{"id":"d8df9bf8-329c-419f-90cc-dac8a00c6dad","arxiv_id":"2606.29349","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"PDE-constrained optimization with reverse-mode AD and L-BFGS jointly recovers full-field displacement and force from sparse sensors, outperforming modal expansion by 2-17x on cantilever plate, elbow pipe, and reactor pressure vessel examples.","lead":"The paper formulates virtual sensing in structural dynamics as a PDE-constrained optimization problem that recovers full-field displacements and forces from sparse sensors by treating the elastodynamic equation as a hard constraint. Engineers in structural health monitoring and nuclear applications might read it for a reported accuracy gain over modal methods on realistic geometries.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"All three numerical examples generate synthetic measurements from the identical elastodynamic PDE used as the optimization constraint, leaving robustness to modeling error untested.","rationale":"The reader's weakest_assumption directly identifies the same modeling-error and uniqueness assumptions that underpin the numerical results. Because the full-text examples remain fully consistent with an exact-model regime, the concern stands and the UNVERDICTED status is appropriate pending robustness checks.","tokens_in":1739,"tokens_out":377,"duration_ms":20257,"concrete_test":"Re-run the RPV example after generating the 'measured' displacements with a forward model that includes 5 % uniform perturbation in Young's modulus plus a small viscous damping term absent from the inversion PDE; keep the same sensor locations and noise level, then recompute both PDE-CO and modal-expansion errors. If the displacement error reduction factor falls below 2×, the modeling-error assumption materially affects the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim is that PDE-CO recovers displacement and force with 2–17× lower error than modal expansion and sub-percent accuracy. This performance is demonstrated exclusively on in-silico data produced by solving the same linear elastodynamic PDE (with the same material properties, geometry, and boundary conditions) that is subsequently imposed as an equality constraint. Under these conditions the inverse problem is well-posed by construction once Tikhonov regularization is applied, but the setup supplies no information on how either method behaves when the true structure deviates from the model (unmodeled damping, geometric imperfections, spatially varying moduli, or sensor noise statistics not captured by the forward operator). Because the relative advantage of the PDE constraint over modal truncation could diminish or reverse under such mismatch, the reported error-reduction factors cannot be taken as evidence that the method will deliver comparable gains on physical structures.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper formulates virtual sensing as a PDE-constrained optimization problem in which the linear elastodynamic PDE is imposed as an equality constraint, the force distribution is the decision variable, and full-field displacement and force are recovered jointly via Tikhonov-regularized L-BFGS minimization with gradients obtained by reverse-mode automatic differentiation. The method is implemented by coupling an offline FEniCSx finite-element stage with an online JAX/GPU stage and is demonstrated on three synthetic examples of increasing geometric complexity (cantilever plate, 90° elbow pipe, reactor pressure vessel). The central claim is that PDE-CO reduces displacement error by factors of 2–17× relative to modal expansion while achieving sub-percent accuracy on every example.","tokens_in":1954,"tokens_out":584,"duration_ms":20503,"significance":"If the performance advantage survives modeling mismatch, the joint PDE-constrained recovery of displacement and force would be a useful addition to structural-health-monitoring toolkits. The hybrid FEniCSx–JAX implementation and explicit GPU timing results are concrete strengths that support reproducibility of the computational workflow.","major_comments":[{"comment":"All three numerical examples generate the synthetic sensor data by solving exactly the same linear elastodynamic PDE (identical geometry, material properties, boundary conditions, and forcing) that is subsequently enforced as the equality constraint in the optimization. Because the inverse problem is therefore well-posed by construction once Tikhonov regularization is applied, the reported 2–17× error reductions cannot be interpreted as evidence that the same advantage will persist when the true structure deviates from the model (unmodeled damping, spatially varying moduli, geometric imperfections, or sensor-noise statistics not captured by the forward operator). This verification gap directly affects the headline claim of consistent outperformance.","section":"Numerical examples (the three cases described in the abstract and §4)"},{"comment":"The Tikhonov regularization parameter is listed as a free hyper-parameter whose value is not derived from first principles. The sub-percent accuracy figures and the factor-of-2–17 improvement are therefore conditional on a specific choice of this parameter; no sensitivity study or a-priori selection rule is supplied that would allow the reader to assess how much of the reported gain is attributable to favorable tuning