{"id":"b00c2f63-991a-4fde-83ee-4dd893d62f65","arxiv_id":"2505.01060","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MPNO learns a nonlocal peridynamic constitutive law with a monotone gradient network, guaranteeing unique solutions in the small-deformation regime and showing improved robustness in downstream simulations.","lead":"The paper introduces MPNO, a neural operator that learns peridynamic material laws while enforcing a monotone stretch-force relation, guaranteeing solution uniqueness for small deformations. It demonstrates the method on synthetic nonlinear materials and on molecular-dynamics data, reporting better downstream simulation robustness than plain neural networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The uniqueness guarantee is proved for the linearized energy (3.3), but the deployed MPNO operator (3.4) and the downstream solver use the exact nonlinear stretch; the abstract's 'guaranteeing solution uniqueness of MPNO' is therefore not established.","rationale":"The reader's weakest_assumption targets the absence of constraints on W in Eq. (3.5), but for scalar λ that concern is not supported: the shared-weight cascaded construction yields a nonnegative derivative for arbitrary real W, so nonnegativity of W is unnecessary. The reader's broader rationale, however, includes the mismatch between the linearized theorem and the nonlinear solved system, which is exactly the central weakness. The paper is explicit in Section 3.5 that the uniqueness theorem is for the small-deformation linearization, but the abstract and the downstream evaluation treat the nonlinear MPNO as the object whose uniqueness is guaranteed. Because the proof uses affine-in-η stretches, it does not cover the exact operator in Eq. (3.4), and there is no continuity or local-uniqueness argument bridging the two. This is a genuine load-bearing gap in the strongest claim, but it is a gap in proof scope rather than an observed failure, and the empirical results support the practical value of the method. A conditional verdict remains appropriate; a concrete multiplicity test would determine whether the gap is merely formal or has observable consequences.","tokens_in":19695,"tokens_out":18234,"duration_ms":206346,"concrete_test":"On a trained MPNO in the Section 4 small-strain setup, fix one test loading and boundary layer. Solve Eq. (3.2) with the exact λ from Eq. (3.4) using (i) the two-phase initialization and (ii) at least 100 random initial guesses drawn uniformly from a 1% ball around the linearized solution, recording all distinct converged fixed points up to rigid motions. If two non-equivalent solutions appear, the paper must either prove local uniqueness of the nonlinear operator or explicitly restrict the uniqueness claim to the linearized model; if no multiplicity appears, the practical concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's specific concern about missing nonnegativity constraints on W does not land. For scalar bond stretch λ, W is a vector and the recursions in Eq. (3.5) give d z_l/dλ = p_l ⊙ W with p_l ≥ 0, so the output derivative is Σ_i W_i^2 α_{L,i} σ'_L,i(z_{L-1,i}) p_{L-1,i} ≥ 0 for any real W. Monotonicity of mGradNet-C therefore does not require W ≥ 0. The load-bearing gap is elsewhere: Theorem 3.2 proves uniqueness only under the linearization (3.3), which makes λ affine in η and lets strict convexity of w(·,ξ) transfer to the energy functional. The MPNO operator in Eq. (3.4) and the nonlinear solver in Algorithm 3.1 use the exact λ = |ξ+η|/|ξ| and exact force direction (ξ+η)/|ξ+η|. Convexity of w in λ does not, in general, make w(|ξ+η|/|ξ|) convex in η, because the norm is nonlinear and w need not be nondecreasing; the Blatz-Ko g(λ)=λ−λ^{-3} used in Section 4 is actually negative for λ<1, so the associated w is not monotone there. Thus no theorem in the paper rules out multiple solutions of the nonlinear equation that is actually solved, and the central uniqueness claim is overbroad as stated. Strict convexity is also not architecturally guaranteed, since mGradNet-C only enforces nondecreasing g and k may vanish on parts of the family.