{"id":"2c51bada-8fac-4d82-a7b4-567aaf6daba9","arxiv_id":"2501.18808","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An autoregressive Hamiltonian neural network coupled with an unscented Kalman filter improves long-term trajectory prediction and uncertainty quantification for unknown Hamiltonian systems.","lead":"The paper trains Hamiltonian neural networks from position-velocity state pairs, adds an autoregressive loss over multiple rollout steps, and feeds the learned model into an unscented Kalman filter that folds in live measurements. It then shows this combination outperforms plain MLP, neural ODE, and HNN baselines on a mass-spring system and on highly elliptical Molniya-type orbits with J2 perturbations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that AHNN learns a true Hamiltonian surrogate is not established because the fixed 60 s RK4 step may alias periapsis dynamics; the learned Hθ could be a step-size-dependent effective Hamiltonian rather than the physical energy.","rationale":"I agree with the reader that the fixed 60 s RK4 step is the weakest load-bearing assumption. The method's central contribution rests on learning a Hamiltonian surrogate that preserves physical invariants; if the learned Hθ is only an effective Hamiltonian matched to h = 60 s, the physical interpretation and the generalization claim collapse, and the UKF advantage could be an artifact of the discretization rather than evidence of a good dynamics model. This concern is independent of the reproducibility issues (no code, no error bars, unreported UKF hyperparameters), which also support a conditional verdict but are not the single load-bearing assumption. The proposed test directly separates a true Hamiltonian model from a step-tuned discrete map by evaluating the same trained model with smaller RK4 substeps. If the test passes, the central claim is on much firmer ground; if it fails, the paper should be reframed as learning a discrete surrogate for a fixed measurement cadence, and the claim of preserving physical invariants would need substantial qualification. The Table 2 anomaly (AHNN5 velocity RMSE 4.596 versus AHNN3 0.648 in the true-initial condition) is an additional reason to require seed-level statistics, but it does not change the primary recommendation: conditional acceptance pending a step-sensitivity study and the other reproducibility checks.","tokens_in":14274,"tokens_out":9513,"duration_ms":108142,"concrete_test":"Train AHNN5 on the same orbit data, then evaluate long-term prediction using the identical Hθ but integrating Eq. (3) with RK4 at 15 s substeps (four substeps per 60 s interval) instead of a single 60 s step. If position/velocity RMSE or the reported energy RMSE changes materially (e.g., more than a factor of 2) or if the energy trace shows periapsis spikes, the learned Hθ is step-size-dependent and the claim of learning continuous Hamiltonian dynamics is not supported. A complementary control is to retrain with a 30 s step and report whether the same W=5 advantage and UKF gains persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper trains Fθ in Eq. (3) using a fourth-order Runge-Kutta integrator with a fixed 60 s step (Section 5.2) and evaluates on Molniya orbits whose speed varies by up to an order of magnitude (Figure 1). Near periapsis, a 60 s step spans a large fraction of the fast segment, so RK4's local error is not controlled. Because Fθ is trained to reproduce 60 s transitions, Hθ can absorb the integrator's discretization error into an effective Hamiltonian for step h = 60 s. Such an Hθ would conserve a modified energy, not the physical Hamiltonian, and would not generalize if the integrator step is changed even though the underlying continuous dynamics are unchanged. The reported energy RMSE and the claim of preserving physical invariants are therefore not sufficient evidence that the continuous Hamiltonian dynamics were learned. The UKF corrections in AHNNKF5 could also mask this step-size bias, so the best reported RMSE does not distinguish a good physical model from a step-tuned discrete surrogate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Hamiltonian Neural Network (HNN) training objective that does not require state-derivative labels, an autoregressive extension (AHNN) that trains over multi-step rollouts, and a UKF-based Bayesian data assimilation scheme (AHNNKF) for state estimation and uncertainty quantification. The authors validate the approach on a frictionless mass-spring system and on highly elliptic, J2-perturbed Molniya orbits, reporting RMSE and energy-conservation metrics. The central claim is that AHNN5 and AHNNKF5 achieve the best long-term