conditional acceptance, pending reproduction and correction of the printed losses, remains the appropriate call.","tokens_in":47767,"tokens_out":14515,"duration_ms":150708,"concrete_test":"Re-implement the two flagship DEM examples literally from the printed equations: the pressurized cylinder with Eqs. (26)-(28) and the 1D phase-field with Eqs. (77)-(78), using the stated architectures (3x30 ReLU^2, 3x50) and stated point counts, and compare with the reported 0.5% and 2.88% L2 errors. Then re-run both with corrected energy forms: boundary work integral(t dot u) dS weighted by boundary length for the cylinder, and integral(phi^2 + l0^2 |grad phi|^2) dx for the phase field. If the literal versions reproduce the reported errors, the concern is moot; if only the corrected versions do, the paper's equations are not the method being validated; if neither does, the benchmark numbers are unreliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the discrete loss minimized by the DNN is a faithful computational proxy for the physical energy. The paper's own equations do not consistently satisfy this. In the 1D phase-field DEM example, Eqs. (77)-(78) define I(phi)=(1/2) integral(phi - l0^2 |grad phi|^2) dx, whereas the phase-field energy used elsewhere (Eqs. (65), (79)) is Gc/(2l0) integral(phi^2 + l0^2 |grad phi|^2) dx plus a history term; the printed 1D functional has the wrong sign on the gradient term, omits phi^2, and its Euler-Lagrange equation is 1 + 2 l0^2 phi'' = 0, not the phase-field equation (72)-(73). The reported 2.88% L2 error therefore does not test the energy method as stated. Eq. (28) is similarly problematic: Lneu uses the area AOmega as the boundary quadrature weight and defines fneu as t_N,x + t_N,y with no displacement factor; for the pressurized cylinder, t_N,x and t_N,y are declared to depend on the unknown displacement u and v (Eq. (26)), so as written Lneu is neither external work nor a traction penalty. If these are typos, the published equations cannot reproduce the claimed 0.5% and 2.88% results; if literal, they solve a different problem. Section 6.7.4's admission that point sampling affects the stability and accuracy of the method further underscores that the discrete loss has not been validated. The underlying energy idea may be sound, but the quantitative support is not yet pinned down.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes and tests the Deep Energy Method (DEM), in which the total potential energy of a mechanical system is used as the loss function for a fully connected deep neural network. The authors argue that, for problems with a variational structure, minimizing the empirical energy (Eq. 12) yields an approximate solution without meshes or labeled data. They present proof-of-concept applications in linear elasticity (pressurized cylinder, plate with a hole, hollow sphere, cube with a spherical hole), elastodynamics via collocation, hyperelasticity, phase-field fracture, piezoelectricity, and Kirchhoff plate bending. Benchmarks with analytical or finite-element reference solutions report relative L2 errors of about 0.5% to 5% for displacements and energies, and the paper contrasts DEM with deep collocation on several examples.","tokens_in":48106,"tokens_out":5258,"duration_ms":57450,"significance":"If the reported results are reproducible, DEM is a valuable extension of the physics-informed neural network literature into computational mechanics, particularly because it replaces residual-based losses with energy minimization and handles natural boundary conditions in a unified way. The paper's breadth—2D/3D linear elasticity, hyperelastic torsion, phase-field fracture, piezoelectricity, and fourth-order plate problems—goes well beyond most existing PINN studies and includes comparisons with analytical and finite-element solutions. The inclusion of code snippets and detailed architecture/optimizer specifications is helpful for reproducibility. However, several printed loss function definitions are not the mechanical energies they claim to be, and the discrete quadrature underlying Eq. (12) is not analyzed. These issues directly affect the quantitative claims and must be corrected before the results can be fully accepted.","major_comments":[{"comment":"The functional being minimized in the one-dimensional phase-field DEM example is not the phase-field energy. As written, I(phi) = (1/2)∫(phi - l0^2 |∇phi|^2) dx, whose Euler-Lagrange equation is 1 + 2 l0^2 phi'' = 0, not the phase-field equilibrium equation in Eqs. (72)-(73) and not the Griffith-type energy in Eqs. (65) or (79). The reported 2.88% L2 error for DEM therefore does not test the energy method as stated, and the comparison with the 70.6% collocation error is not a valid demonstration for this benchmark. The functional must be corrected, or the relation of the printed functional to the phase-field energy must be explained.","section":"6.5.1, Eqs. (77)-(78)"},{"comment":"The Neumann boundary loss Lneu is not the work of external traction: fneu is defined as t_N,x + t_N,y (plus t_N,z in 3D) with no factor of the displacement, and it is weighted by the area/volume of the whole domain rather than by the boundary length/area. In the pressurized cylinder example, Eq. (26) defines t_N,x = P u and t_N,y = P v, so the traction itself depends on the unknown displacement, and the total loss Lint - Lneu is not the total potential energy of the system. As a result, the reported 0.5% and 1.8% displacement errors for the cylinder and plate-with-hole benchmarks are not tied to the minimization problem that the paper claims to solve. These equations need to be rewritten or clarified.","section":"6.2.1, Eq. (28); 6.2.2; 6.2.3, Eqs. (37), (40)"},{"comment":"The paper does not provide a quadrature error analysis or convergence study for the replacement of the energy integral by a weighted sum over sample points. This is load-bearing because if the sum is not an accurate quadrature, the network minimizes a different functional from the physical energy. The paper itself states in Section 6.7.4 that point sampling 'affects the stability and accurate of this deep energy method' and that a particular collocation point distribution was needed for the annular plate. A systematic study of the discrete loss as a function of the number and distribution of sampling points is needed to support the general claims made for DEM.","section":"4.2, Eq. (12); 6.7.4"},{"comment":"The initial history amplitude is defined by B = 1/(1 - phi) for phi < 1, but phi is the unknown phase field that the method is supposed to compute. This makes H(x,0) in Eq. (69) depend on the solution before it is known, which is a circular definition. The authors should state how B is evaluated in practice, for example using an initial guess or a previous iteration, and whether the reported results are sensitive to that choice.","section":"6.5, Eq. (70)"}],"minor_comments":[{"comment":"In the analytical solution for the pressurized cylinder, the expression for vexact is printed as urad cos(theta), which is the same as uexact; it should presumably be urad sin(theta).","section":"6.2.1, Eq. (25)"},{"comment":"The text refers to the linear-elastic minimization problem as 'Eq. (65)', but Eq. (65) is the phase-field energy; the intended reference is likely the problem statement in Eq. (24).","section":"6.2.1, 6.2.2"},{"comment":"In Eq. (28), the boundary term uses A_Omega, the area of the domain, as the quadrature weight; for a line boundary in 2D this should be the length of the boundary (or the appropriate boundary measure), as is done for the volume versus surface terms in Eq. (37).","section":"6.2.1, Eq. (28)"},{"comment":"The sentence 'the proper point sampling inside physical domain affects the stability and accurate of this deep energy method' should read 'affects the stability and accuracy', and the paragraph would benefit from a quantitative description of the sampling strategy used for the annular plate and of its effect on the reported errors.","section":"6.7.4"}],"recommendation":"major_revision","confidential_remarks":"The central energy-minimization idea is plausible and the paper covers an unusually wide range of computational mechanics benchmarks, so I see no reason to reject outright. However, the misprinted phase-field functional and the incorrect Neumann loss definition make several reported numbers unreproducible as written; these are correctable but require re-derivation and possibly re-running the affected examples. The lack of quadrature/convergence analysis is also a substantive gap for a method whose loss is a discrete energy estimate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful application paper built on Deep Ritz, not a new method; treat the phase-field and Neumann-boundary equations as needing correction before the headline numbers can be trusted. What's actually new: the energy-as-loss idea is E et al.'s Deep Ritz method [14], properly cited. The contribution is the breadth of computational mechanics applications: linear elasticity with internal pressure, hyperelastic torsion, phase-field fracture, piezoelectricity, and Kirchhoff plates with an autoencoder and problem-specific sinusoidal activation. That is a genuinely useful proof-of-concept collection, and the DEM-vs-collocation comparison for the 1D phase-field problem is instructive. What it does well: the linear elasticity benchmarks (0.5% for the cylinder, 1.8% for the plate with hole, 5.3% strain-energy error for the cube) are validated against analytical solutions and are plausible. The hyperelastic example is checked against a FEniCS reference. The paper is transparent about implementation details, gives code snippets, and explicitly acknowledges remaining caveats, including non-convexity and the sensitivity of the energy method to point sampling in Section 6.7.4. The soft spots are real but not all equal. The stress-test note lands: Eq. (77)-(78) define a 1D phase-field DEM loss that is not the phase-field energy used elsewhere in the paper; its Euler-Lagrange equation is a simple parabola, not the exponential profile of Eq. (71). As written, the reported 2.88% L2 error does not test the method as stated. Likewise, Eq. (28) defines the Neumann loss as a sum of tractions without a displacement multiplier, so it is dimensionally inconsistent with an external-work term. These look like typos, but the paper cannot be reproduced from the printed equations until they are fixed. The B parameter in Eq. (70) is self-referential, and the paper offers no quadrature