{"id":"6c8420c2-e05d-470a-aebc-a188ca7fbddf","arxiv_id":"2509.07474","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A differentiable Kalman filter that learns the state-transition operator by field inversion and a neural closure model is demonstrated on rocket and Allen-Cahn models, with reported 90% error reductions over a fixed-model Kalman filter.","lead":"This paper trains a Kalman filter to fix its own imperfect physics model by learning missing dynamics from noisy data. The authors report at least 90% lower state-estimation error than a classical Kalman filter in two synthetic test problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 90% improvement claim is not supported by the evaluation: the same observation sequence is used to fit F_k (Eq. 24, 31) and to report reconstruction errors (Figs. 1, 4, 5), so the margin may be interpolation rather than predictive skill.","rationale":"The reader's weakest assumption is exactly the load-bearing issue: the optimization objective in Eqs. (24) and (31) is also the quantity used to report errors in Figs. 1, 4, and 5. Since the rocket inversion is overparameterized relative to scalar observations, the reported improvements do not establish predictive skill. I agree with the reader's REJECT verdict because the central quantitative claim is unsupported by the evaluation design, even though the adjoint Jacobian derivation is independently checked via finite differences and automatic differentiation in Appendix C. The proposed held-out check would directly test whether the DKF's advantage survives on unseen data, which is the minimal requirement for the abstract's claim.","tokens_in":18618,"tokens_out":6115,"duration_ms":72789,"concrete_test":"Split each experiment's observation sequence in time: optimize F_k / d(v) via Eqs. (24) and (31) on the first half of the time steps, then freeze the optimized operator and compute the reported RMSE on the held-out second half. For Allen-Cahn, additionally run one new trajectory with a different phase φ or initial condition and report the reconstruction error there. If the ≥90% reduction over classic Kalman filtering persists on held-out time steps or new trajectories, the concern is refuted; if the margin drops substantially, the current numbers are in-sample fits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DKF reduces state reconstruction error by at least 90% versus classic Kalman filtering. The evidence for this rests on Figs. 1, 4, and 5, where the reported errors are computed on the same observation sequence that was optimized in Eqs. (24) and (31). In the rocket case, F_k has four free entries per time step while the observation is scalar, so the inverse problem is underdetermined (4 n_t unknowns against n_t scalar measurements); minimizing Eq. (24) can therefore match the training observations without recovering true dynamics. The apparent closeness of the optimized matrix in Table 2 may only reflect the initialization as small perturbations around a physically reasonable baseline, not identifiability. Likewise, the Allen-Cahn diffusivity table is tuned on the same trajectory used for Figs. 4 and 5, with single realizations and no error bars. Thus the headline margin is a fitting result unless demonstrated on unseen data. The adjoint Jacobian verification in Appendix C supports the gradient computation, but it does not establish that the optimized operator generalizes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Differentiable Kalman Filter (DKF), a two-level optimization framework in which the state transition matrix F_k of a Kalman filter is treated as a design variable and optimized via analytically derived adjoint gradients against observation mismatch. After inversion, a neural-network closure model is trained to reproduce the optimized operators. The method is demonstrated on a rocket dynamics model and an Allen–Cahn reaction–diffusion problem, and the authors claim that DKF reduces state reconstruction error by at least 90% relative to the classical Kalman filter while maintaining robust uncertainty quantification. The adjoint Jacobian derivation is presented in Appendix B and verified against automatic differentiation and finite differences in Appendix C.","tokens_in":18871,"tokens_out":6937,"duration_ms":85557,"significance":"If the central numerical claim were supported, the paper would make a useful contribution: an analytic adjoint formulation for differentiating through the Kalman filter is valuable, and the two-level field-inversion-plus-closure workflow is an attractive way to combine data assimilation with machine learning. The Jacobian verification in Appendix C is a concrete strength and gives confidence that the gradient machinery is correct. However, the reported 90% improvement is the paper's headline result, and the current evaluation does not substantiate it as predictive performance because the same observation sequence is used for fitting and for reporting errors. The contribution is therefore promising but not yet demonstrated.","major_comments":[{"comment":"The evaluation is circular for the paper's central claim. In the rocket case, Eq. (24) minimizes the squared observation error over F_k on the same sequence z_k used to report reconstruction errors in Fig. 1. In the Allen–Cahn case, Eq. (31) minimizes the whitened innovation on the trajectory whose reconstructions are shown in Figs. 4 and 5. With per-time-step free parameters, the optimizer can