prediction accuracy while preserving physical invariants such as energy, and that the UKF integration improves robustness and provides uncertainty estimates. The paper does not ship code, and the empirical validation has several gaps that currently prevent full confidence in the central claim.","tokens_in":14516,"tokens_out":6847,"duration_ms":72364,"significance":"If the claims are substantiated, the paper makes a useful algorithmic contribution: it replaces derivative labels with one-step transitions, adds an autoregressive training objective that is particularly relevant to long-horizon prediction, and embeds the learned surrogate in a Bayesian filter to obtain state estimates with uncertainty. The orbital benchmark on highly elliptic Molniya orbits is a meaningful and nontrivial test case, and the energy-RMSE comparisons against MLP and NODE baselines are informative. However, the load-bearing empirical claims are not yet fully supported: the fixed 60-second RK4 step may alias the fast periapsis dynamics, no error bars are reported around the headline RMSE values, and the gravitational potential used for data generation is underspecified. The paper is potentially valuable, but it needs substantial additional validation and clarification before the main claims can be accepted.","major_comments":[{"comment":"The fixed 60 s RK4 step used to define and train Fθ may alias the periapsis segment of the Molniya orbits, so the claim that the learned Hθ preserves the physical Hamiltonian is not established. Section 5.1.2 reports periapsis altitudes of 540–560 km and eccentricities 0.7–0.8, and Figure 1 notes speed variations of an order of magnitude; with h=60 s the RK4 local error near periapsis is uncontrolled, and because the training samples are exactly 60 s apart, Hθ can absorb the discretization error into an effective step-dependent Hamiltonian. The reported energy RMSE (Figure 7) therefore does not distinguish a physical Hamiltonian from a step-tuned discrete surrogate. I request an explicit discretization-convergence study: retrain and evaluate with h=30 s and h=120 s, or with an adaptive integrator at a fixed tolerance, and show that the induced energy error and test RMSE are stable across these choices.","section":"§5.2 and Eq. (3)"},{"comment":"No error bars or variance information are reported for the central comparisons. Section 5.2 states that models are evaluated with three random seeds, but Tables 1 and 2 report single RMSE values; it is not stated whether these are the best, median, or mean of the seeds, and no standard deviation or confidence interval is given. This matters because several advertised improvements are small: for example, in Table 1, AHNNKF5 position RMSE (0.6596) is only about 10% below HNNKF (0.7261) in the true-initial mass-spring case. Please report the mean and standard deviation over seeds and over test orbits for all entries in Tables 1 and 2, and similarly for the energy RMSE values reported in the text.","section":"§5.3, Tables 1–2"},{"comment":"The gravitational potential used to generate the orbital data is underspecified. Equation (13) presents a full spherical-harmonic expansion with unspecified degrees Nz and Nt, defines longitude via arccos, and does not specify the rotation between the Earth-centered inertial frame in which (q,p) are expressed and the body-fixed frame in which the potential is evaluated; the text then states that only J2 is used, but does not say whether the training and evaluation data were generated with only J2 or with additional GGM05C terms. Please state precisely which potential was used for data generation and training, give the J2 value and frame conventions, and replace the arccos longitude definition with atan2(y,x) if any longitude-dependent tesseral terms are retained.","section":"§4.2, Eq. (13)"},{"comment":"In Table 2, the NODE row reports perturbed-initial RMSE values of 14374 and 25121, which are two orders of magnitude larger than the true-initial values (124.66 and 84.067) and appear inconsistent with the scaling column header; the text discusses this only qualitatively as a lack of robustness. Please clarify how these values were computed, whether any divergent filtered trajectories were excluded or clipped, and whether the scaling of the table entries is the same for all models.","section":"Table 2, NODE row"}],"minor_comments":[{"comment":"Tables 1 and 2 contain repeated digit strings such as '1.55281.55281.5528' and '1.73811.73811.7381'; if these are not formatting artifacts in the manuscript, the table entries need to be corrected.","section":"Tables 1–2"},{"comment":"The sentence 'We use a fourth-order Runge-Kutta integrator with a