{"id":"b9467312-d5cc-4cf6-bb5e-1474c360670c","arxiv_id":"2506.10352","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HANO, an autoregressive Fourier neural operator with attention and U-Net branches, predicts path-dependent material stress from short observable strain-stress windows and stays accurate across loading resolutions and partial histories.","lead":"This paper introduces a neural network model, HANO, that predicts how materials respond to changing loads using only a short window of recent strain and stress measurements, without needing to know the material's internal state. It is designed for path-dependent materials like metals and composites, and aims to work reliably even when loading data are irregular, noisy, or start partway through a load history.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central hidden-variable-free claim relies on an unverified identifiability premise: a fixed k-step strain-stress window must uniquely encode the internal state; this is trivially true for the 1D benchmark but is not established for the 6D Hashin damage benchmark, where individual damage modes…","rationale":"I read the paper as an empirical proposal: HANO predicts the next stress autoregressively from a fixed window of observable strain-stress history, and the headline contributions are the 1D elastoplastic results, the partial-history robustness, and consistent outperformance of RNN/FNO/UFNO baselines. The reader's weakest assumption, that the finite window uniquely and sufficiently encodes internal state, is indeed the mathematical foundation of the whole approach. I partially agree with the reader: that assumption is genuinely load-bearing, but not uniformly. In the 1D benchmark, the linear kinematic hardening model makes the state identifiable from a single pair, so the k=10 window is more than sufficient. The concern becomes decisive for the 6D anisotropic damage benchmark, where individual Hashin damage mode variables are not directly recoverable from current stress-strain data and the paper provides no identifiability analysis. The empirical saturation in Figure 13 shows only that longer windows stop improving the trained model, not that the window-to-next-stress map is single-valued. The 21.09% NRMSE on that benchmark further suggests that the mapping may be imperfectly identified. I nonetheless give credit for the parameter-matched comparisons, the clear architecture description, and the strong 1D results, which are not threatened by this concern. The reader already returned CONDITIONAL, so I do not move the verdict; I would add the identifiability check on the Hashin benchmark as an explicit condition for the stronger generalization claims. The proposed test is concrete and feasible because the data generator is public and Appendix B/C specifies the model.","tokens_in":30515,"tokens_out":10438,"duration_ms":138909,"concrete_test":"Use the open Hashin data generator (or reimplement Appendix B/C) to run an identifiability test: for many random trajectories, collect input windows W = {epsilon_i, sigma_i}_{n-k+1}^n plus Delta_epsilon_{n+1}; cluster near-identical windows after normalization; for any cluster where the next stress sigma_{n+1} differs beyond a small tolerance, Equation (12) is violated. More directly, perform an adversarial optimization: fix a target next-strain increment, initialize two different damage states, and optimize their preceding strain histories so their k stress-strain windows coincide; if the resulting next stresses differ materially, the window does not encode the state. A pass with no large spread would settle the concern; a fail would require limiting the hidden-variable-free claim or adding state information.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (12) postulates that the next stress is a function of the last k strain-stress pairs and the next strain increment, and Remark 3.3 elevates this to a design principle. If this map is not single-valued, HANO is not well defined regardless of training. For the 1D elastoplastic benchmark the premise is actually safe: with linear kinematic hardening, sigma = E(epsilon - epsilon_p) and alpha = H*epsilon_p, so any single pair (epsilon, sigma) already determines the internal state. The load-bearing case is the anisotropic damage benchmark (Section 5). In the modified Hashin model the combined degradation factors d_f and d_m may be recoverable from current stress and strain, but future evolution depends on the individual mode variables d_FT, d_FC, d_MT, d_MC and their thresholds; the paper gives no argument that a window of k=20 observations disambiguates these states. The only evidence is the empirical plateau in Figure 13, which is a model-capacity and training observation, not an identifiability check. The high 21.09% NRMSE in Table 5 is consistent with an ill-posed mapping. Thus the central hidden-variable-free claim is not established for the more complex benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HANO, an autoregressive neural operator for data-driven constitutive modeling of path-dependent inelastic materials. HANO maps a short window of k recent strain–stress pairs plus the next strain increment to the next stress, using a Fourier neural operator backbone augmented by U-Net and hierarchical self-attention layers. The authors