{"id":"311ce369-e175-4c33-ab6a-3dac6f371488","arxiv_id":"2506.16918","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A physics-guided POD-DeepONet surrogate predicts microscale displacements in viscoelastic composites with about 2-5% field errors and about 100x speedup over the reference FE solver.","lead":"The authors train a POD-DeepONet to replace the expensive microscale finite element solve in FE2 simulations of a viscoelastic fiber composite, then reconstruct strains, internal variables, and stresses from the predicted displacements using physical equations. A smart generalist might read it because it is a concrete test of whether physics-informed operator learning can accelerate rate-dependent multiscale materials simulation by about 100x.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Agrees with the reader: the model never enforces balance of linear momentum (Eq. 2), and the reported accuracy is demonstrated only on the sinusoidal, fixed-frequency, fixed-step test family of Eq.","rationale":"The reader's weakest assumption correctly identifies the two linked vulnerabilities in the strongest claim: (i) the architecture never enforces the balance of linear momentum, and (ii) the error statistics are produced on test inputs drawn from the same sinusoidal family as the training inputs. Both are load-bearing because the stated goal is use as a microscale surrogate inside FE2, where the driving strains are determined by a macroscale Newton iteration and are not restricted to smooth single-frequency sinusoids with constant time steps. If the network has not learned equilibrium well enough to generalize, the homogenized stress used by the macroscale will be inaccurate even if the in-distribution reconstruction errors are small. The physical post-processing of internal variables is credible and is a genuine strength: Eq. (19) is a standard viscoelastic evolution equation, and solving it with the predicted strain field is a reasonable way to carry history forward. The released training code and dataset are also positive evidence, although the absence of evaluation routines limits independent verification of the reported numbers. I do not see a reason to reject the paper: the in-distribution results are plausible, the constitutive model is standard, and the conditional verdict already demands exactly the additional checks that would resolve the concern. The single most decisive check is to evaluate the released model on out-of-distribution loading histories with variable time steps; if that test passes, the equilibrium concern can be addressed by a residual check and the claim is substantially stronger. If it fails, the reported accuracy and speedup are properly interpreted as restricted to sinusoidal, fixed-step inputs.","tokens_in":9924,"tokens_out":5478,"duration_ms":73876,"concrete_test":"Download the released GitHub code and Zenodo dataset, and evaluate the trained model on 100 new loading histories that are random walks (or superpositions of two sine waves) within the same +/-0.04 strain amplitude but with time steps drawn uniformly from Delta t = 0.05-0.2 s. Recompute the fraction of test trajectories whose homogenized-stress relative error, measured by Eq. (22), is below 6%. If that fraction is materially below the reported 94%, or if the median error is substantially larger than on the sinusoidal test family, the central claim must be restricted to the tested input family until equilibrium-residual and out-of-distribution checks are supplied.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a POD-DeepONet with physics-based post-processing can replace the microscale FE solve in an FE2 framework for rate-dependent viscoelastic materials. For that claim to hold, the predicted displacement field must satisfy, at least approximately, balance of linear momentum (Eq. 2) whenever the model is queried. The training loss (Eq. 12) is a pure regression on POD coefficients; no equilibrium residual enters the loss, and the subsequent kinematic, constitutive, and internal-variable updates (Eqs. 13-20) do not restore equilibrium. The numerical evidence is also confined to inputs of the form Eq. (21): sinusoidal macroscale strain with fixed frequency omega = 2 pi x 0.05 s^-1, fixed Delta t = 0.1 s, and amplitudes in U(-0.04, 0.04). Training and test sets are drawn from this same family, so the reported 94%-below-6% homogenized-stress statistic and the ~100x speedup do not establish behavior for the non-sinusoidal, variable-time-step strain histories that a coupled FE2 macroscale solver would produce, including during Newton iterations. The manuscript itself lists arbitrary trajectories and variable time steps as future work, which is an explicit limitation. This is a validation/generalization gap rather than an internal inconsistency, and the withheld evaluation code makes independent reproduction of the headline error statistics impossible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid microscale surrogate for FE2 multiscale simulations of rate-dependent viscoelastic composites. A POD-DeepONet maps the current macroscale strain and a low-dimensional latent representation of the internal variables to POD coefficients of the microscale displacement field; the microscale strain, internal-variable evolution, and stress are then computed from the kinematic relation, the viscoelastic constitutive ODE, and the constitutive law, respectively. The model is trained on 1000 finite-element RVE trajectories with sinusoidal macroscale strain histories (Eq. 21), and evaluated on 100 held-out trajectories. The authors report displacement, strain, internal-variable, and stress field errors in the 2--6% range, with about 