{"id":"30514f7c-5459-4c6c-9e81-cf15d878c3ed","arxiv_id":"2502.05463","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A recurrent Fourier neural operator can learn memory- and microstructure-dependent homogenized constitutive laws, with a universal approximation theorem for 1D Kelvin-Voigt viscoelasticity and demonstrations on viscoelastic and elasto-viscoplastic materials.","lead":"What if we could train a neural network once, on many tiny material samples, and then simulate an entire composite material without re-solving the microscopic structure each time? This paper introduces a recurrent neural operator that learns how stress evolves with strain history and with the material's internal microstructure, with a proof that it can approximate the exact homogenized law for a standard viscoelastic model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.7 relies on an unproven trade-off between FNM approximation error e_G and Lipschitz constant L_G; without controlling L_G(e_G), the Gronwall factor e_G exp(L_G T)/L_G cannot be made small.","rationale":"The reader's weakest-assumption analysis is correct, and I find no stronger objection. The proof of Theorem 4.7 rests on three ingredients: Lipschitz regularity of the cell-problem solution map (Lemmas 4.1 and 4.2), piecewise-constant approximation of BV microstructures (Proposition 4.4), and universal approximation of the affine ODE maps Fpc and Gpc by FNMs (Proposition C.3). The first two ingredients are sound and well supported by the paper's analysis. The gap is entirely in the third: the underlying universal approximation theorem (Proposition C.2, based on [31]) does not control the Lipschitz constant of the approximating FNM as the approximation tolerance tends to zero. Since Gpc is affine in ξ, a specialized construction might give bounded L_G, and the theorem could be repaired; the authors do not provide such a construction. This is not an objection to the numerical experiments, which are substantial and support the practical value of FNM-RNO, but it means the advertised universal approximation theorem is unproven as stated. The appropriate disposition remains the reader's CONDITIONAL verdict, with the missing error-versus-Lipschitz estimate as the condition. The proposed check is the minimal experiment that would either convert the gap into a lemma or expose a counterexample.","tokens_in":42798,"tokens_out":5109,"duration_ms":56921,"concrete_test":"Attempt to prove a quantitative strengthening of Proposition C.3: for every δ>0 there exists an FNM GFNM with sup error ≤δ and Lipschitz constant L_G ≤ C on the compact set in (4.27), with C independent of δ. If this holds, insert it into (4.32) and Theorem 4.7 goes through. If it fails, construct the approximating FNM that minimizes L_G for δ = 10^{−1}, 10^{−2}, …, 10^{−6} and check whether R(δ) = (δ/L_G(δ)) exp(L_G(δ) T) decreases to 0 for T=1. This distinguishes a repairable gap from a false theorem.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper, Theorem 4.7, is not established by the proof as written. The argument approximates Gpc by an FNM GFNM with pointwise error e_G (Proposition C.3), then applies Gronwall to control ξpc−ξRNO, obtaining in (4.32)–(4.34) a stress-error contribution of order (e_G/L_G) exp(L_G T), where L_G is the Lipschitz constant of GFNM in the hidden-state variable ξ. The proof then states that e_F, e_G can be chosen sufficiently small to make this term smaller than e/2. However, Proposition C.3, via Proposition C.2, guarantees only existence of GFNM with sup error less than e_G; it gives no control on L_G as a function of e_G. If the universal-approximation construction forces L_G to grow as e_G shrinks, for example L_G ∼ e_G^{−α} with α>0, then (e_G/L_G) exp(L_G T) diverges for every T>0, and the claimed uniform bound cannot be achieved. The same issue enters the a priori bound on ξRNO used to define the compact set on which FFNM is applied. This is a genuine gap at equations (4.32)–(4.34). The theorem may be repairable through a bespoke bounded-Lipschitz FNM construction for the affine map Gpc, but the paper supplies no argument for the necessary error-versus-Lipschitz trade-off.