claim two principal advantages over RNN-based surrogates: discretization-invariance and independence from hidden state variables, the latter enabling accurate predictions from pre-stressed or partially observed histories. The method is tested on 1D elastoplasticity with kinematic hardening and on a 6D anisotropic damage dataset generated with a modified Hashin model. Reported NRMSEs are 0.6% on full 1D histories and 0.4% from truncated pre-stressed histories (vs. 35.9% for RNN1), and 21.09% on the 6D damage benchmark (vs. 35.23% for FNO and 24.67% for UFNO). Parameter-matched comparisons in Appendix D and Section 5.3 support the architectural conclusions.","tokens_in":30713,"tokens_out":4854,"duration_ms":60166,"significance":"If the central claims hold, HANO would be a valuable and practical surrogate: it avoids recurrent hidden-state initialization, degrades gracefully under missing or noisy history, and is backed by parameter-matched comparisons (Appendix D, Table 6) that separate architectural benefit from model capacity. The 1D elastoplasticity results are strong and the use of a publicly available damage dataset is a strength. However, the absolute 21.09% NRMSE on the damage benchmark materially tempers the abstract's 'high predictive accuracy' language, and the central hidden-variable-free premise is not rigorously established beyond the 1D case. The paper would be stronger if the identifiability assumption were checked and the discretization-invariance claims qualified.","major_comments":[{"comment":"The central claim that HANO 'eliminates the need for hidden state variables by directly utilizing short segments of observable strain-stress history' presumes that the map in Eq. (12) is single-valued: a k-step window of observable strain-stress pairs and the next strain increment must uniquely determine the next stress. This is safe for the 1D elastoplastic benchmark, where any pair (epsilon, sigma) with linear kinematic hardening fixes epsilon_p and the back stress. It is not established for the modified Hashin damage benchmark of Section 5, where future evolution depends on the individual mode damage variables d_FT, d_FC, d_MT, d_MC and their thresholds; the paper gives no argument or experiment showing that a window of k=20 (or even k=6, as suggested by Fig. 13) distinguishes different internal states with identical observable windows. The empirical error plateau in Fig. 13 is a training observation, not an identifiability check. I recommend either adding a concrete uniqueness test, e.g., searching generated trajectories for distinct internal states that share the same input window but lead to different future stresses and quantifying the discrepancy, or providing an observability argument for the damage model. As written, the hidden-variable-free design principle is load-bearing and unverified for the more complex benchmark.","section":"Section 3.3, Eq. (12), Remark 3.3"},{"comment":"The paper states that HANO 'inherently achieves discretization invariance' because the FNO filters index physical wavenumbers rather than grid points. However, the input to HANO is a fixed-length window of k observations. Consequently, the physical time span covered by the window scales with the sampling resolution, and the claim is unconditional only if the material's memory is exactly k steps regardless of the sampling rate. Section 5.4 itself acknowledges that the physical time span covered by the history window is the theoretically relevant quantity, citing Coleman-Gurtin. The experiments in 4.3.2 and 4.3.3 vary the number of increments per cycle over a moderate range (roughly 57-150 steps per cycle) while keeping k fixed, which does not separate the effect of k from the effect of physical time span. The claim should be qualified or supplemented by an experiment that varies the physical window length at fixed k to test whether HANO remains invariant, or by a theoretical condition on the required window length as a function of sampling density.","section":"Remark 3.2 and Section 4.3"},{"comment":"The abstract and conclusion characterize HANO as achieving 'high predictive accuracy' and 'accurately captures complex path-dependent responses,' but the best result on the anisotropic damage benchmark is NRMSE = 21.09% (Table 5). While HANO substantially outperforms FNO (35.23%) and UFNO (24.67%), a 21% normalized error is not high absolute accuracy, and the claim should be calibrated to the difficulty of the task. Please either soften the absolute-accuracy language to 'state-of-the-art relative to the tested baselines' or provide evidence that this level of error is appropriate for the target application.","section":"Abstract and Section 5, Table 5"},{"comment":"The abstract claims HANO 'naturally accommodates varying path discretizations and irregular sampling,' but no experiment in the manuscript uses irregular (non-uniform) time step sampling. Section 4.3 tests uniform resolutions of 60-150 steps per cycle and variable-resolution training data with a fixed uniform resolution per trajectory; Figure 7 shows 57, 83, 112, and 135 increments per cycle, all uniformly sampled within each cycle. The FNO backbone as implemented uses a 1-D FFT, which presumes a regular temporal grid; no mechanism for handling non-uniform sampling is described. Either add experiments