fit the training observations without learning a generalizable model. The reported at-least-90% error reduction is therefore a training-fit result unless the authors evaluate on held-out observation sequences or on independent trajectories. A train/test split should be added, or the claims should be explicitly reframed as fitting/reconstruction accuracy.","section":"Section 4, Eqs. (24), (31), Figs. 1, 4, 5"},{"comment":"The rocket inversion is underdetermined: F_k has 4 free entries per time step, while each observation is a scalar (H_k = [1 0], Eq. (23)), giving 4 n_t unknowns against n_t measurements. Minimizing Eq. (24) can therefore match the observed altitude without recovering the true dynamics. Table 2 shows the optimized F_k close to the true F, but this is not a consequence of identifiability; it likely reflects the initialization as 'small perturbations around a physically reasonable baseline' (Sec. 4.1.1). The authors should provide an identifiability or regularization analysis, report reconstruction error for the unobserved velocity component, and validate on unseen data before claiming that the dynamics have been discovered.","section":"Section 4.1, Eq. (24), Table 2, Fig. 1"},{"comment":"The uncertainty-quantification claim is not quantitatively supported. The DKF posterior variance is produced by a filter whose model was optimized on the same trajectory; matching the supposedly true [P_k]_{ii} on that trajectory is a consistency check, not a calibration test. No coverage probabilities, reliability metrics, or ensemble-spread diagnostics are reported, and each noise level appears to use a single realization without error bars. The phrase 'robust uncertainty quantification' in the abstract requires statistical calibration evidence on independent data.","section":"Section 4.2.5, Fig. 5"},{"comment":"The Allen–Cahn experiments use a single trajectory for both inversion and evaluation. The tabulated diffusivity values are optimized against the same observations used to produce the reconstructions in Fig. 4 and the variance comparison in Fig. 5; the DNN is then trained on those same inversion results. Consequently, the DNN reconstructions demonstrate interpolation on the training trajectory, not generalization. Please add a train/test split across initial conditions or noise realizations, and report the corresponding test errors.","section":"Section 4.2.2-4.2.4"}],"minor_comments":[{"comment":"The table marks DEKF with '×' for Learning/Adaptation and End-to-end Gradient, but the text states that the differentiable EKF 'allows the entire state estimation pipeline to be trained via gradient-based optimization.' This inconsistency should be resolved.","section":"Section 1, Table 1"},{"comment":"The notation n_t is used both for the number of time steps and as the upper index for the tabulated diffusivity values d~(v_{n_t}). Since these are different quantities, please use separate symbols (e.g., n_v for the number of table points).","section":"Eq. (31), Section 4.2.2"},{"comment":"The captions do not specify exactly how the 'DNN predictions' are obtained: is the neural-network surrogate used inside the Kalman filter as the transition operator, or is it used to reconstruct the state directly? Please clarify in the captions and text.","section":"Figs. 4 and 5"},{"comment":"The experimental setup is underspecified: the measurement noise covariance R, process noise covariance Q, thrust force, mass, burn time, and the L-BFGS convergence criteria are not reported. These details are needed to reproduce the results and to interpret the noise-level comparison.","section":"Section 4.1"},{"comment":"The conclusion states that the Allen–Cahn case achieves an RMSE of d(v) below 1e-2 and an improvement of 'at least two orders of magnitude' compared to the classical Kalman filter, while the abstract reports a '90%' reduction in state reconstruction error. The metrics should be aligned and clearly defined.","section":"Section 5, abstract"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands: the central numerical claim is undermined by the train/test circularity identified at Eqs. (24) and (31). I agree with the reader that the current evidence is not sufficient for acceptance. However, the adjoint derivation is internally consistent and verified, so the issue is fixable with a proper holdout evaluation. If the authors cannot provide such an evaluation, or if the 90% margin disappears on held-out data, the paper should be rejected. I recommend major revision with the holdout experiments as a required condition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is analytic: the block Jacobians in Appendix B are derived cleanly and cross-checked against automatic differentiation and finite differences in Appendix C, which held up to inspection. The two-stage workflow—field-invert the transition operator inside a Kalman filter, then fit a neural closure to the inverted operators—is a legitimate extension of FIML into a sequential filtering setting, even if the individual pieces are not new. The Allen–Cahn operator construction is also sensibly physics-informed rather than a black box.