94% of test trajectories having homogenized-stress error below 6%, and a roughly 100x speedup relative to a Fortran RVE solver.","tokens_in":10187,"tokens_out":4083,"duration_ms":50489,"significance":"If the claims hold, the paper makes a useful contribution to hybrid machine-learning-based computational homogenization: it combines a data-driven displacement operator with physics-based constitutive and internal-variable updates, stores internal variables in a low-dimensional latent form, and demonstrates the approach on a nonlinear rate-dependent material with a small training set. The release of training code and data on Zenodo is a strength for reproducibility, although the evaluation code is withheld. I agree with the reader that the training procedure is standard supervised learning and not circular. However, the central claim that the model can replace the microscale FE solve in an FE2 framework is not yet established: equilibrium is not enforced or checked, and the numerical evidence is confined to the sinusoidal, fixed-frequency, fixed-time-step family on which the model was trained.","major_comments":[{"comment":"The balance of linear momentum, Eq. (2), is the governing equation for the microscale problem, but it does not enter the training loss (Eq. 12) and it is not restored by the post-processing steps (Eqs. 13--20). The predicted displacement field is therefore not guaranteed to satisfy equilibrium, and no equilibrium residual is reported. Since the claimed application is replacement of the microscale FE solve in FE2, this is a load-bearing validation gap. I recommend reporting the divergence of the predicted stress field on test trajectories (e.g., mean and maximum residual relative to the FE reference) or adding a residual-based term to the loss; without one of these, the surrogate's stress predictions are only regression outputs, not equilibrium solutions.","section":"Section 2.2, Eqs. (2), (10)--(14), and Section 3.3"},{"comment":"All training and test inputs are sinusoidal macroscale strain histories with amplitude A sampled from U(-0.04, 0.04), fixed angular frequency omega = 2*pi*0.05 s^-1, and fixed time step Delta t = 0.1 s. A coupled FE2 macroscale solver would generate non-sinusoidal strain histories, variable time-step increments, and trial increments during Newton iterations, all outside the tested distribution. The reported 94%-below-6% homogenized-stress statistic and the field errors therefore do not establish performance in an actual FE2 simulation. The manuscript itself lists arbitrary trajectories and variable time steps as future work, which is an explicit limitation; please either add tests on non-sinusoidal and variable-step histories or revise the claims to state that the surrogate is currently validated only on the training-family inputs.","section":"Section 3.2, Eq. (21), and Section 3.3"},{"comment":"The claimed ~100x speedup (0.59 s vs 53 s per trajectory) is not sufficiently documented to be verifiable. The comparison appears to be between a GPU-based Python implementation and a Fortran solver run with 16 MPI processes and 1 OpenMP thread per process, but the hardware, compiler flags, timing methodology, number of repeated runs, and whether the time includes the full 100-step trajectory or a shorter sequence are not stated. The computational cost of training and POD construction is also omitted. Please report a more controlled timing study with the same physical problem and comparable computational resources, or at minimum provide full wall-clock details so the speedup claim can be evaluated.","section":"Section 3.3, speedup comparison"}],"minor_comments":[{"comment":"The text states that 100 test trajectories 'each consisting of ten time steps' were analyzed, whereas the training data consist of trajectories of 100 time steps. If the test trajectories were truncated to 10 steps, the reported errors do not reflect recursive accumulation over the full temporal horizon; if this is a typo, please correct it.","section":"Section 3.3"},{"comment":"The notation is overloaded: s denotes both the total number of snapshots in Eq. (7) and the number of samples in the loss function (12), and the subscript of a_i,n in Eq. (8) conflicts with the vector a_n defined in Eq. (9). Please clarify.","section":"Section 2.2, Eqs. (8)--(12)"},{"comment":"The definition of the homogenized-stress error used in Figure 6 is not given; please state the exact metric (e.g., relative l2 norm over the trajectory, per-component or combined).","section":"Section 3.3, Figure 6"},{"comment":"The statement that evaluation and visualization routines are 'reserved for future work' prevents independent verification of the headline error statistics. I encourage the authors to release the evaluation code alongside the training code and data.","section":"Code Availability"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis one is worth reading but not a breakthrough. The authors extend their Equino hybrid operator framework to rate-dependent viscoelastic composites by adding POD-latent internal variables and physics-based post-processing. That specific combination is new, and the paper does it cleanly: the network predicts displacement POD coefficients, then kinematic relations, constitutive laws, and the ODE for viscous strains are applied at the integration points. This keeps the model flexible and physically interpretable. The results on a 2D fiber-matrix RVE show mean relative errors around 2–5% for displacements, strains, stresses, and internal variables, with about 94% of test trajectories below 6% homogenized-stress error. That is plausible for the tested distribution, and the ~100x speedup versus a Fortran solver is in the right ballpark.