{"id":"1ffa051e-5587-43aa-9367-75d7c02b3aff","arxiv_id":"2412.02864","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper derives an adjoint-based PDE-constrained optimization method to identify elasto-viscoplastic constitutive parameters from full-field displacement and force data, validated only on synthetic noise-free experiments.","lead":"This paper presents a computational method to infer a material's stress-strain behavior from full-field measurements like digital image correlation, by solving an optimization problem constrained by the balance laws of mechanics. The method uses adjoint-based gradients so the cost grows only linearly with the number of material parameters, which could make it practical for complex models including neural networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-uniqueness inside the assumed model family: recovered parameters differ from ground truth by up to 125% while the objective drops to ~1e-9, so matching the calibrated tests does not by itself identify the constitutive relation.","rationale":"The reader's weakest assumption was model-form error: the true material might lie outside the chosen parametric family. My concern is closely related but more specific and is supported by the paper's own results: even when the data are generated inside the model family and without noise, the recovered parameters are far from the generating parameters even though the calibration objective is minuscule. For the dynamic example, m is recovered as 4.50 versus the true 2.00, and the independent uniaxial test still shows 2-8% RMS error; for the quasistatic example, epsilon_p0 is off by roughly 75% while O reaches 5.59e-9. This means the map from parameters to the observed data has a large near-flat valley, so the optimization problem is ill-posed and the calibrated model is not certified for unseen loading conditions. The authors are honest about parameter non-uniqueness and provide a zero-shot uniaxial validation, which is real evidence, but it covers only one loading path and does not establish that the recovered constitutive relation generalizes to the complex, multi-axial paths the method is intended to probe. I do not recommend rejecting the paper: the adjoint derivation appears standard, the mesh sensitivity studies are useful, and the method clearly can fit the calibration data. However, the central claim that the method identifies the constitutive relation should be conditioned on a demonstration of predictive generalization beyond the calibrated experiments, ideally with noise and model-form error. This is essentially the same condition the reader imposed, so the verdict remains CONDITIONAL; my read does not move it.","tokens_in":21272,"tokens_out":9150,"duration_ms":134164,"concrete_test":"Simulate a hold-out experiment not used in the inversion, for example a non-proportional loading path (e.g., compression followed by shear) or a cyclic loading test, using the generating parameters P_gen and the recovered parameters P_QS / P_DC from Tables 1-3. Compare the full force-displacement or stress-strain responses and compute an Oind-style relative RMS error. If this hold-out error exceeds a few percent (say 5%), the recovered constitutive relation does not generalize despite fitting the calibration data, confirming the concern; if it remains near the calibration-level error, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the method identifies constitutive behavior from full-field observations. The results in Tables 1-3 show that the parameter-to-observation map is highly non-injective even in the noise-free, in-family synthetic setting. For the dynamic example, the recovered rate exponent m = 4.50 versus the generating m = 2.00 and epsilon_p0 differs by roughly 27%, yet the optimization objective falls by about three orders of magnitude. For the quasistatic example, epsilon_p0 differs by roughly 75% while the objective reaches 5.59e-9. The authors acknowledge this degeneracy and add a zero-shot uniaxial test, but that test checks only one loading path and still yields 2-8% RMS errors in the dynamic case. Thus the load-bearing assumption is not merely that the true material lies in the Perzyna family; it is that the chosen observations select a unique, or at least predictively equivalent, constitutive response. No identifiability analysis or richer hold-out validation is provided. If different parameter vectors that fit the calibration data produce materially different predictions on untested paths, the method has not learned the constitutive relation but has only fit one member of an equivalence class.