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Monotone Peridynamic Neural Operator (MPNO), a bond-based peridynamic constitutive model learned from full-field displacement/force data. The constitutive relation is parameterized by a monotone gradient network for g(λ) and an MLP with a final ReLU for the kernel k(ξ). The authors prove a uniqueness theorem (Theorem 3.2) for the energy functional under a small-deformation linearization, propose a two-phase solver, give a truncation-error analysis for kernel and stretch learning, and report experiments on a 1D Blatz-Ko synthetic model and a 2D molecular-dynamics homogenization task.","tokens_in":20044,"tokens_out":6495,"duration_ms":60975,"significance":"If the claims held, MPNO would be a valuable step toward data-driven constitutive models with a priori well-posedness guarantees in small-deformation regimes. The paper's positive contributions are a rigorous, if elementary, uniqueness condition for the linearized bond-based peridynamic energy; an architecture that provably produces monotone gradient maps (for scalar stretch, monotonicity does not require positivity of W); and an experimental demonstration of convergence with grid refinement and of robust downstream solution. The application to molecular dynamics data on an unseen domain is a strong practical result. However, the central guarantee as stated is not established for the actual nonlinear operator used in the solver, and the theoretical convergence analysis is conditional on unquantified training residuals. These issues require a substantial revision.","major_comments":[{"comment":"Theorem 3.2 proves uniqueness only for the linearized energy with λ ≈ 1 + ξ·η/|ξ|², but the MPNO operator in Eq. (3.4) and the solver in Algorithm 3.1 use the exact stretch λ = |ξ+η|/|ξ| and force direction (ξ+η)/|ξ+η|. Strict convexity of w in λ does not in general imply convexity of w(|ξ+η|/|ξ|, ξ) in η, since the norm is nonlinear and w need not be nondecreasing; indeed, the Blatz-Ko g(λ)=λ−λ⁻³ used in Section 4 is negative for λ<1, so its primitive is not monotone there. Consequently, no result in the paper rules out multiple solutions of the nonlinear equation that is actually solved, and the abstract and Section 6 statements promising 'guaranteed solution uniqueness of MPNO' are overbroad. At minimum, the uniqueness claim should be explicitly restricted to the linearized small-deformation model, and the relation to the nonlinear solver should be presented as empirical.","section":"§3.1–§3.2, Theorem 3.2 and Eq. (3.4)"},{"comment":"The monotonicity of mGradNet-C does not require a nonnegativity constraint on W: for scalar λ, the recursion gives d z_l/dλ = p_l ⊙ W with p_l ≥ 0, so the output derivative is a sum of squared entries times nonnegative factors. However, strict convexity of the energy density is not guaranteed. Theorem 3.2 requires w(·,ξ) to be strictly convex for every ξ, but the architecture only enforces g' ≥ 0; intervals where α_l or the relevant W components vanish give g' = 0 and hence only weak convexity. Moreover, kNN is constrained only by a final ReLU to be nonnegative, so where kNN(ξ)=0 the micropotential is independent of λ and strict convexity fails. The paper does not state additional conditions (e.g., positive lower bounds on the scaling weights and on kNN) under which the theorem's hypothesis is satisfied, so the claimed architectural guarantee is not established.","section":"§3.2, Eq. (3.5)"},{"comment":"The theoretical convergence statement is weaker than the abstract suggests. Lemma 4.1 bounds ∥A(kNN−ktrue)∥ by O(Δx^q)+O(∥e_k∥), where e_k is the training residual; the second term is not quantified and depends on the optimization, and Remark 4.2 notes that the condition number of A^T A grows as Δx→0, so the bound does not imply convergence of kNN to ktrue in the reported norm. The numerical convergence shown in Fig. 5 is encouraging, but the claim in the abstract that convergence 'is shown theoretically' should be softened to a conditional estimate.","section":"§4.2, Lemma 4.1 and Remark 4.2"}],"minor_comments":[{"comment":"The strict-convexity argument should spell out that η ↦ 1 + ξ·η/|ξ|² is affine, so composition with strictly convex w yields strict convexity in η; the 'if and only if' statement for equivalence also needs an explicit justification.","section":"§3.1, proof of Theorem 3.2"},{"comment":"The empirical measures ρξ and ρλ use weights w^{(i)} and v^{(i)} that appear to omit the other factor of the product g(λ)k(ξ); please clarify the definitions.","section":"§3.3, Eqs. (3.11)–(3.12)"},{"comment":"In the neighbor search in line 3, the exclusion of the point xk = xj should be stated explicitly.","section":"§3.6, Algorithm 3.1"},{"comment":"The sentence in the text that the one-phase solution has 'a higher average error' is vague; please quote the numerical values from Table 1.","section":"§4.1, Table 1"},{"comment":"References [51] and [52] are the same paper; also [8] should include the volume/page or DOI for the Gradient networks article.","section":"References"},{"comment":"The notation (G^NN_{Δx})^{-1} in Eq. (3.10) is nonstandard; define it as the numerical solution map of the learned model.","section":"§3.4, Eq. (3.10)"},{"comment":"There are several typos, e.g., 'trianing' in §3.3 and 'oftaining' in §6; a careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central uniqueness theorem is correct but narrow; the paper's framing overclaims its reach. I believe the issues are fixable with revision: restrict the well-posedness claim to the linearized model, add conditions for strict convexity, and calibrate the convergence claims. The experimental part, especially the MD application, is solid and within the journal's scope. The novelty of Theorem 3.2 is modest—it is a direct convex-analysis consequence—but the combination with the architecture and the MD demonstration may be sufficient. The reader's concern about missing nonnegativity of W does not land, but the strict-convexity and nonlinear-operator gaps are real. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a solid, well-written engineering paper with a genuinely useful architectural idea, but its headline claim is a step ahead of its theorems. Theorem 3.2 proves uniqueness for the energy functional under the small-deformation linearization (3.3), which makes λ affine in η. The MPNO operator in (3.4) and the solver in Algorithm 3.1 use the exact nonlinear stretch λ = |ξ+η|/|ξ| and the exact force direction. Strict convexity of w(λ,ξ) in λ does not generally make w(|ξ+η|/|ξ|, ξ) convex in η — that composition is convex only if w is also nondecreasing, and the paper's own Blatz-Ko example g(λ)=λ−λ^{-3} is negative for λ<1. So no theorem in the paper rules out multiple solutions of the equation that is actually solved.\n\nWhat is genuinely good: the mGradNet-C construction is correct for scalar bond stretch. The reader's worry about missing W constraints is a red herring: for a scalar λ, W is a vector and the output derivative is a sum of squared W components times nonnegative factors, so monotonicity holds for any real W. The synthetic convergence study is careful and the two-phase initialization strategy makes practical sense. The MD experiments show real generalization to a new domain and unseen loadings. The convergence analysis (Lemma 4.1) is honest but conditional: the bound contains an unquantified training residual, so it is consistency given perfect optimization, not a free convergence theorem.\n\nThe main soft spot is the theory/experiment mismatch. Either prove uniqueness for the exact nonlinear operator under a small-deformation regime, or restate the contribution as 'a linearized model with guaranteed uniqueness, used as an initialization and regularization prior.' Secondary issues: the architecture only forces g nondecreasing, not strictly increasing, and k can vanish, so strict convexity of w is not architecturally guaranteed. Also, no code or data is provided, which makes reproduction harder. The citation pattern looks fine.\n\nThis deserves a serious referee. I would send it out, but with a clear expectation of major revision. The experiments are strong enough that the paper should survive, but the framing must be aligned with what is actually proved.\n\nBest","headline":"Solid architecture and strong experiments, but the uniqueness guarantee is proved only for a linearized proxy, not for the nonlinear operator actually trained and solved.","tokens_in":20567,"tokens_out":5848,"would_cite":false,"duration_ms":58831,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q25","68R10","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A learned material law can be made well-posed by forcing its bond force to come from a convex energy, so that in small deformations any two solutions differ only by rigid motion.","keywords":["neural operators","peridynamics","data-driven constitutive modeling","solution uniqueness","convex energy","monotone gradient network","nonlocal models","molecular dynamics homogenization"],"falsifier":"Take any instance of the mGradNet-C architecture, random or trained, with nonnegative $\\alpha_l,\\beta_l$ and arbitrary shared weights and biases, and evaluate $g_{\\text{NN}}(\\lambda)$ on a fine grid spanning the observed stretch range. If any increment shows the output decreasing as $\\lambda$ increases, or equivalently if $\\partial g_{\\text{NN}}/\\partial\\lambda$ is negative anywhere, the claimed monotonicity of the architecture fails and the convexity and uniqueness guarantee cannot be invoked for that model. A corresponding physical check would be to solve the small-deformation model and exhibit two non-equivalent displacement fields with the same boundary data and loading.","tokens_in":19466,"feed_emoji":"🧱","tokens_out":8295,"duration_ms":85902,"temperature":0.7,"pith_summary":"This paper claims that a learned constitutive model for nonlinear materials can be made well-posed by design: in the small-deformation regime, the model's predicted displacement field is unique up to rigid motions, rather than branching into non-physical solutions. The mechanism is to represent the bond-force law as the derivative of a convex energy, enforced by using a monotone-gradient neural network for the stretch dependence. A uniqueness theorem shows that strict convexity of the micropotential in the bond stretch rules out distinct minimizers, and