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a recurrent neural operator architecture, FNM-RNO, for learning homogenized constitutive laws that depend both on strain history and on the material microstructure, and it analyzes this architecture in the setting of one-dimensional Kelvin-Voigt viscoelasticity. The theoretical part establishes Lipschitz continuity of the cell-problem solution map with respect to material coefficients, proves that continuous bounded-variation microstructures can be approximated by piecewise-constant ones, and states a universal approximation theorem (Theorem 4.7) claiming that FNM-RNO can uniformly approximate the exact homogenized stress-strain map. The numerical part trains the model on viscoelastic and elasto-viscoplastic cell-problem data and demonstrates accurate stress prediction, generalization from piecewise-constant to continuous microstructures, robustness to spatial and temporal resolution changes, and deployment in macroscale simulations.","tokens_in":43113,"tokens_out":5187,"duration_ms":53565,"significance":"If Theorem 4.7 is established, the paper would provide the first universal approximation guarantee for a microstructure-dependent, memory-dependent learned constitutive law, and it would justify the proposed architecture beyond heuristics. The supporting analytical results are valuable in themselves: the Lipschitz regularity of the cell-problem map is carefully proved, the piecewise-constant approximation result is clean, and the explicit closed-form expression for the Prony coefficients in Lemma 3.4 is a useful contribution. The numerical study is extensive, including 2,500-sample test sets, two material families for viscoelasticity, an elasto-viscoplastic extension, and macroscale deployment. The main caveat is that the proof of the central approximation theorem contains a load-bearing gap concerning the Lipschitz constant of the learned dynamics; this needs to be repaired before the universal approximation claim can be accepted.","major_comments":[{"comment":"The proof of Theorem 4.7 assumes that the term (e_G/L_G) exp(L_G T) can be made arbitrarily small by choosing e_G sufficiently small. Proposition C.3 guarantees only the existence of an FNM GFNM with uniform approximation error below e_G; it provides no control on L_G, the Lipschitz constant of GFNM in the hidden-state variable ξ. Universal-approximation constructions can have Lipschitz constants that grow as the approximation error shrinks, in which case the Gronwall factor in (4.32) may diverge for every positive T. The same issue affects the a priori bound on ξ_RNO in (4.33), which is used to define the compact set on which FFNM is approximated. As written, the claimed bound (4.34) does not follow from the stated results; a repair would require an explicit bounded-Lipschitz FNM construction for the affine map G_pc, or a different argument that controls the error-versus-Lipschitz trade-off.","section":"§4.3, Eqs. (4.32)–(4.34)"},{"comment":"The proof implicitly identifies the internal-state dimension L of the FNM-RNO with the dimension n(e/2) of the hidden state ξ_pc in the piecewise-constant model Ψ_pc. This identification is not stated, and the bounds (4.30)–(4.34) use √L for both ξ_pc and the linear term ⟨1, ξ_RNO − ξ_pc⟩. If L and n(e/2) differ, those expressions compare vectors of different dimensions. Setting L = n(e/2) explicitly in the theorem statement would fix this, but it is a necessary clarification for the proof to be coherent.","section":"§4.3, Theorem 4.7"}],"minor_comments":[{"comment":"The opening sentence of Theorem 4.7 states a pointwise claim for any E, ν, and ϵ satisfying Assumptions 1, but the proof establishes a uniform statement over the bounded-variation classes described in (4.28). Please align the wording with the uniform claim.","section":"Theorem 4.7"},{"comment":"The symbol T is used both for the time interval [0, T] and for the number of Fourier layers in the FNM. This is confusing in definitions such as the layer index t ∈ [T ]; a different symbol for the number of layers would improve readability.","section":"Definition 2.4"},{"comment":"The stress-prediction panels in Figure 3 are repeated across columns in a way that makes it difficult to see which column corresponds to which model. Please simplify the layout or use clearer separation between the true, FNM-RNO, and no-memory curves.","section":"Figure 3"},{"comment":"The phrase 'discretization agnostic' is stronger than what the experiments show: the results in Figure 4 indicate sensitivity to temporal resolution, as the authors acknowledge. A phrase such as 'robust