error analysis for the discrete loss; Section 6.7.4 concedes that sampling affects stability and accuracy. No code or seeds are released, and repeated-run variance is not reported. Overall: the central idea is not new but is sound, and the applications are worth having. The paper deserves a serious referee, but the authors need to correct the phase-field energy, fix the Neumann loss, and release code before the quantitative claims can be accepted.","headline":"Useful application-driven Deep Ritz paper with a clear phase-field equation error; the mechanics benchmarks are plausible but need corrected losses and released code.","tokens_in":48676,"tokens_out":1739,"would_cite":true,"duration_ms":19125,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Using the system's own energy as the loss function, the Deep Energy Method solves mechanics boundary-value problems without meshes or labeled data.","keywords":["Deep Energy Method","energy-based loss function","partial differential equations","computational mechanics","neural network solvers","phase-field fracture","Kirchhoff plate bending","hyperelasticity"],"falsifier":"Run the pressurized-cylinder benchmark with the same network but with deliberately inadequate sampling (for example 10×10 interior points instead of 80×80) and compare the optimized displacement to the analytical solution; if the relative $L^2$ error grows far beyond 0.5% while the exact solution still minimizes the true energy integral, the discrete loss is not an accurate surrogate and the reported accuracy reflects the chosen sampling, not the method itself.","tokens_in":47571,"feed_emoji":"⚙️","tokens_out":6086,"duration_ms":60538,"temperature":0.7,"pith_summary":"The paper establishes the Deep Energy Method (DEM): instead of enforcing a PDE at collocation points, a neural network is trained by minimizing the total energy of the mechanical system, so the PDE is solved through its variational (Euler–Lagrange) structure. The authors argue this is the natural loss for mechanics, since the energy functional already encodes the governing equations, and they show a proof of concept across linear elasticity, hyperelasticity, phase-field fracture, piezoelectricity, and Kirchhoff plate bending. Reported relative $L^2$ displacement errors are 0.5% for a pressurized thick cylinder, 1.8% for a plate with a circular hole, and 1.05% for a hollow sphere, with strain-energy errors near 3–5%. For a 1D phase-field crack, DEM reaches a 2.88% error where a collocation method gives 70.6%, demonstrating the benefit of the energetic format.","feed_headline":"Deep Energy Method solves mechanics PDEs with no mesh or labeled data","feed_subtitle":"Bending, fracture, and piezoelectric benchmarks match analytical solutions to within a few percent.","key_machinery":"The load-bearing object is the total energy functional $\\mathcal{E}[u]$, e.g. $\\mathcal{E}[u]=\\int_\\Omega \\Psi(\\epsilon(u))\\,d\\Omega$ for linear elasticity (with fracture, piezoelectric, and plate terms in later examples), whose minimum over admissible displacements is the solution of the governing PDE by the Euler–Lagrange argument. The network output $u_p(x)$ is the trial function, and automatic differentiation supplies the strain and stress fields, so the loss is the discrete weighted sum $\\mathcal{L}(p)=\\sum_i \\Psi(\\epsilon(u_p(x_i)))w_i$ over sampling points. Dirichlet conditions are enforced exactly by multiplying the network output by functions that vanish on the constrained boundary. This converts the BVP into a finite-sum, non-convex optimization problem that is then minimized by gradient-based optimizers, connecting variational mechanics to deep learning.","core_discovery":"The central claim, stated on the paper's own terms, is that a deep neural network can approximate the solution of a boundary value problem in computational mechanics by minimizing the system's total potential energy rather than the residual of the strong-form PDE. The energy is taken as the loss function, Neumann (traction) conditions arise naturally from the variational principle, and Dirichlet conditions are built into the trial function, so no labeled data, mesh, or explicit weak-form derivation is needed. The paper demonstrates this Deep Energy Method on plane-stress and three-dimensional linear elasticity, hyperelastic torsion, phase-field fracture, piezoelectric coupling, and Kirchhoff plate bending, reporting relative $L^2$ displacement errors below about 2% for the classical elastic benchmarks and a 2.88% error on a 1D phase-field crack profile versus 70.6% for collocation. The authors present the method as a proof of concept that DNNs can serve as the approximation space in computational mechanics when the loss is the energy.","pith_inferences":["The reported accuracy depends on the sampling points acting as a quadrature rule; an obvious extension the paper leaves implicit is adaptive or error-driven point placement, which could make DEM robust on domains with stress concentrations or cutouts.","The same energy-as-loss principle should extend to dissipative or rate-dependent problems by using incremental potentials or rate functionals, though the paper only treats conservative systems.","Because a trained network gives a differentiable field at