versus the PDE constraint itself.","section":"Optimization formulation and numerical results"}],"minor_comments":[{"comment":"The abstract states that force is recovered simultaneously through the PDE constraint, yet the precise definition of the force variable (distributed body force versus boundary traction) and its discretization are not restated in the results section; a short clarifying sentence would remove ambiguity when comparing to modal-expansion force back-calculation.","section":"Abstract and §3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments. We address each major point below, acknowledging where the examples are verification studies and where additional analysis can strengthen the presentation.","responses":[{"response":"We agree that the three examples constitute verification studies in which sensor data are generated from the identical PDE later used as the equality constraint. This controlled setting isolates the contribution of the PDE constraint and joint recovery of displacement and force, yielding the reported 2–17× error reductions relative to modal expansion under exact model match. The manuscript’s claims are confined to these synthetic cases; we do not assert robustness to model mismatch. A clarifying sentence will be added to §4 and the discussion to state explicitly that the results are verification results on consistent forward/inverse models. The PDE-CO formulation itself remains applicable to approximate models, but quantitative performance under mismatch is outside the scope of the present work.","revision_made":"partial","referee_comment":"[Numerical examples (the three cases described in the abstract and §4)] All three numerical examples generate the synthetic sensor data by solving exactly the same linear elastodynamic PDE (identical geometry, material properties, boundary conditions, and forcing) that is subsequently enforced as the equality constraint in the optimization. Because the inverse problem is therefore well-posed by construction once Tikhonov regularization is applied, the reported 2–17× error reductions cannot be interpreted as evidence that the same advantage will persist when the true structure deviates from the model (unmodeled damping, spatially varying moduli, geometric imperfections, or sensor-noise statistics not captured by the forward operator). This verification gap directly affects the headline claim of consistent outperformance."},{"response":"The regularization parameter λ is chosen empirically for each example to achieve the reported accuracy. We will add a short sensitivity study (new figure or table in §4) showing displacement and force error versus λ for the cantilever-plate case, together with a brief statement that, in the absence of ground truth, λ may be selected by the discrepancy principle or cross-validation on held-out sensors. This addition will make the dependence on λ explicit without altering the core claims.","revision_made":"yes","referee_comment":"[Optimization formulation and numerical results] The Tikhonov regularization parameter is listed as a free hyper-parameter whose value is not derived from first principles. The sub-percent accuracy figures and the factor-of-2–17 improvement are therefore conditional on a specific choice of this parameter; no sensitivity study or a-priori selection rule is supplied that would allow the reader to assess how much of the reported gain is attributable to favorable tuning versus the PDE constraint itself."}],"tokens_in":1538,"tokens_out":559,"duration_ms":20583,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this paper delivers a concrete PDE-constrained optimization method for virtual sensing that jointly recovers full-field displacement and force by treating force as the optimization variable under the elastodynamic equality constraint. On the three synthetic cases it reports clear error reductions over modal expansion.\n\nThe formulation is straightforward: Tikhonov-regularized objective, reverse-mode AD through the forward solver, L-BFGS updates, with an offline FEniCSx assembly stage feeding a JAX GPU stage. The examples progress from a cantilever plate to an elbow pipe to a reactor pressure vessel model, and the numbers show sub-percent displacement accuracy plus the claimed speedups. The joint recovery avoids the separate back-calculation step that modal methods require.\n\nThe implementation details are reproducible in principle and the GPU acceleration is a practical plus. The comparison to modal expansion is direct and the performance gap is quantified on each geometry.\n\nThe soft spot is that every measurement set is produced by solving the identical linear elastodynamic PDE with the same material properties and boundary conditions later used as the hard constraint. This leaves open how either method behaves once the real structure deviates through damping, geometric imperfections, or spatially varying moduli. The reported factors of improvement therefore apply strictly inside the matched-model regime.\n\nThe work is aimed at structural-health-monitoring engineers who need full-field estimates from sparse sensors on engineering geometries. Readers already working on PDE-constrained inverse problems will find the software stack and the RPV-scale example useful.