fixed time step of 60 s' appears to apply to the orbital experiments, but the mass-spring experiment uses a 0.01 s sampling interval; clarify that the 60 s step applies only to the orbital experiments and specify the integration step used for the mass-spring system.","section":"§5.2"},{"comment":"The reference for Brown (2004) is incomplete because it lacks a title, and the AdamW citation lists only Loshchilov; please provide the full references.","section":"References"},{"comment":"The 'ExpHyperbolicLR' scheduler is cited to a preprint by the authors but is not described in the paper; either define the scheduler or remove the dependency on an unpublished external method.","section":"§5.2"},{"comment":"In Eq. (7), Yk is described as the set of measurements 'until the time instant tk+1'; it should be tk, and the quantity δxk+1 should be defined before it is used in the covariance expression.","section":"Eq. (7)"},{"comment":"Algorithm 1 has no return statement; specify the outputs (x̂k+1, Pk+1) to make the pseudocode self-contained.","section":"Algorithm 1"},{"comment":"The sigma-point generation is described as taking 'the i-th row of the matrix square root'; specify whether a Cholesky factor or a symmetric square root is used, since this affects the spread of the sigma points.","section":"§3.2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a machine-learning or physics-informed modeling venue, and I do not see a fundamental circularity problem in the experimental design: the HNN is trained on external trajectory data, and the baselines are standard. The main obstacles are the fixed-step discretization issue around periapsis and the absence of uncertainty quantification on the reported metrics; both are addressable with additional experiments and reporting. I would also encourage the authors to release code and data, as the current level of detail is not sufficient for a fully independent reproduction of the orbital experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kolya,\n\nShort version: this is a plausible, useful integration of three known pieces—Hamiltonian neural networks, unrolled/multi-step training, and the unscented Kalman filter—applied to Molniya-class orbit prediction with J2. The empirical story is coherent: AHNN with window 5 plus UKF beats MLP, NODE, and plain HNN on the reported RMSE and preserves energy better. That result, if it holds up, is genuinely useful for orbit determination.\n\nWhat's new is the specific combination and the application. The autoregressive rollout loss (Eq. 4) is a natural extension of Greydanus, and pairing it with UKF gives you a practical way to do online state refinement with an NN process model. The mass-spring and Molniya experiments are appropriate benchmarks. The paper reads well, and the authors clearly know the astrodynamics.\n\nNow the soft spots, in proportion. First, the evaluation has a few concrete deficits: no error bars or seed-wise spread in any RMSE table; UKF hyperparameters (α, β, κ) are TPE-tuned on the validation set but the chosen values are never reported; and there is a clear internal inconsistency in Table 2—AHNN5 velocity RMSE (4.5962) is larger than AHNN3 (0.6476) for the true-initial case, which conflicts with the SMA figures showing AHNN5 best. The \"NODE 14374\" entry in the same table is almost certainly a missing decimal point, which makes one worry about proofreading. Appendix A.2 has a misplaced/contradictory sentence claiming HNN does not improve with UKF, while Table 2 shows mixed evidence (position improves dramatically, velocity gets worse). None of these by itself kills the paper, but together they mean the claims outrun the evidence as presented.\n\nSecond, and more substantive: everything is trained and evaluated at a fixed 60-second RK4 step. For a Molniya orbit with order-of-magnitude speed variation, that step very likely does not resolve periapsis, so the learned Hθ can absorb discretization error into an effective Hamiltonian. The paper's claim to have learned the physical H and to conserve physical energy is therefore not established. The authors should show step-size insensitivity, run an adaptive integrator, or at least discuss this. This is fixable, but it matters for the \"preserving physical invariants\" claim.\n\nThere is also a minor overstatement that HNN is \"unexplored\" in derivative-free training; that's not quite true in the literature. And no code/data is released, which makes verification harder.