\n\nThat said, the central numerical claim does not survive contact with the evaluation design. Eq. (24) and Eq. (31) minimize the observation error on the same trajectory that is then used to report the reconstruction errors in Figs. 1, 4, and 5. So the \"at least 90% reduction\" is a measure of how well the optimized filter fits the data it was tuned on, not how well it predicts unseen observations. In the rocket case the inverse problem is overparameterized—four entries of F_k per scalar observation—so matching the observations does not imply discovering the true dynamics. The fact that the optimized F_k in Table 2 stays near a reasonable baseline could simply reflect the initialization. The Allen–Cahn results are single realizations, no error bars, and the paper never reports a numeric error table for any of the claimed margins, which makes it hard to check the 90% claim quantitatively even on the training data.\n\nThere is also a smaller internal inconsistency: the abstract says \"at least 90%\" error reduction, while the conclusion claims \"at least two orders of magnitude\". Those are not the same thing, and neither is backed by a number in the text.\n\nOn the plus side, the paper is honest about the method's limits in a few places, and the derivations are verifiable. The failure is in the evaluation, not in the mathematics. A revised version with held-out trajectories, proper baselines from the differentiable-filter literature (e.g., Kloss et al., KalmanNet), numeric error tables, and error bars could well change the verdict.\n\nWho is this for? Researchers working on closure modeling or differentiable data assimilation might find the adjoint derivations useful, but they should not cite the 90% claim as evidence. I would send it to a referee because the analytic work is substantive and worth engaging with, but I would expect major revision or rejection on the current evidence.","headline":"Solid adjoint-derivation work, but the headline 90% improvement is a training-set fit: same observations are optimized and scored, with no held-out evaluation, no baselines, and no numeric error tables.","tokens_in":19371,"tokens_out":1733,"would_cite":false,"duration_ms":23071,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E11","65K10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Kalman filter whose transition operator is optimized by adjoint gradients, then approximated by a neural network, recovers missing dynamics and cuts reconstruction error by at least 90% on two test systems.","keywords":["differentiable Kalman filter","field inversion","adjoint sensitivity","model-form error","closure model","Allen-Cahn equation","state estimation","uncertainty quantification"],"falsifier":"Use one trajectory: run the DKF field inversion on the first half of the time series and evaluate reconstruction on the held-out second half, at sigma=0.005, 0.025, and 0.125. If the at-least-90% error reduction observed in the paper does not persist on the held-out segment, the optimized operator is fitting the evaluation sequence rather than discovering the dynamics. The same test applies to the Allen-Cahn case with independent noise draws; additionally, compare per-step parameter count (four F_k entries per scalar rocket observation) with attained recovery of the true F_k.","tokens_in":18482,"feed_emoji":"📡","tokens_out":6059,"duration_ms":69355,"temperature":0.7,"pith_summary":"The paper claims that the classical Kalman filter's fixed dynamics model can be made adaptive by treating the state-transition operator as an optimization variable. A field-inversion stage minimizes the mismatch between predicted and observed measurements using analytic adjoint gradients, and a second stage trains a neural network to reproduce the optimized operators as a state-dependent closure model. On a rocket trajectory problem and an Allen-Cahn reaction-diffusion boundary value problem, the two-stage DKF lowers reconstruction error by at least 90% across tested noise levels and keeps the filter's covariance-based uncertainty estimates close to the true errors. The contribution matters because it sketches a way to combine data assimilation with machine learning while retaining the Kalman filter's structure and uncertainty propagation.","feed_headline":"Kalman filter learns missing physics, cutting error by 90%","feed_subtitle":"Optimizing the transition operator inside the filter, then fitting a neural net, recovers dynamics from noisy data.","key_machinery":"The differentiable Kalman filter (DKF): the standard Kalman predict-update loop is reinterpreted as a root-finding system whose residuals r_k stack state and covariance equations, with the transition matrices F_k as learnable parameters. The argument is carried by analytic adjoint Jacobian blocks (Eq. 17)—expressions for A_{k,k-1} and A_{k,k} with respect to state x and covariance P—which provide gradients of the observation-mismatch loss without unrolling the recursion. A neural closure DNN(x,d;theta) trained by Frobenius loss on the optimized operators then serves as the state-dependent transition law.","core_discovery":"The central claim is that model-form error—the gap between the approximate transition operator carried by the filter and the true dynamics—can be removed by solving a least-squares field inversion problem at filter level. The filter's predict-update recursion is treated as a differentiable residual system, the transition matrices F_k are the design variables, and adjoint equations derived from the residual Jacobian blocks give the gradient of the observation mismatch. After the optimized F_k sequence is found, a DNN trained to map state and design variables to those operators serves as the new transition law. In the paper's two examples the optimized filter produces nearly