the architecture is chosen to satisfy that hypothesis automatically. On synthetic data the learned law converges to the manufactured ground truth as the measurement grid refines, and on molecular-dynamics data the model transfers to a new geometry and new loadings. This matters because unconstrained neural constitutive laws often produce divergent or nonphysical downstream simulations, and the paper offers a path to data-driven material models with a well-posedness guarantee analogous to classical elasticity.","feed_headline":"Monotone network gives learned material laws unique answers","feed_subtitle":"Peridynamic models trained from data stay well-posed under small deformations, so new loadings yield one solution, not many.","key_machinery":"The load-bearing object is the cascaded monotone gradient network (mGradNet-C), specified in Eq. (3.5): a layered network with nonnegative scaling weights $\\alpha_l, \\beta_l$, a weight matrix $W$ shared across all layers, and increasing activation functions, designed so that its output is the gradient of a convex potential. MPNO uses this network for $g(\\lambda)$ in the separable bond force $g(\\lambda)k(\\xi)(\\xi+\\eta)/|\\xi+\\eta|$, while a separate multilayer perceptron with a ReLU output learns the nonnegative kernel $k(\\xi)$. An increasing $g$ makes the micropotential $w(\\lambda,\\xi)$ convex in the bond stretch $\\lambda$, which by the paper's Theorem 3.2 makes the total energy strictly convex up to rigid displacements and thereby guarantees solution uniqueness in the small-deformation regime. The same network also supports the two-phase solver, because the unique small-deformation solution can be computed first and then refined for the full nonlinear model.","core_discovery":"The central claim is that nonlinear bond-based peridynamics has a simple sufficient condition for solution uniqueness, and that condition can be hard-wired into a neural operator. For small deformations, where the bond stretch linearizes as $\\lambda \\approx 1 + \\xi\\cdot\\eta/|\\xi|^2$, the paper proves that if the micropotential $w(\\lambda,\\xi)$ is strictly convex in $\\lambda$ for every bond vector $\\xi$, then any two minimizers of the total energy are equivalent: they differ only by a rigid translation, a rigid rotation, or both. Because the pairwise force in a microelastic material is $\\partial w/\\partial\\eta$, convexity is equivalent to the stretch-dependent part $g(\\lambda)$ being a monotonically increasing function. The proposed Monotone Peridynamic Neural Operator therefore parameterizes $g$ with a monotone gradient network and the influence function $k(\\xi)$ with a nonnegative-output network, learning both from full-field displacement-loading data. The paper then demonstrates convergence of the learned model to the ground truth as the grid spacing decreases and shows that the unique small-deformation solution serves as a robust initial guess for large-deformation simulations.","pith_inferences":["If monotonicity of the mGradNet-C architecture can be rigorously proved, or replaced by a provably monotone parameterization, the convex-energy template should extend to state-based peridynamics and to dynamic problems, where uniqueness would prevent spurious oscillations in long-time simulations.","The strict-convexity condition is sufficient but likely not necessary; a weaker monotone-operator condition on the force could give existence and uniqueness in larger deformation regimes, which is a testable reformulation of the theorem.","Because the kernel $k(\\xi)$ carries material-specific microstructure while $g(\\lambda)$ is shared, MPNO could serve as a foundational material model: pretrain $g$ once, then fit only $k$ per material, with the uniqueness guarantee carrying over unchanged.","The two-phase solver's success suggests that the small-deformation unique solution acts as a natural branch selector in the large-deformation regime; whether that selection always picks the physically preferred branch is a question the paper leaves open."],"forward_implications":["Downstream simulations with new, unseen loadings will return a single displacement field, up to rigid motion, in the small-deformation regime, eliminating divergent or nonphysical solution branches.","Learning the constitutive relation as a monotone stretch function plus a nonnegative kernel does not destroy expressive power: on the synthetic hyperelastic dataset the learned product $g(\\lambda)k(\\xi)$ reaches sub-percent errors and converges as the measurement grid is refined.","The two-phase solver, which uses the unique small-deformation solution as an initial guess, extends robustness to large-deformation settings where strict convexity is not required.","The same architecture transfers to real data: a homogenized continuum model learned from molecular dynamics simulations predicts