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a PDE-constrained optimization framework for identifying constitutive parameters of history-dependent materials from full-field displacement measurements and macroscopic force histories. The forward problem is a boundary value problem for a J2 elasto-viscoplastic material with power-law hardening and rate dependence; the sensitivity of the objective is computed with an adjoint method, linear in space and quasilinear in time, and the parameters are updated with a gradient-based optimizer. The method is demonstrated on two synthetic benchmarks: a quasistatic compression of a plate with a hole using only the final displacement snapshot and force history, and a dynamic compression of a thin annular specimen reduced to a two-dimensional problem. The authors report that the recovered parameters differ from the generating parameters even though the training objective is very small, and they introduce a zero-shot uniaxial stress-strain test to show predictive equivalence. They conclude that the method is accurate and efficient, scales nearly independently of the number of parameters, and is suited to hyperparameterized constitutive models.","tokens_in":21553,"tokens_out":9821,"duration_ms":108887,"significance":"If the central claims are established, this would be a valuable contribution to full-field inverse material characterization: a systematic adjoint formulation for a non-smooth, history-dependent constitutive model, with explicit numerical schemes in the appendices and careful robustness studies over initial guesses, objective weights, mesh size, and specimen geometry. The authors deserve credit for acknowledging the degeneracy of the parameter-to-observation map and for testing predictive equivalence with an independent uniaxial objective. The scalability claim, if supported, would motivate extensions to neural-network constitutive models. However, the reported parameter non-uniqueness and the absence of noise and model-form validation mean the paper currently demonstrates a useful parameter-fitting framework rather than the stronger claim of learning constitutive relations from experiments.","major_comments":[{"comment":"The claim that the method 'accurately recovers elasto-viscoplastic material parameters' is not supported by the reported results. In the quasistatic example the recovered εp0 is 0.0393 versus the generating value 0.0225 (about 75% error) while the objective is 5.59×10⁻⁹; in the dynamic example the recovered m is 4.50 versus 2.00 and εp0 is 0.0286 versus 0.0225 while the objective is 1.44×10⁻⁴. Table 3 shows that different initial guesses converge to materially different parameter vectors (m from 1.09 to 4.50, εp0 from 0.0277 to 0.0781) with final objectives spanning 0.87×10⁻⁴ to 8.57×10⁻⁴. The independent uniaxial test, Eq. (21), gives Oind = 0.0069 for the quasistatic case and 0.021 for the dynamic case and probes only a single loading path. Thus the paper has not established that the optimized parameters identify the constitutive relation; it has shown that the calibration data admit an equivalence class of parameters. The authors acknowledge this degeneracy, but the central claim of learning constitutive relations requires either an identifiability/observability analysis of the parameter-to-observation map or a multi-path, multi-axial hold-out validation demonstrating that the recovered class members predict equivalent responses outside the calibration set.","section":"§3, Tables 1–3; §3.1 concluding sentence"},{"comment":"The cost model Cost(NP) = A + B NP with A >> B is asserted rather than demonstrated. The discussion omits the handling of the forward trajectory during the backward adjoint solve: the adjoint equations use forward fields at each time step, and neither a storage nor a recomputation strategy is described. All demonstrations use only five parameters, so the claimed near-independence of the number of parameters—which is the paper's motivation for neural-network extensions—is not verified. I recommend adding a scaling experiment with larger NP (for example using a smooth surrogate problem) or providing a more precise analysis of the constants in Eq. (11) including trajectory management and the cost of the parameter update.","section":"§2.1.4, Eq. (11)"},{"comment":"The synthetic data are generated by the same forward solver and the same material model used in the inversion, with no noise, and the dynamic experiment is reduced by the uniform-axial-strain assumption (Eq. (23)). This setup cannot reveal whether the method tolerates measurement noise, DIC interpolation error, or model-form error, all of which are central to accurate identification from experiments. The dynamic test is self-validating in that both data generation and inversion use Eq. (51). The authors should either add a noise robustness study (for example, perturbing uexp and f_R at levels representative of DIC) and a model-form-error test (for example, generating data with a slightly different hardening law), or explicitly scope the claim to noise-free, in-family synthetic benchmarks.","section":"§3.1 and §3.2 validation setup"}],"minor_comments":[{"comment":"The text states that the objective remains on the order of 10⁻⁹ after the initial-guess sensitivity study, but Table 3 reports final objectives around 10⁻⁴; the text and table should be made consistent.","section":"§3.2, Sensitivity to initial guess"},{"comment":"Equation (18) contains a typographical artifact: '= − ∂o/∂u · on