to changes in resolution' would be more accurate.","section":"§5.3"},{"comment":"Appendix D is an honest and useful discussion, but its main message—that training without the penalty term can give smaller relative L2 error while having much larger relative L∞ error—should be summarized in the main text, since it directly informs the loss-function choice in (5.3).","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the gap in the proof of Theorem 4.7 is real and load-bearing, but it appears repairable: G_pc is affine in the hidden state, so a bespoke bounded-Lipschitz FNM approximation should be feasible. The Lipschitz regularity results and the numerical work are strong, and the paper fits the journal's scope. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The architecture in Section 2 is the real novelty: a recurrent neural ODE forced by strain and with the full microstructure function as an input to the Fourier neural map. This is a clean way to build memory and material dependence into one operator, and the numerical results justify it. The paper trains on piecewise-constant microstructures and transfers to continuous high-memory microstructures; the macroscale deployment experiments are well designed and the accuracy is genuinely good. The Lipschitz regularity results for the 1D Kelvin-Voigt cell problem are careful and useful.\n\nThe problem is Theorem 4.7. The proof approximates the piecewise-constant right-hand side G_pc by an FNM G_FNM with pointwise error e_G, then applies Gronwall to control xi_pc - xi_RNO. The bound contains the factor (e_G/L_G) exp(L_G T), where L_G is the Lipschitz constant of G_FNM. The paper says \"choose e_G sufficiently small\" and lets it go at that. But Proposition C.3 gives no control on L_G as a function of e_G. If L_G grows faster than log(1/e_G), the term blows up and the uniform bound fails. This is a real gap, at equations (4.32)-(4.34), and it is load-bearing: the universal approximation theorem is the main theoretical claim. The theorem may be salvageable, since G_pc is a simple affine map and a bounded-Lipschitz FNM construction with explicit error-vs-Lipschitz trade-off is plausible, but the paper does not provide it.\n\nEverything else holds up. The numerics are extensive, the elasto-viscoplastic experiment extends the method beyond the theory, and the limitations (1D theory, no released code/data) are stated honestly. The citation pattern is appropriate; the architecture builds directly on prior work [5,6,27,39] without overclaiming.\n\nVerdict: worth a serious referee. The gap should be fixed or the theorem stated conditional on a bounded-Lipschitz approximation property. I would send it to peer review with that request.","headline":"The FNM-RNO architecture is a useful contribution and the numerics are convincing, but the main universal approximation theorem has a genuine proof gap that needs fixing before the theory can be cited.","tokens_in":43680,"tokens_out":3506,"would_cite":true,"duration_ms":33566,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35B27","65M60","68T07","74D05","74D10","74Q10","74Q15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A recurrent neural operator can approximate the exact homogenized viscoelastic constitutive law to any prescribed accuracy.","keywords":["constitutive modeling","homogenization","memory","microstructure","Kelvin-Voigt","viscoelasticity","neural operator","universal approximation"],"falsifier":"Train the FNM-RNO on the piecewise-constant Kelvin-Voigt cell problem at decreasing error tolerances and measure the Lipschitz constant of the learned right-hand side with respect to the internal state; if that constant grows faster than the reciprocal of the training error, the Gronwall factor in the proof cannot be driven to zero and the uniform approximation bound of Theorem 4.7 would fail for long time horizons.","tokens_in":42566,"feed_emoji":"🧠","tokens_out":9553,"duration_ms":85229,"temperature":0.7,"pith_summary":"Heterogeneous materials are usually modeled by homogenization, which replaces fine microstructure with an effective macroscale law, but the cell problems that define that law are expensive and must be redone for every new material. This paper constructs a recurrent neural operator, FNM-RNO, that takes both the strain history and the microstructure function as inputs and outputs the averaged stress. The central theoretical