\n\nThe paper is coherent on its own terms and supplies specific numerical claims, so it deserves a serious referee.","headline":"PDE-CO recovers displacement and force jointly via the elastodynamic PDE on synthetic data and beats modal expansion by 2-17x, but all tests use data generated from the exact same model.","tokens_in":2419,"tokens_out":407,"would_cite":false,"duration_ms":21321,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"PDE-constrained optimization recovers full-field displacements and forces from sparse sensors by enforcing the elastodynamic equation as a constraint.","keywords":["PDE-constrained optimization","virtual sensing","structural dynamics","elastodynamics","inverse problem","full-field reconstruction","sparse sensors","finite element method"],"falsifier":"Application of the method to a physical structure whose material properties, boundary conditions, or damping deviate from those assumed in the PDE model, followed by comparison of recovered fields against independent full-field measurements.","tokens_in":2645,"feed_emoji":"📐","tokens_out":702,"duration_ms":25827,"temperature":0.7,"pith_summary":"This paper formulates virtual sensing as a PDE-constrained optimization problem in which the governing elastodynamic equation acts as an equality constraint and the applied force distribution serves as the optimization variable. The framework jointly recovers consistent full-field displacements and forces that satisfy both the physics and the sparse sensor data. Gradients of the Tikhonov-regularized objective are obtained via reverse-mode automatic differentiation through the forward PDE solver, with L-BFGS used to find the optimum. The method is tested on a cantilever plate, a 90-degree elbow pipe, and a reactor pressure vessel model, where it reduces displacement errors by factors of 2 to 17 relative to modal expansion while reaching sub-percent accuracy.","feed_headline":"PDE optimization recovers full fields from sparse sensors","feed_subtitle":"Joint recovery of displacement and force by enforcing the elastodynamic equation cuts errors by 2-17x versus modal expansion on plates, pipe","key_machinery":"PDE-constrained optimization in which the elastodynamic equation serves as an equality constraint, the force distribution is the optimization variable, and gradients are computed by reverse-mode automatic differentiation through the PDE solver.","core_discovery":"By casting virtual sensing as an optimization problem with the elastodynamic PDE as an equality constraint and the force as the variable to be determined, the framework recovers full-field displacements and forces that satisfy both the physics and the sparse data, achieving sub-percent accuracy and outperforming modal methods by factors of 2 to 17 in displacement error.","pith_inferences":["The same constrained-optimization structure could be applied to other linear PDE-governed systems such as heat conduction or acoustics if the forward solver is available.","When independent full-field data become available, the recovered force fields could serve as an additional consistency check beyond displacement error alone.","Real-time deployment would require further reduction of the online solve time or warm-starting strategies for successive time windows."],"forward_implications":["Displacement recovery errors are reduced by factors of 2 to 17 compared to modal expansion across all tested geometries.","Force distributions are recovered simultaneously through the PDE constraint rather than derived from truncated modal coordinates.","Sub-percent accuracy is obtained on complex geometries including a representative reactor pressure vessel.","GPU acceleration through the JAX stage offsets the computational cost of the optimization relative to CPU implementations."],"fun_headline_variants":["PDE-constrained optimization recovers displacement and force from sensors","Virtual sensing as PDE optimization outperforms modal expansion","Elastodynamic PDE enables full-field recovery with sparse data","PDE-CO jointly recovers displacement and force via optimization"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The elastodynamic PDE is assumed to be an exact and complete model of the structure with no modeling error, and the Tikhonov-regularized inverse problem is assumed to admit a unique physically meaningful solution from the given sparse measurements.","fun_headline_variants_meta":{"raw":{"variants":["PDE-constrained optimization recovers displacement and force from sensors","Virtual sensing as PDE optimization outperforms modal expansion","Elastodynamic PDE enables full-field recovery with sparse data","PDE-CO jointly recovers displacement and force via optimization"]},"model":"grok-4.3","cost_usd":0.005445,"raw_usage":{"total_tokens":2631,"prompt_tokens":691,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":54449500,"prompt_tokens_details":{"text_tokens":691,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1879,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":691,"tokens_out":61,"duration_ms":18495,"temperature":1.0,"reasoning_tokens":1879,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T01:57:40.609451+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Application of the method to a physical structure whose material properties, boundary conditions, or damping deviate from those assumed in the PDE model, followed by comparison of recovered fields against independent full-field measurements.","supporting_citations":[],"review_version":1}