Ω' should read '= −∂o/∂u on Ω'.","section":"Eq. (18)"},{"comment":"The caption of Figure 3(f) says 'the rsults of the independent stress-strain test'; this should be corrected to 'results'.","section":"Figure 3 caption"},{"comment":"The definition of Oind integrates over strain, but the strain interval is not specified; adding the integration range would make the reported numbers reproducible.","section":"Eq. (21)"},{"comment":"The adjoint formulation for elasto-viscoplasticity is presented for the smooth flow rule, but the treatment of the elastic unloading/yield boundary is not discussed; a sentence describing how the active plastic region is handled in the adjoint evolution would improve the reproducibility of the method.","section":"§2.2.2 and Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid methodology contribution for a solid-mechanics journal, and the promised Parts 2 and 3 could address experimental data and neural-network representations. The key issue is that the paper's own results demonstrate substantial parameter non-uniqueness even in the noise-free, in-family setting, so the wording of the central claim ('accurately recovers material parameters', 'learning constitutive relations') goes beyond what is shown. I would not require experimental data in this Part 1, but I would require either a quantitative identifiability analysis or a significantly softened claim, plus a brief treatment of noise/model-form error in the synthetic validation. The scalability claim in Eq. (11) also needs at least a scaling experiment or a more careful cost analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new piece is the adjoint sensitivity derivation for a general history-dependent internal-variable constitutive model, carried through a J2 elasto-viscoplastic law with yield, together with a careful finite-element implementation for both quasistatic and dynamic problems. Prior adjoint-based parameter identification exists for elasticity, viscoelasticity, and Norton-Hoff viscoplasticity, but the general formulation and the treatment of the yield surface here are worked out in more detail than anything I've seen in one place. The authors are also unusually honest: they report that recovered parameters differ substantially from the generating parameters (e.g., m = 4.50 vs 2.00 in the dynamic case) even when the objective drops to 1e-9, and they add a zero-shot uniaxial check. That transparency is real credit.\n\nThe soft spots are the ones the stress-test flags. All validation is synthetic, noise-free, and generated by the same forward solver used in the inversion, so inverse crime is in play. More importantly, the parameter-to-observation map is demonstrably non-injective inside the assumed Perzyna family: different parameter vectors fit the calibration data almost perfectly yet differ by tens of percent in individual parameters. The zero-shot uniaxial test is a single loading path and does not discriminate between the equivalence-class members. So the paper shows the method can fit one member of an equivalence class, not that it can identify the constitutive relation. The authors acknowledge this degeneracy but do not analyze identifiability or test hold-out predictions on more diverse loading paths. That gap is load-bearing for the title's promise.\n\nThe scaling claim (cost ~ A + B*N_P) is plausible and is the main practical motivation, but it is not demonstrated here beyond the small parameter set; neural-network scaling is deferred to Part 3. No code is provided, which makes the numerical details harder to verify independently.\n\nBottom line: this is a solid, well-executed methods paper with a careful derivation and honest reporting. Its central claim is narrower than the title suggests, and the validation is too clean. A serious referee should engage with it; the authors should be pushed on noise robustness, identifiability, and at least one independently generated or experimental data set. I'd send it to review.","headline":"A competent adjoint-based parameter identification paper with an honest treatment of non-uniqueness, but the validation is too clean to support the 'learning constitutive relations' framing.","tokens_in":22025,"tokens_out":2778,"would_cite":true,"duration_ms":29287,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74C10","74S05","65N21"],"pacs":[],"model":"deepseek-v4-flash","headline":"Constitutive behavior can be identified from full-field experiments by solving a PDE-constrained inverse problem whose adjoint gradient costs about one extra forward solve per iteration, nearly independent of the number of material…","keywords":["constitutive relation identification","PDE-constrained optimization","adjoint method","elasto-viscoplasticity","full-field measurements","inverse problems","digital image correlation","internal variables"],"falsifier":"Run the identical inversion on data generated by a materially different constitutive law, such as a different hardening form or a model with damage or softening, or on real experimental data with known measurement