result is a universal approximation theorem: in one-dimensional Kelvin-Voigt viscoelasticity, such a network can approximate the exact homogenized constitutive law uniformly to any prescribed error over a large class of microstructures and strain inputs. If correct, the learned law can be deployed in macroscale simulations and applied to new microstructures without retraining, and the numerical experiments show it also works for elasto-viscoplastic materials beyond the theory.","feed_headline":"Learned neural operator emulates viscoelastic laws to any accuracy","feed_subtitle":"Trained once, it predicts stress for unseen microstructures and strain histories.","key_machinery":"The central object is the FNM-RNO, a neural differential equation with internal state vector $\\xi(t)$, stress output $\\sigma(t)=F_{\\mathrm{FNM}}(\\epsilon(t),\\dot{\\epsilon}(t),\\xi(t);M)$, and dynamics $\\dot{\\xi}(t)=G_{\\mathrm{FNM}}(\\epsilon(t),\\xi(t);M)$, where $M$ is the microstructure function and $F_{\\mathrm{FNM}},G_{\\mathrm{FNM}}$ are Fourier Neural Mappings. The proof machinery has three parts: the cell problem is Lipschitz in the microstructure, so small changes in the material fields produce controlled changes in stress; any bounded-variation microstructure can be replaced by a piecewise-constant one whose homogenized memory kernel is a finite sum of exponentials, reducing the constitutive law to a linear ODE with continuous coefficients; and Fourier Neural Mappings universally approximate the ODE right-hand sides on compact sets, which, combined with Gronwall's inequality, yields the theorem.","core_discovery":"The central claim, Theorem 4.7, is that for any tolerance $e>0$ there exist Fourier Neural Mappings $F_{\\mathrm{FNM}}$ and $G_{\\mathrm{FNM}}$ such that the FNM-RNO map uniformly approximates the exact homogenized Kelvin-Voigt constitutive law on the set of microstructures with bounded variation and bounded strain trajectories. The proof isolates three properties: Lipschitz continuity of the cell problem in the microstructure, the fact that piecewise-constant microstructures produce a memory kernel that is a finite sum of exponentials and hence a finite-dimensional linear ODE, and the universal approximation property for Fourier Neural Mappings. The theorem is stated in one-dimensional linear viscoelasticity; the numerical section extends the architecture to continuous high-memory microstructures and to nonlinear elasto-viscoplasticity.","pith_inferences":["If the Gronwall factor is the bottleneck, the uniform guarantee degrades with time horizon, so long-time deployment may need a stability analysis of the recurrent operator itself.","The same internal-variable construction could extend to other history-dependent homogenized laws, such as thermo-viscoelastic or poroelastic composites, wherever a Markovian internal-state ODE representation exists.","A direct empirical check of the theorem's assumption would be to train FNM-RNOs at decreasing error thresholds and measure the sensitivity of the learned right-hand side to the internal state; if that sensitivity grows too fast, the uniform approximation claim would need a modified architecture or a different error decomposition."],"forward_implications":["A single trained FNM-RNO can predict stress for microstructures outside its training set, as long as they satisfy the bounded-variation assumptions, with error uniformly controlled by the theorem.","The model can be embedded in macroscale simulations, replacing repeated cell-problem solves; in the paper's experiments its displacement error is comparable to a multiscale simulation with 20 cells per wavelength.","The architecture is discretization agnostic: predictions remain accurate when evaluated at spatial and temporal resolutions different from the training data.","The same architecture learns elasto-viscoplastic constitutive laws with mean relative errors around 1-3% on stress and plastic strain, even though the theoretical guarantee is for linear viscoelasticity.","Including at least a few internal variables substantially reduces error compared to a memoryless stress response, confirming that the recurrent structure captures the homogenized memory kernel."],"supporting_citations":[{"why":"Derives the Kelvin-Voigt cell problem, the piecewise-constant memory kernel form, and the weighted H1 norms used throughout.","marker":"[6]"},{"why":"Establishes