with genuinely irregular temporal increments or remove the irregular-sampling claim from the abstract and introduction.","section":"Abstract and Section 4.3"}],"minor_comments":[{"comment":"The sentence 'HANO approximates the continuous mapping for the continuous operator mapping for the constitutive relation Gz' contains a duplicated phrase and should be edited to 'HANO approximates the continuous operator mapping for the constitutive relation Gz'.","section":"Section 3.3"},{"comment":"The appendix heading appears as 'T raining and V alidation Losses' with unintended spacing; it should be 'Training and Validation Losses'.","section":"Appendix E heading"},{"comment":"The text says the variable-resolution training paths have peak strain amplitudes from roughly 0.015 to 0.025, whereas Section 4.1 sets loading amplitudes to [0.008, 0.015]. Clarify whether these are different amplitude ranges or whether Figure 6's description is imprecise.","section":"Section 4.3.3 and Figure 6"},{"comment":"The text says accuracy 'plateaus around k=6,' but Figure 13 shows a gradual increase and fluctuations in NRMSE beyond k=6. Consider describing the trend as 'reaching a minimum near k=6 with mild fluctuations for larger k' rather than a plateau.","section":"Section 5.4"},{"comment":"The name 'HANOreduced' is written as both 'HANOreduced' and 'HANO reduced' in the same paragraph; use a single consistent notation.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical core is solid for the 1D elastoplasticity case and the architectural ablations are well designed. The main risk is the unverified identifiability premise behind the hidden-variable-free claim in the anisotropic damage benchmark, which is central to the paper's novelty. I would be supportive of a revised version that either proves or empirically tests uniqueness of the k-window map, qualifies the discretization-invariance and irregular-sampling claims, and softens the 'high predictive accuracy' language for the 21.09% damage result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The first: this is a competent, well-run application of FNO plus attention to history-dependent constitutive modeling, and the 1D elastoplastic results are genuinely good—0.6% NRMSE on full histories, 0.4% when starting mid-trajectory, versus roughly 36% for an RNN. The parameter-matched comparison in Appendix D makes that comparison fair. The second: the central claim—that a fixed window of strain–stress history eliminates the need for internal state variables—is safe for the 1D linear-hardening case (any single (ε, σ) pair fixes the state) but is not established for the 6D Hashin damage benchmark. The 21% NRMSE there, while better than FNO (35%) and UFNO (25%), is not a high-accuracy result, and no identifiability argument is given for the four damage variables.\n\nWhat's new: this is the first neural-operator treatment of history-dependent constitutive relations that I know of, and the architecture is sensible—Fourier layers for global temporal patterns, a U-Net path for local features, attention for reweighting history. The variable-resolution training experiment in Section 4.3.3 is a nice robustness check. The paper is also honest: it acknowledges the window-length limitation in Section 5.4 and the conclusion.\n\nSoft spots, in order of softness. The identifiability premise is the load-bearing one. For the 1D problem it's trivially true, but for the modified Hashin model the map from a 20-step window to the four damage variables is not analyzed. The plateau in Figure 13 is a training observation, not a proof of well-posedness. It could be fixed with a short numerical experiment—construct two different damage states that yield the same recent window and show HANO's prediction is ambiguous. Second, the discretization-invariance claim borrows FNO's theory, but FNO's invariance is for uniform grids and the window here is fixed-length; changing resolution changes which physical time span the window covers. The empirical results are encouraging, but the theory is stretched. Third, no code or data is actually released—the GitHub link is a promise, and the numbers are single-run with no seed variance reported.\n\nThis paper is for computational mechanics researchers building data-driven constitutive surrogates, and for people working on operator learning for time-series. It deserves a serious referee. The authors should be asked to release code and data, to address the identifiability question directly, and to either moderate the hidden-state-free claim or provide evidence that the window uniquely encodes the internal state.","headline":"HANO delivers strong 1D elastoplastic results and fair parameter-matched comparisons, but the hidden-state-free premise is only proven for the 1D case; the 6D damage benchmark (21% NRMSE) and the unexamined identifiability question temper the central claim.","tokens_in":31278,"tokens_out":3469,"would_cite":true,"duration_ms":41949,"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":"A history-aware neural operator predicts path-dependent material stress without hidden state variables.","keywords":["operator learning","neural operator","constitutive modeling","path-dependent materials","elastoplasticity","anisotropic damage","autoregressive forecasting"],"falsifier":"Generate a set