noise, and then test the recovered model on an independent non-proportional loading path not used in the objective. A concrete version: add roughly one percent noise to the synthetic displacements and reaction forces in the quasistatic plate-with-hole test and check whether the zero-shot uniaxial error stays below a few percent.","tokens_in":1799,"feed_emoji":"🔬","tokens_out":2393,"duration_ms":81931,"temperature":0.7,"pith_summary":"Learning a material's constitutive relation is an indirect inverse problem: measure displacements and forces from an experiment, then find the stress-strain law that makes a simulation reproduce those observations. This paper shows that the problem can be posed as a PDE-constrained optimization, with the balance laws as constraints, and that the adjoint method supplies the parameter gradient at the cost of roughly one extra forward solve. The central scaling claim is Eq. (11): each optimization iteration costs A + B*NP with A >> B, so the cost barely grows with the number of parameters, unlike gradient-free or finite-difference approaches. The method is demonstrated on synthetic elasto-viscoplastic data from a quasistatic plate-with-hole compression and a dynamic impact test; the recovered parameters reproduce the observations and pass an independent uniaxial stress-strain test, even though some individual hardening parameters differ from the generating values. If this holds, constitutive relations could be identified from complex, heterogeneous experiments rather than idealized uniform tests, and the approach extends to hyperparameterized models such as neural networks.","feed_headline":"Adjoint method recovers material laws from full-field experiments","feed_subtitle":"Per-iteration cost barely depends on the number of parameters, opening the door to neural-network constitutive models.","key_machinery":"The machinery is the adjoint of the initial-boundary value problem that governs the experiment. The objective is augmented with the weak form of the balance laws and internal-variable evolution (Eq. 5), and the adjoint states are chosen to satisfy a backward-in-time system (Eqs. 10 and 18) that removes all terms containing unknown solution sensitivities. This leaves the parameter gradient (Eqs. 9 and 17) as a single quadrature, so each gradient evaluation costs one forward solve plus one adjoint solve, independent of the parameter count. The parameter update is handled by the Method of Moving Asymptotes, and the forward-adjoint pair is discretized with P1 finite elements in space with quadrature-point plastic variables.","core_discovery":"The paper claims that full-field experimental observations, combined with the balance laws, carry enough information to identify the constitutive behavior of an inelastic material through a PDE-constrained optimization. The forward problem is the initial-boundary value problem of the experiment; the objective compares computed and measured displacements and reaction forces; and the adjoint method yields the parameter gradient without forming solution sensitivities. For a general Perzyna-type internal variable theory, the adjoint system (10) is linear in space and quasilinear in time, solved backward in time, and for J2 elasto-viscoplasticity with power-law hardening the sensitivity reduces to (17). Two synthetic demonstrations, one quasistatic and one dynamic, drive the objective down by several orders of magnitude from a poor initial guess, show insensitivity to initial guess, objective weights, and mesh resolution, and yield parameters that match the yield strength while differing in hardening parameters yet agreeing on independent uniaxial response. The paper interprets this as observational degeneracy: the method identifies models that reproduce the measured response, not necessarily the underlying parameter values.","pith_inferences":["If the per-iteration cost really is dominated by the forward and adjoint solves, then the practical bottleneck for neural-network parameterizations will be backpropagating through the adjoint update rather than the raw number of network weights; the announced part 3 of this series is where that claim gets tested.","The validation uses noise-free synthetic data generated by the same forward solver used in the inversion, so real-world applicability hinges on tolerance to measurement noise and model-form error; a natural companion study would contaminate synthetic data with DIC-like noise and test an out-of-family material law.","The quasistatic result that the final snapshot plus force history is enough suggests specimen geometry can be engineered to maximize information content, for example asymmetric holes or multi-rate loading paths, to shorten experiments or reduce the number of tests needed.","The observed non-uniqueness implies that reporting a single best-fit parameter vector will understate uncertainty; prediction-oriented validation, such as the zero-shot uniaxial test used here, is the correct criterion for whether an identified law is useful."],"forward_implications":["Material