forces and displacements on a different domain under discontinuous loadings.","The convergence analysis provides a practical diagnostic: at least first-order, and under smoothness assumptions second-order, decay of model error with mesh refinement, so grid-convergence tests can validate learned nonlocal laws."],"supporting_citations":[{"why":"Supplies the cascaded monotone gradient network architecture used to parameterize $g(\\lambda)$ as the gradient of a convex potential.","marker":"[8]"},{"why":"Defines the peridynamic neural operator framework that MPNO extends, providing the operator architecture and the training and evaluation setup.","marker":"[19]"},{"why":"Provides the classical result that the gradient of a convex function is monotone, which bridges monotone $g$ to convex energy in Remark 3.3.","marker":"[38]"},{"why":"Introduces the kernel exploration measures and conditioning analysis used in the mesh-convergence error estimates.","marker":"[31]"},{"why":"Supplies the classical convexity and polyconvexity context for existence and uniqueness in nonlinear elasticity and the buckling counterexample motivating the small-deformation restriction.","marker":"[3]"},{"why":"Supplies the molecular-dynamics data generation and coarse-graining procedure used in the homogenized-model application.","marker":"[44]"},{"why":"Establishes the data-driven peridynamic upscaling methodology from molecular dynamics that the MD application builds on.","marker":"[51]"}],"fun_headline_variants":["Monotone peridynamic operator yields unique solutions","Guaranteed uniqueness for data-driven material laws","Hard-wired convexity gives unique material responses","Neural operator with built-in unique-solution guarantee","Monotone network ensures one solution for material models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The uniqueness guarantee rests on the assertion that the special monotone-gradient network outputs a monotonically increasing function of the bond stretch for every allowed set of weights; the paper cites the architecture's design for this property but does not prove it or constrain the shared weight matrix, so if training ever produces a non-monotone realization, the convex-energy hypothesis, and with it the uniqueness theorem, no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["Monotone peridynamic operator yields unique solutions","Guaranteed uniqueness for data-driven material laws","Hard-wired convexity gives unique material responses","Neural operator with built-in unique-solution guarantee","Monotone network ensures one solution for material models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1731,"prompt_tokens":1045,"completion_tokens":686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":612}},"tokens_in":661,"tokens_out":686,"duration_ms":6736,"temperature":1.0,"reasoning_tokens":612,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:28:07.620059+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any instance of the mGradNet-C architecture, random or trained, with nonnegative $\\alpha_l,\\beta_l$ and arbitrary shared weights and biases, and evaluate $g_{\\text{NN}}(\\lambda)$ on a fine grid spanning the observed stretch range. If any increment shows the output decreasing as $\\lambda$ increases, or equivalently if $\\partial g_{\\text{NN}}/\\partial\\lambda$ is negative anywhere, the claimed monotonicity of the architecture fails and the convexity and uniqueness guarantee cannot be invoked for that model. A corresponding physical check would be to solve the small-deformation model and exhibit two non-equivalent displacement fields with the same boundary data and loading.","supporting_citations":[{"cited_title":"Chaudhari, S","cited_arxiv_id":null,"evidence_quote":"Supplies the cascaded monotone gradient network architecture used to parameterize $g(\\lambda)$ as the gradient of a convex potential."},{"cited_title":"Jafarzadeh, S","cited_arxiv_id":null,"evidence_quote":"Defines the peridynamic neural operator framework that MPNO extends, providing the operator architecture and the training and evaluation setup."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the classical result that the gradient of a convex function is monotone, which bridges monotone $g$ to convex energy in Remark 3.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the kernel exploration measures and conditioning analysis used in the mesh-convergence error estimates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical convexity and polyconvexity context for existence and uniqueness in nonlinear elasticity and the buckling counterexample motivating the small-deformation restriction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the molecular-dynamics data generation and coarse-graining procedure used in the homogenized-model application."}],"review_version":1}