\n\nOverall: for a practitioner tracking objects in highly elliptical orbits, this is worth reading. It deserves a serious referee, but I'd send it back for major revision, primarily for the step-size analysis, error bars, and cleaning up the tables/appendix. I wouldn't cite it yet until those are addressed.","headline":"Useful combination of HNN+UKF for Molniya orbit prediction, but the evaluation needs error bars, a step-size check, and table cleanup before the claims are solid.","tokens_in":15056,"tokens_out":3071,"would_cite":false,"duration_ms":31217,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M15","70H05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Hamiltonian neural network trained on multi-step rollouts, coupled to an Unscented Kalman Filter, predicts long-horizon orbits while conserving energy.","keywords":["Hamiltonian neural networks","autoregressive training","unscented Kalman filter","orbit determination","data assimilation","energy conservation","symplectic integrator","long-term prediction"],"falsifier":"Re-run the orbit experiment on a test set where periapsis passages fall between the 60-second sample points, and compare AHNN5 predictions against a reference trajectory produced with an adaptive-step symplectic integrator; if the error spikes on those orbits or the energy drifts, the paper's claim that AHNN preserves invariants under its tested discretization fails.","tokens_in":14010,"feed_emoji":"🛰️","tokens_out":7274,"duration_ms":71568,"temperature":0.7,"pith_summary":"This paper claims that a Hamiltonian neural network trained without derivative labels can learn the dynamics of an unknown mechanical system and, when coupled with an Unscented Kalman Filter, deliver long-horizon predictions that conserve energy and stay close to the true trajectory. The core proposal is the Autoregressive Hamiltonian Neural Network (AHNN), whose loss unrolls the learned flow over a window of $W$ steps and penalizes the accumulated prediction error against the true sequence, rather than the one-step derivative mismatch used by standard HNNs. On a spring-mass system and on J2-perturbed Molniya orbits, the authors report that AHNN5 outperforms MLP, NODE, and standard HNN baselines in state RMSE, and that the AHNN-based UKF gives the lowest errors while producing uncertainty estimates. If correct, this matters because it shows that physically structured neural surrogates can serve as the process model inside a Bayesian filter for orbit determination and other Hamiltonian systems.","feed_headline":"AHNN-KF cuts orbit prediction error by an order of magnitude","feed_subtitle":"Multi-step training preserves energy and a Kalman filter keeps long-horizon orbit forecasts on track.","key_machinery":"The load-bearing object is the autoregressive Hamiltonian loss of Eq. (4), which treats the network as a map $F_\\theta(x_k) = x_k + \\int_{t_k}^{t_{k+1}} J\\nabla H_\\theta\\,dt$ and trains it by unrolling $W$ composed copies of $F_\\theta$ against the true trajectory segment. This makes the training objective match the closed-loop use of the model and penalizes error growth over exactly the horizon the model will be asked to predict. The second mechanism is the Unscented Kalman Filter, which propagates a set of $\\sigma$ points through $F_\\theta$, reconstructs mean and covariance, and corrects the state with measurements; this supplies both real-time refinement and uncertainty quantification. The paper also uses a Huber loss and a fourth-order Runge-Kutta integrator with a fixed 60-second step for the orbit experiments.","core_discovery":"The paper's central claim is that the Hamiltonian structure of an unknown system can be learned from state sequences alone and then used as a predictive model that both conserves energy and maintains accuracy over long rollouts. The authors replace the standard HNN loss on state derivatives with an autoregressive loss, Eq. (4), that applies the learned flow $F_\\theta$ recursively for $W$ steps and averages the distance between the predicted and observed states. They then insert this flow into an Unscented Kalman Filter that propagates $\\sigma$ points through $F_\\theta$ and assimilates noisy position measurements. In their numerical experiments, AHNN5 achieves the best RMSE among MLP, NODE, HNN, and AHNN variants, with the UKF version AHNNKF5 best overall, and the reported energy RMSE for AHNN5 is around two orders of magnitude smaller than for MLP and NODE. The paper thereby claims a practical recipe for learning Hamiltonian dynamics where derivative measurements such as accelerations are unavailable.","pith_inferences":["Editorial inference: the fixed 60-second Runge-Kutta step is part of what the network learns, so retraining with a symplectic or adaptive-step integrator is a direct test of whether the reported gains survive a different discretization.","Editorial