characterization no longer requires specimens that achieve uniform stress and strain; a complex geometry that generates many strain paths can substitute for many separate idealized tests.","Because per-iteration cost is nearly independent of the number of parameters, constitutive models with hundreds or thousands of parameters, including neural networks, become practical targets for identification from experiments.","The formulation applies to history-dependent, rate-dependent inelastic response with a yield surface, a setting where numerical differentiation of the objective would be fragile or prohibitive.","A single inversion can combine data from multiple tests, such as different strain rates or geometries, in one objective; in the quasistatic example a single final displacement snapshot plus a force history sufficed to fix the observed response.","Two materially different parameter sets can fit the same observations, so the meaningful output of the method is a model that reproduces measured and independently tested response, not a unique parameter vector."],"supporting_citations":[{"why":"Supplies the PDE-constrained optimization and adjoint framework used to derive the sensitivity (Eq. 9).","marker":"[24]"},{"why":"Provides the adjoint-state method for computing gradients of the objective with respect to parameters.","marker":"[45]"},{"why":"Gives the variational formulation of viscoplastic constitutive updates that the J2 forward and adjoint plastic updates are built on.","marker":"[41]"},{"why":"Method of Moving Asymptotes is the gradient-based optimization scheme used to update material parameters.","marker":"[52]"},{"why":"Finite element updating is the earlier iterative inverse-identification approach this work extends with an adjoint-based PDE constraint.","marker":"[27]"},{"why":"Digital image correlation supplies the full-field displacement measurement technique motivating the experimental data model.","marker":"[13]"}],"fun_headline_variants":["Infer material laws from full-field tests with adjoint PDE-constrained optimization","PDE-constrained inverse method learns constitutive relations from experiment data","Adjoint-based identification of constitutive laws from full-field measurements","Recovering inelastic material behavior via adjoint-optimized PDE models"],"cache_read_input_tokens":24192,"weakest_assumption_plain":"The load-bearing premise is that the true material's response lies inside the chosen parametric family of Perzyna-type internal-variable models, since the inversion only tunes those parameters and the noise-free synthetic data come from the very same forward solver used in the inversion.","fun_headline_variants_meta":{"raw":{"variants":["Infer material laws from full-field tests with adjoint PDE-constrained optimization","PDE-constrained inverse method learns constitutive relations from experiment data","Adjoint-based identification of constitutive laws from full-field measurements","Recovering inelastic material behavior via adjoint-optimized PDE models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000966,"raw_usage":{"total_tokens":4086,"prompt_tokens":896,"completion_tokens":3190,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":3114}},"tokens_in":512,"tokens_out":3190,"duration_ms":21770,"temperature":1.0,"reasoning_tokens":3114,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:00:16.516612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical inversion on data generated by a materially different constitutive law, such as a different hardening form or a model with damage or softening, or on real experimental data with known measurement noise, and then test the recovered model on an independent non-proportional loading path not used in the objective. A concrete version: add roughly one percent noise to the synthetic displacements and reaction forces in the quasistatic plate-with-hole test and check whether the zero-shot uniaxial error stays below a few percent.","supporting_citations":[{"cited_title":"Hinze, R","cited_arxiv_id":null,"evidence_quote":"Supplies the PDE-constrained optimization and adjoint framework used to derive the sensitivity (Eq. 9)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the adjoint-state method for computing gradients of the objective with respect to parameters."},{"cited_title":"Ortiz and L","cited_arxiv_id":null,"evidence_quote":"Gives the variational formulation of viscoplastic constitutive updates that the J2 forward and adjoint plastic updates are built on."},{"cited_title":"Svanberg","cited_arxiv_id":null,"evidence_quote":"Method of Moving Asymptotes is the gradient-based optimization scheme used to update material parameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Finite element updating is the earlier iterative inverse-identification approach this work extends with an adjoint-based PDE constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Digital image correlation supplies the full-field displacement measurement technique motivating the experimental data model."}],"review_version":1}