\n\nThe soft spots are clear. The training loss is pure regression on POD coefficients; there is no equilibrium residual from Eq. (2) in the objective or in the post-processing. So balance of linear momentum is only implicitly learned from data. The test inputs are the same family as training: sinusoidal strain with fixed frequency and fixed time step (Eq. 21). The paper itself lists arbitrary trajectories and variable time steps as future work, which is an honest limitation but undermines the abstract's unqualified \"<6% error\" claim. The speedup comparison is under-documented; we only get 0.59 s vs 53 s without wall-clock details for the neural model or how the Fortran time was measured. And while code and data are on GitHub/Zenodo, the evaluation routines are withheld, so independent reproduction is not currently possible.\n\nThese are standard concerns for this kind of work, not fatal flaws. The central approach is coherent and the errors on in-distribution paths are believable. The paper would benefit from equilibrium residual checks, out-of-distribution tests, a clearer benchmark setup, and full evaluation code. The \"first\" claim is a bit strong, but the specific architecture is a legitimate new application.\n\nWho should read it: anyone building neural operator surrogates for history-dependent multiscale mechanics, especially with FE2 integration in mind. I'd take a look myself, and I'd send it to review—conditional acceptance with those revisions, not rejection. Serious referee time is warranted.","headline":"Solid incremental step for neural-operator surrogates in rate-dependent FE2, but the equilibrium and generalization claims are not as strong as the abstract implies.","tokens_in":10702,"tokens_out":2629,"would_cite":true,"duration_ms":28799,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that a POD-DeepONet hybrid can replace the microscale finite-element solve inside FE2 for rate-dependent viscoelastic composites, keeping homogenized stresses under 6% error on about 94% of held-out trajectories while…","keywords":["operator learning","POD-DeepONet","multiscale simulation","FE2","viscoelasticity","internal variables","computational homogenization","rate-dependent materials"],"falsifier":"Feed the trained hybrid microscale model strain histories produced by a coupled macroscale finite-element computation, including non-sinusoidal loading, unloading, and rate changes, and compare homogenized stresses and internal-variable states against full FE2 reference solutions over many time steps; if the errors exceed the reported 6% or accumulate with recursion, the central claim is falsified.","tokens_in":9699,"feed_emoji":"⚙️","tokens_out":7920,"duration_ms":86533,"temperature":0.7,"pith_summary":"The paper tries to establish that the expensive microscale finite-element solve in an FE2 multiscale simulation can be replaced by a hybrid neural-operator model without losing physical consistency for rate-dependent materials. The surrogate combines a POD-DeepONet that predicts displacement coefficients from the macroscopic strain and a compressed representation of internal variables with explicit kinematic, constitutive, and evolution equations that reconstruct the microscale strain, stress, and internal-variable fields. On unseen loading trajectories it reports displacement errors around 2.3%, strain errors around 3%, internal-variable errors around 4.5 to 5%, and homogenized stress errors below 6% for roughly 94% of test trajectories. A trajectory that takes about 53 seconds with a conventional Fortran microscale solver is evaluated in about 0.59 seconds with the hybrid model. If the claim holds, the principal computational bottleneck of FE2 becomes cheap enough to make concurrent nonlinear multiscale simulation practical.","feed_headline":"Neural surrogate runs microscale FE2 solves 100x faster, <6% error","feed_subtitle":"Keeps internal variables physical while cutting a trajectory's cost from 53 s to 0.59 s.","key_machinery":"The load-bearing object is a POD-DeepONet with a latent internal-variable loop. A branch network maps the pair of macroscale strain and projected internal-variable coefficients to coefficients of a displacement field expanded in POD modes, with the trunk network replaced by precomputed POD bases. The displacement is then differentiated through the kinematic relation projected onto the modes, and the evolution equation for viscous strain, the internal variable, plus the viscoelastic constitutive law are solved as physical post-processing, so the model is not a black-box stress predictor. The POD truncation, using 16 modes for both the displacement and internal-variable fields, sets the latent dimension and the reconstruction error floor.","core_discovery":"The central claim is that a POD-DeepONet can serve as the microscale constitutive engine in a hybrid model for rate-dependent materials: it learns a map from the current macroscale strain and the latent projection of internal variables at time $t_n$ to the displacement field at $t_{n+1}$, while the strain is obtained by applying the linearized kinematic relation to the POD modes, the internal variables are advanced by solving their thermodynamically consistent evolution ordinary differential equation, and the stresses follow from the viscoelastic constitutive law. Because the internal variables are represented by their coefficients on POD modes, the high-dimensional state is compressed and updated recursively. The paper reports mean relative $\\ell^2$ errors on held-out trajectories of about 2.3% for displacements, about 3% for strains, about 4.5 to 5% for internal variables, and about 3 to 5.5% for stresses, with