true trajectories","pith_inferences":["The paper does not test out-of-sample generalization; we infer that splitting the observation window into fit and hold-out segments would reveal how much of the reported margin is memorization of the fitted trajectory.","The analytic Jacobian blocks of Eq. (17) are generic to Kalman recursion and, we infer, could accelerate training of other learned filtering components such as measurement models or noise covariances.","The '90% better than classical KF' comparison uses a deliberately un-optimized baseline; we infer the margin would be smaller, and more informative, against a well-tuned extended or unscented filter.","In the rocket case the four free entries of F_k per time step exceed the single scalar observation per step; we infer that identifiability is the limiting factor and that richer observations or regularization would be needed for reliable dynamics discovery."],"forward_implications":["If DKF works as claimed, any Kalman-filter application with an imperfect model can replace its fixed transition operator with a data-corrected one, improving tracking without giving up recursive filtering.","The learned closure model, trained on optimized operators, carries the correction into new states, so the benefit is not confined to the exact trajectory used for inversion.","The analytic adjoint gradient path avoids storing a full unrolled computation graph, which would make the approach more scalable to long horizons.","Since the covariance recursion is retained, the corrected filter still outputs principled uncertainty around its state estimate rather than a point forecast.","The two-stage workflow gives an interpretable sequence of discovered transition operators that can be inspected as data-driven corrections of the physics."],"supporting_citations":[{"why":"Supplies the classical Kalman predict-update recursion that DKF differentiates and optimizes.","marker":"[41]"},{"why":"Introduces the field-inversion-and-machine-learning two-stage workflow that DKF adapts to sequential filtering.","marker":"[40]"},{"why":"Provides the adjoint/control-theoretic sensitivity method used to derive the analytic gradients of the filter residuals.","marker":"[42]"},{"why":"Establishes the differentiable-filter line of work the paper positions against; the approach inherits end-to-end gradient training.","marker":"[29]"},{"why":"Differentiable extended Kalman filter application used as a baseline and comparison in the method table.","marker":"[32]"},{"why":"Differentiable UKF/KalmanNet baseline for end-to-end adapted filtering, used as a comparison point.","marker":"[33]"},{"why":"Differentiable ensemble Kalman filter baseline for high-dimensional estimation, used as a comparison point.","marker":"[34]"}],"fun_headline_variants":["Differentiable Kalman filter corrects model error, cuts error 90%","Kalman filter learns missing physics via adjoint optimization","Field inversion plus neural net closure cuts Kalman error 90%","Adjoint-based DKF recovers true dynamics, 90% less error","Neural closure model fixes Kalman filter, error down 90%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The evaluation assumes that optimizing the transition operator on the very observation sequence used for the reported errors is a valid proxy for reconstruction on unseen data; with more fitted parameters per time step than independent observations, the filter can fit noise without learning real dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable Kalman filter corrects model error, cuts error 90%","Kalman filter learns missing physics via adjoint optimization","Field inversion plus neural net closure cuts Kalman error 90%","Adjoint-based DKF recovers true dynamics, 90% less error","Neural closure model fixes Kalman filter, error down 90%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000679,"raw_usage":{"total_tokens":2920,"prompt_tokens":738,"completion_tokens":2182,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":2089}},"tokens_in":482,"tokens_out":2182,"duration_ms":16761,"temperature":1.0,"reasoning_tokens":2089,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:09:22.315522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use one trajectory: run the DKF field inversion on the first half of the time series and evaluate reconstruction on the held-out second half, at sigma=0.005, 0.025, and 0.125. If the at-least-90% error reduction observed in the paper does not persist on the held-out segment, the optimized operator is fitting the evaluation sequence rather than discovering the dynamics. The same test applies to the Allen-Cahn case with independent noise draws; additionally, compare per-step parameter count (four F_k entries per scalar rocket observation) with attained recovery of the true F_k.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical Kalman predict-update recursion that DKF differentiates and optimizes."},{"cited_title":"Kloss, G","cited_arxiv_id":null,"evidence_quote":"Establishes the differentiable-filter line of work the paper positions against; the approach inherits end-to-end gradient training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Differentiable extended Kalman filter application used as a baseline and comparison in the method table."},{"cited_title":"Revach, N","cited_arxiv_id":null,"evidence_quote":"Differentiable UKF/KalmanNet baseline for end-to-end adapted filtering, used as a comparison point."}],"review_version":1}