inference: the same AHNN-UKF recipe should transfer to other systems with known symplectic structure but unknown potentials, such as rigid-body rotation or charged-particle motion in a magnetic field, though the paper does not demonstrate this.","Editorial inference: the paper's uncertainty bands come from the UKF's Gaussian approximation and may be optimistic under model mismatch; comparing the reported $3\\sigma$ intervals against empirical error across many test orbits would quantify that."],"forward_implications":["If the central claim is correct, position-velocity telemetry alone is enough to train a prediction model that respects Hamiltonian structure, removing the need for noisy acceleration labels.","The best-performing model, AHNNKF5, provides both a low-error state estimate and a covariance, which is what real orbit determination and collision-assessment pipelines need.","Larger autoregressive windows improve prediction accuracy in the paper's ablations, at the cost of more expensive backpropagation through time; this gives practitioners a direct trade-off to tune.","The reported energy conservation means predicted trajectories will not exhibit the artificial spiral-in or spiral-out seen with MLP and NODE baselines over long horizons."],"supporting_citations":[{"why":"Introduces Hamiltonian neural networks and the derivative-based training baseline that AHNN extends and compares against.","marker":"(Greydanus et al., 2019)"},{"why":"Supplies the unscented transform and the UKF algorithm used to assimilate measurements.","marker":"(Julier & Uhlmann, 1997)"},{"why":"Provides the sigma-point weighting and Gaussian-assumption details used in the UKF implementation.","marker":"(Wan & Van Der Merwe, 2000)"},{"why":"Introduces neural ODEs, the ODE-solver-based training paradigm, and the NODE baseline the paper compares against.","marker":"(Chen et al., 2018)"},{"why":"Provides the high-order symplectic integrator used to generate the training trajectories for the orbit experiments.","marker":"(Kahan & Li, 1997)"},{"why":"Supplies the two-body gravitational potential model, including J2, used to define the Molniya-orbit Hamiltonian.","marker":"(Gurfil et al., 2016)"},{"why":"Provides the mean gravity field model that supplies the J2 coefficient used in the orbit simulations.","marker":"(Cheng et al., 2011)"}],"fun_headline_variants":["Autoregressive HNN plus Kalman filter cuts orbit error","Learn Hamiltonians from positions alone with AR-HNN and UKF","Bayesian data assimilation improves learned Hamiltonian models","AR-HNN with UKF conserves energy and beats baselines","Autoregressive Hamiltonian networks with UKF reduce long-horizon error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that a fixed 60-second Runge-Kutta step resolves the sharp periapsis acceleration of the Molniya orbit, so the learned Hamiltonian is not biased by the discretization.","fun_headline_variants_meta":{"raw":{"variants":["Autoregressive HNN plus Kalman filter cuts orbit error","Learn Hamiltonians from positions alone with AR-HNN and UKF","Bayesian data assimilation improves learned Hamiltonian models","AR-HNN with UKF conserves energy and beats baselines","Autoregressive Hamiltonian networks with UKF reduce long-horizon error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001406,"raw_usage":{"total_tokens":5630,"prompt_tokens":841,"completion_tokens":4789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":4703}},"tokens_in":457,"tokens_out":4789,"duration_ms":31919,"temperature":1.0,"reasoning_tokens":4703,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:23:41.989337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the orbit experiment on a test set where periapsis passages fall between the 60-second sample points, and compare AHNN5 predictions against a reference trajectory produced with an adaptive-step symplectic integrator; if the error spikes on those orbits or the energy drifts, the paper's claim that AHNN preserves invariants under its tested discretization fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the unscented transform and the UKF algorithm used to assimilate measurements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the sigma-point weighting and Gaussian-assumption details used in the UKF implementation."},{"cited_title":"and Li, R.-C","cited_arxiv_id":null,"evidence_quote":"Provides the high-order symplectic integrator used to generate the training trajectories for the orbit experiments."},{"cited_title":"K., et al","cited_arxiv_id":null,"evidence_quote":"Supplies the two-body gravitational potential model, including J2, used to define the Molniya-orbit Hamiltonian."}],"review_version":1}