that the homogenized viscoelastic strain-to-stress map has the memory-kernel form the paper approximates.","marker":"[19]"},{"why":"Supplies the universal-approximation and bounded-variation compactness arguments for FNOs in elliptic homogenization that the proof adapts.","marker":"[5]"},{"why":"Defines Fourier Neural Mappings and proves the universal approximation result extended in Proposition C.2.","marker":"[27]"},{"why":"Provides the FNO universal approximation theorem used inside the proof of the FNM approximation result.","marker":"[31]"},{"why":"Motivates encoding material dependence in the differential-equation right-hand side and supplies the elasto-viscoplastic internal-variable cell problem.","marker":"[39]"},{"why":"Shows homogenization introduces memory in Kelvin-Voigt materials via semigroup theory.","marker":"[51]"},{"why":"Shows the homogenized memory kernel can be represented as a sum of exponentials, underpinning the piecewise-constant ODE reduction.","marker":"[55]"}],"fun_headline_variants":["Universal neural operator for viscoelastic constitutive laws","Neural operator learns memory-dependent material laws","From microstructure to stress: neural operator learns laws","No retraining: neural operator generalizes to new microstructures","Arbitrary accuracy: neural operator learns viscoelastic laws"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that making the neural-network approximation of the cell-problem dynamics arbitrarily accurate does not make it so sensitive to the internal state that small errors are amplified beyond control over the time interval; the paper does not establish this joint scaling.","fun_headline_variants_meta":{"raw":{"variants":["Universal neural operator for viscoelastic constitutive laws","Neural operator learns memory-dependent material laws","From microstructure to stress: neural operator learns laws","No retraining: neural operator generalizes to new microstructures","Arbitrary accuracy: neural operator learns viscoelastic laws"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001546,"raw_usage":{"total_tokens":6169,"prompt_tokens":918,"completion_tokens":5251,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":5175}},"tokens_in":534,"tokens_out":5251,"duration_ms":36365,"temperature":1.0,"reasoning_tokens":5175,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:14:36.193198+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the FNM-RNO on the piecewise-constant Kelvin-Voigt cell problem at decreasing error tolerances and measure the Lipschitz constant of the learned right-hand side with respect to the internal state; if that constant grows faster than the reciprocal of the training error, the Gronwall factor in the proof cannot be driven to zero and the uniform approximation bound of Theorem 4.7 would fail for long time horizons.","supporting_citations":[{"cited_title":"Bhattacharya, B","cited_arxiv_id":null,"evidence_quote":"Derives the Kelvin-Voigt cell problem, the piecewise-constant memory kernel form, and the weighted H1 norms used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that the homogenized viscoelastic strain-to-stress map has the memory-kernel form the paper approximates."},{"cited_title":"Bhattacharya, N","cited_arxiv_id":null,"evidence_quote":"Supplies the universal-approximation and bounded-variation compactness arguments for FNOs in elliptic homogenization that the proof adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Fourier Neural Mappings and proves the universal approximation result extended in Proposition C.2."},{"cited_title":"Kovachki, S","cited_arxiv_id":null,"evidence_quote":"Provides the FNO universal approximation theorem used inside the proof of the FNM approximation result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates encoding material dependence in the differential-equation right-hand side and supplies the elasto-viscoplastic internal-variable cell problem."},{"cited_title":"S´anchez-Palencia, Non-homogeneous media and vibration theory, Lecture Note in Physics, Springer-Verlag, 320 (1980), pp","cited_arxiv_id":null,"evidence_quote":"Shows homogenization introduces memory in Kelvin-Voigt materials via semigroup theory."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows the homogenized memory kernel can be represented as a sum of exponentials, underpinning the piecewise-constant ODE reduction."}],"review_version":1}