arbitrary query points, DEM could be paired with inverse problems or design loops that repeatedly evaluate stresses, a use not explored here.","A direct testable extension compares DEM's convergence rate against FEM on a problem with a known singular solution; if the quadrature is adequate, the energy minimizer should converge at a rate governed by network expressiveness and optimizer performance."],"forward_implications":["Mechanical boundary value problems become mesh-free: no mesh generation or element quadrature is required, only evaluation points in the domain.","No labeled data or external simulation is needed for training; the energy functional and boundary conditions are the only problem data, so the approach transfers to new materials or loads by changing the energy.","Natural (traction) boundary conditions are satisfied automatically through the variational principle, while essential (Dirichlet) conditions are built into the trial function, simplifying the loss.","Fourth-order problems like Kirchhoff plate bending bypass the $C^1$-continuity requirement that complicates mesh-based methods, because the network and automatic differentiation provide smooth derivatives.","For phase-field fracture, the energy formulation captures the sharp crack profile far more accurately than collocation with the same architecture, pointing toward fracture simulation without remeshing."],"supporting_citations":[{"why":"Supplies the core idea of using an energy functional as the neural-network loss (deep Ritz method), which DEM adapts and extends to mechanics problems.","marker":"[14]"},{"why":"Provides the collocation/physics-informed baseline whose strong-form residual loss DEM is compared against.","marker":"[9]"},{"why":"Establishes that ReLU networks can reproduce linear finite element spaces, motivating the use of DNNs as the approximation space.","marker":"[20]"},{"why":"Supplies the variational phase-field fracture energy whose minimization DEM uses to compute displacement and crack fields.","marker":"[33]"},{"why":"Gives the analytical plate solutions used to validate the deep energy and deep collocation results for Kirchhoff plates.","marker":"[36]"},{"why":"Shows trial functions that exactly satisfy boundary conditions, used to build the Dirichlet conditions into the network output.","marker":"[24]"},{"why":"Provides the unified network construction for PDEs in complex geometries with boundary conditions enforced a priori.","marker":"[25]"},{"why":"Provides the finite-element reference solution for the hyperelastic torsion benchmark and the plate benchmark comparison.","marker":"[29]"}],"fun_headline_variants":["Deep energy method: neural nets solve mechanics PDEs without meshes","No mesh, no labels: deep energy method solves mechanics PDEs","Neural network minimizes potential energy to crack PDE benchmarks","Deep Energy Method matches analytical solutions within 2%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The discrete sum over sampling points is assumed to be a faithful quadrature of the energy integral, so that minimizing it yields a function close to the true energy minimizer; the paper does not analyze this quadrature error and even notes that point sampling affects the stability and accuracy of the method for the annular plate.","fun_headline_variants_meta":{"raw":{"variants":["Deep energy method: neural nets solve mechanics PDEs without meshes","No mesh, no labels: deep energy method solves mechanics PDEs","Neural network minimizes potential energy to crack PDE benchmarks","Deep Energy Method matches analytical solutions within 2%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000547,"raw_usage":{"total_tokens":2659,"prompt_tokens":1031,"completion_tokens":1628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1559}},"tokens_in":647,"tokens_out":1628,"duration_ms":12246,"temperature":1.0,"reasoning_tokens":1559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:44:51.520234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pressurized-cylinder benchmark with the same network but with deliberately inadequate sampling (for example 10×10 interior points instead of 80×80) and compare the optimized displacement to the analytical solution; if the relative $L^2$ error grows far beyond 0.5% while the exact solution still minimizes the true energy integral, the discrete loss is not an accurate surrogate and the reported accuracy reflects the chosen sampling, not the method itself.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the analytical plate solutions used to validate the deep energy and deep collocation results for Kirchhoff plates."},{"cited_title":"Enhancing approximation abilities of neural networks by training derivatives","cited_arxiv_id":"1712.04473","evidence_quote":"Shows trial functions that exactly satisfy boundary conditions, used to build the Dirichlet conditions into the network output."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the unified network construction for PDEs in complex geometries with boundary conditions enforced a priori."},{"cited_title":"Logg, K.-A","cited_arxiv_id":null,"evidence_quote":"Provides the finite-element reference solution for the hyperelastic torsion benchmark and the plate benchmark comparison."}],"review_version":1}