of loading paths for the same material model in which two different internal histories (for example, different amounts of plastic strain and back stress) arrive at the same last k strain–stress pairs, then run the trained HANO forward and compare its next-stress predictions with the reference. If the predictions diverge substantially for such pairs, the fading-memory window assumption fails.","tokens_in":30250,"feed_emoji":"🧮","tokens_out":1560,"duration_ms":20374,"temperature":0.7,"pith_summary":"HANO learns a continuous operator that maps a short window of recent strain–stress observations, plus the next strain increment, to the next stress value. By conditioning directly on observable history rather than on a latent hidden state, it avoids the discretization and initial-state sensitivity problems of recurrent neural network surrogates. The paper demonstrates on elastoplasticity with kinematic hardening and on progressive anisotropic damage that this design maintains accuracy under variable resolution, partial loading histories, multi-cycle extrapolation, and noisy inputs, with errors far below RNN baselines.","feed_headline":"One neural operator predicts pre-stressed material stress at 0.4% error","feed_subtitle":"By feeding short strain–stress history windows instead of hidden states, HANO stays accurate even when loading history begins mid-path.","key_machinery":"The central object is the history-aware neural operator, built as a mapping $G^H_\\phi = Q \\circ (L^{AEUF}_M \\circ \\cdots \\circ L^{AEUF}_1) \\circ (L^F_L \\circ \\cdots \\circ L^F_1) \\circ P$ that acts on a function input $a(x)$ representing the strain–stress history window and the next strain increment. The Fourier layers index physical wavenumbers rather than grid points, which the paper argues yields discretization invariance; the U-Net and self-attention branches recover local high-frequency features and re-weight the history segments most relevant to the next stress. The autoregressive loop slides the window forward using the predicted stress, which is what allows prediction to start at an arbitrary intermediate point.","core_discovery":"The paper claims that a neural operator with Fourier layers, U-Net multiscale paths, and multi-stage self-attention can model path-dependent constitutive behavior autoregressively from a fixed window of the k most recent strain–stress pairs and the forthcoming strain increment, without needing internal state variables or initialization from an undeformed reference state. In the 1D elastoplastic benchmark, HANO achieves NRMSE of 0.6 percent on full histories and 0.4 percent on pre-stressed truncated histories, where RNN1 gives 3.0 and 35.9 percent respectively. In the 6D anisotropic damage benchmark, HANO achieves 21.09 percent NRMSE versus 35.23 percent for FNO and 24.67 percent for UFNO, with parameter-matched comparisons preserving the advantage.","pith_inferences":["The fading-memory assumption is not tested for identifiability: the paper finds k empirically but does not establish that the windowed observable states uniquely determine the internal state for kinematic hardening or Hashin-type damage, so adversarial loading paths that produce the same recent window from different internal states could break the mapping.","The operator's projection effect described for noisy inputs suggests a testable property: HANO should act as a low-pass filter on stress inputs, so injecting high-frequency perturbations at different frequencies should reveal a cutoff related to the Fourier mode truncation.","The method should extend naturally to rate-dependent materials, since the window input can carry strain-rate information implicitly through the spacing of strain values, but the paper does not validate this claim; a natural next experiment would be on viscoelastic or viscoplastic data.","A cautious reading of the 1D results is that the elastoplastic benchmark may be comparatively simple because the kinematic-hardening model has a compact memory determined by back stress and plastic strain, so the strong accuracy numbers may not transfer to materials with longer-memory or hidden microstructural state."],"forward_implications":["A surrogate trained this way can be initialized from a pre-stressed configuration by applying a few controlled loading increments and observing the response, rather than by replaying the full history from zero stress. This makes the model usable when early data are missing or when a finite element simulation begins from a previously loaded integration point.","Because the operator is resolution-invariant in principle, the same trained model can be applied to loading paths sampled at different rates, including irregular sampling, without retraining or re-calibration.","The autoregressive design with a short observable-history window removes the need for the user to define internal variables, yield functions, flow rules, or damage evolution equations, so the modeling effort shifts to choosing the window length k and generating representative training paths.","The reported performance suggests the trained surrogate could be embedded into differentiable solvers for gradient-based inverse analysis or real-time simulation, since the stress update uses only local windowed state plus strain increment.","The attention-enhanced U-Fourier