homogenized stresses below 6% error for about 94% of test trajectories. The authors describe this as the first hybrid model to integrate internal variables using a combination of surrogate and assistant strategies within a multiscale framework.","pith_inferences":["Going beyond the paper: a coupled macroscale solver will produce strain histories that are not single-frequency sinusoids, so the reported 6% bound would need to be revalidated for out-of-distribution paths before relying on the speedup in full FE2.","Going beyond the paper: the tests cover a limited recursive window, so long rollouts could accumulate error in the latent internal-variable coefficients; re-projecting onto reference finite-element states or adding an equilibrium residual term would be a natural safeguard.","Going beyond the paper: the low-dimensional internal-variable coefficients could be exploited as compressed memory in large structural simulations, a benefit the authors mention but do not quantify."],"forward_implications":["Within the tested family of sinusoidal loading paths, homogenized stress error stays below 6% for about 94% of unseen trajectories, so the surrogate can replace microscale finite-element solves for that input distribution.","Per-trajectory cost drops from about 53 seconds with the reference Fortran solver to about 0.59 seconds, roughly a 100x speedup, making concurrent nonlinear multiscale simulation more practical.","Internal variables are stored as projection coefficients on POD modes instead of full fields, reducing the memory footprint at every macroscale integration point.","The framework is modular with respect to the constitutive law: the same kinematic, constitutive, and evolution pipeline could host other material models, including data-driven ones.","The methodology is demonstrated in two dimensions but the authors state it generalizes to three dimensions."],"supporting_citations":[{"why":"Defines the DeepONet operator-learning architecture whose branch network is preserved in the POD-DeepONet used here.","marker":"[17]"},{"why":"Introduces POD-DeepONet, the architecture with POD bases replacing the trunk network, and provides the benchmarks motivating its use.","marker":"[18]"},{"why":"Proposes the prior hybrid physics-informed neural-operator approach for quasi-static multiscale problems that this work extends to rate-dependent viscoelasticity.","marker":"[6]"},{"why":"Establishes the FE2 concurrent multiscale method whose expensive microscale solves are being replaced.","marker":"[7]"},{"why":"Supplies the Latin Hypercube Sampling method used to generate the training macroscale strain trajectories.","marker":"[19]"},{"why":"Provides the reduction to a scalar nonlinear equation and the Pegasus iteration used to solve the internal-variable evolution equation at integration points.","marker":"[8]"}],"fun_headline_variants":["Neural hybrid model speeds rate-dependent multiscale simulations 100x","POD-DeepONet microsurrogate: 100x speedup, <6% error","Physics-guided neural micro engine: 100x faster, <6% error","Hybrid neural-physics microscale: 100x speedup for viscoelasticity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model never enforces the physical equilibrium condition, the balance of linear momentum, in its architecture; it assumes the neural network learned that condition from the training data, and the only tests use sinusoidal macroscale strains of fixed frequency with amplitudes inside a narrow range.","fun_headline_variants_meta":{"raw":{"variants":["Neural hybrid model speeds rate-dependent multiscale simulations 100x","POD-DeepONet microsurrogate: 100x speedup, <6% error","Physics-guided neural micro engine: 100x faster, <6% error","Hybrid neural-physics microscale: 100x speedup for viscoelasticity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2166,"prompt_tokens":1002,"completion_tokens":1164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1086}},"tokens_in":618,"tokens_out":1164,"duration_ms":10696,"temperature":1.0,"reasoning_tokens":1086,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:37:06.624099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the trained hybrid microscale model strain histories produced by a coupled macroscale finite-element computation, including non-sinusoidal loading, unloading, and rate changes, and compare homogenized stresses and internal-variable states against full FE2 reference solutions over many time steps; if the errors exceed the reported 6% or accumulate with recursion, the central claim is falsified.","supporting_citations":[{"cited_title":"Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators","cited_arxiv_id":null,"evidence_quote":"Defines the DeepONet operator-learning architecture whose branch network is preserved in the POD-DeepONet used here."},{"cited_title":"Equino: A physics-informed neural operator for multiscale simulations.arXiv preprint arXiv:2504.07976, 2025","cited_arxiv_id":null,"evidence_quote":"Proposes the prior hybrid physics-informed neural-operator approach for quasi-static multiscale problems that this work extends to rate-dependent viscoelasticity."},{"cited_title":"Multiscale FE2 elastoviscoplastic analysis of composite structures","cited_arxiv_id":null,"evidence_quote":"Establishes the FE2 concurrent multiscale method whose expensive microscale solves are being replaced."},{"cited_title":"A remark on the application of the Newton-Raphson method in non- linear finite element analysis","cited_arxiv_id":null,"evidence_quote":"Provides the reduction to a scalar nonlinear equation and the Pegasus iteration used to solve the internal-variable evolution equation at integration points."}],"review_version":1}