layers, rather than raw model size, account for most of the accuracy gain over FNO and UFNO, so the architectural mechanism transfers to other operator backbones if the windowed-input autoregressive formulation is kept."],"supporting_citations":[{"why":"Supplies the Fourier neural operator backbone whose spectral filter indexed by physical wavenumbers is claimed to give discretization invariance.","marker":"[43]"},{"why":"Establishes that recurrent neural networks can predict path-dependent plasticity without explicit yield criteria, the baseline capability that HANO aims to improve.","marker":"[53]"},{"why":"Provides RNN1, the recurrent baseline used for comparison in the elastoplasticity study.","marker":"[22]"},{"why":"Documents the self-consistency (discretization-dependence) failure of recurrent constitutive models that HANO is designed to fix.","marker":"[6]"},{"why":"Gives the internal-state-variable thermodynamic foundation and the fading-memory principle that justifies using only recent windowed history.","marker":"[10]"},{"why":"Releases the open dataset of six-dimensional anisotropic damage trajectories used for the damage benchmark.","marker":"[18]"},{"why":"Defines the modified Hashin criteria that generate the damage benchmark data.","marker":"[24]"},{"why":"Provides the neural ODE framework whose continuous-time perspective inspires HANO's autoregressive operator formulation.","marker":"[9]"},{"why":"Introduces the U-Fourier layer combining Fourier and U-Net paths, the basis of the AEUF layer used in HANO.","marker":"[69]"}],"fun_headline_variants":["HANO: path-dependent material models from 0.4% error, no hidden states","Neural operator predicts pre-stressed stress from short history, no RNN drift","History-aware operator beats RNNs on inelastic paths with 0.4% error","HANO: Fourier operator models damage and elastoplasticity from history windows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that the most recent k strain–stress observations uniquely and sufficiently encode the current internal state of the material, so that predicting the next stress from just that window is well defined. If two different internal states can produce the same recent window, the prediction is ambiguous and the model's accuracy on such paths is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["HANO: path-dependent material models from 0.4% error, no hidden states","Neural operator predicts pre-stressed stress from short history, no RNN drift","History-aware operator beats RNNs on inelastic paths with 0.4% error","HANO: Fourier operator models damage and elastoplasticity from history windows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1327,"prompt_tokens":1010,"completion_tokens":317,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":228}},"tokens_in":626,"tokens_out":317,"duration_ms":4599,"temperature":1.0,"reasoning_tokens":228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:28:46.680102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a set of loading paths for the same material model in which two different internal histories (for example, different amounts of plastic strain and back stress) arrive at the same last k strain–stress pairs, then run the trained HANO forward and compare its next-stress predictions with the reference. If the predictions diverge substantially for such pairs, the fading-memory window assumption fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that recurrent neural networks can predict path-dependent plasticity without explicit yield criteria, the baseline capability that HANO aims to improve."},{"cited_title":"On the potential of recurrent neural networks for modeling path dependent plasticity","cited_arxiv_id":null,"evidence_quote":"Provides RNN1, the recurrent baseline used for comparison in the elastoplasticity study."},{"cited_title":"On the importance of self-consistency in recurrent neural network models representing elasto-plastic solids","cited_arxiv_id":null,"evidence_quote":"Documents the self-consistency (discretization-dependence) failure of recurrent constitutive models that HANO is designed to fix."},{"cited_title":"Thermodynamics with internal state variables","cited_arxiv_id":null,"evidence_quote":"Gives the internal-state-variable thermodynamic foundation and the fading-memory principle that justifies using only recent windowed history."},{"cited_title":"Data-driven constitutive models for brittle solids displaying progressive anisotropic damage","cited_arxiv_id":null,"evidence_quote":"Releases the open dataset of six-dimensional anisotropic damage trajectories used for the damage benchmark."},{"cited_title":"Failure criteria for unidirectional fiber composites","cited_arxiv_id":null,"evidence_quote":"Defines the modified Hashin criteria that generate the damage benchmark data."},{"cited_title":"Neural ordinary differential equations","cited_arxiv_id":null,"evidence_quote":"Provides the neural ODE framework whose continuous-time perspective inspires HANO's autoregressive operator formulation."},{"cited_title":"U-fno—an enhanced fourier neural operator-based deep-learning model for multiphase flow","cited_arxiv_id":null,"evidence_quote":"Introduces the U-Fourier layer combining Fourier and U-Net paths, the basis of the AEUF layer used in HANO."}],"review_version":1}