{"id":"bd8c3516-9e38-4048-9742-0fbf1f48639e","arxiv_id":"1909.01074","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Learning a physical system's energy from data is shown to work better when the data include dissipation, because dissipation breaks the hidden-symmetry (Casimir) degeneracy that pure reversible dynamics cannot reveal.","lead":"This paper views machine learning as a thermodynamic process: the algorithm removes unimportant details, just as entropy drives a gas toward equilibrium. The authors show that when learning the energy of a spinning body from its motion, adding friction reveals parts of the energy that pure mechanics hides.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. 6 claim that dissipation generically lifts Casimir degeneracy is only demonstrated for the energetic Ehrenfest dissipation in Eq. (23); for other Casimir-conserving dissipations the full energy may remain unidentifiable.","rationale":"I agree with the reader's verdict. The reversible half of the claim is solid: for the bracket (19), any function of m^2 is a Casimir, so E and E + lambda m^2 generate identical Hamiltonian trajectories, and the paper's tau = 0 result is a correct numerical illustration. The dissipative half is the load-bearing point. The demonstration uses exactly one dissipation mechanism, Eq. (23), and the conclusion is stated in Sec. 6 without qualification. The counterexample above shows that another Casimir-conserving dissipation that still dissipates energy leaves the Casimir component of the energy exactly unidentifiable; if a real system dissipates through that mechanism, the central claim's promise does not hold. The public code is a useful reproducibility check, but a single in-sample fit cannot establish a general identifiability theorem. Since this is precisely the weakness the reader identified, no verdict change is needed; CONDITIONAL remains appropriate.","tokens_in":17146,"tokens_out":8971,"duration_ms":98292,"concrete_test":"Generate a noiseless synthetic trajectory from m_dot = m x A m - gamma (I - m_hat m_hat^T) A m with a known symmetric positive-definite A of nonzero trace, using the same Crank-Nicolson integrator and timestep as Sec. 5. Then fit the parameters of A (and gamma) by least squares, assuming this alternative dissipation family and known Poisson bivector. Repeat over several initial conditions and gamma values. If the fitted A is determined only up to adding 2 lambda I, so that the trace component cannot be recovered, the Sec. 6 generalization fails; if lambda is identified, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The general claim in Sec. 6, that 'when having both the reversible and irreversible (dissipative) terms, the whole energy can be learned,' is established only for the energetic Ehrenfest dissipation of Eq. (23). That term is special because it contains the Hessian of E. Under E -> E + lambda m^2, the reversible part is unchanged since L m = 0, but the dissipative part changes by -tau lambda L^T L grad E, which is generically nonzero because L^T L = |m|^2 I - m m^T. The Casimir shift therefore becomes observable, but only because the dissipation mechanism is built from the full Hessian. The property is not generic. For a Casimir-conserving gradient-type dissipation, m_dot = m x grad E - gamma (I - m_hat m_hat^T) grad E, which also conserves |m|^2 and dissipates E, the shift E -> E + lambda m^2 changes grad E by 2 lambda m, and this lies in the kernel of the projector; the observed trajectory is identical for all lambda. Hence the full energy remains unidentifiable for every gamma > 0. The Sec. 6 conclusion should therefore be read as applying to the dissipation model of Eq. (23), not to dissipative dynamics in general. The paper itself states in Sec. 5 that the Poisson bivector, timestep, quadratic Ansatz, and dissipation mechanism are assumed known, so the generality of the identifiability result is not demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a thermodynamic interpretation of machine learning, in which learning/reduction is viewed as a dissipative dynamics driven by entropy within the GENERIC framework, while reversible Hamiltonian evolution is interpreted as propagation within a level of description. The authors illustrate the idea with POD, where the Rayleigh quotient is interpreted as an entropy and eigendecomposition as a gradient flow, and with a rigid-body example in which the energy is learned from angular-momentum trajectories. The central new claim is that for dynamics with non-canonical Poisson brackets, the energy can be reconstructed only up to Casimirs of the bracket from reversible trajectories, but when dissipation is present the full energy becomes identifiable; this is demonstrated numerically for the energetic Ehrenfest regularization of Eq. (23).","tokens_in":17512,"tokens_out":6855,"duration_ms":65402,"significance":"If the general identifiability claim were established, the paper would provide a useful bridge between non-equilibrium thermodynamics and data-driven modeling: the Casimir obstruction is a real and often ignored issue in learning Hamiltonian systems, and the idea that dissipation of the specific Ehrenfest type reveals Casimirs is an interesting and testable observation. The paper ships reproducible code for the rigid-body experiment and states a falsifiable prediction (full energy identifiable with dissipation of the form (23)). The mathematical facts invoked about Casimirs and about the gradient-flow representation of eigenproblems are correct. However, the paper is primarily conceptual; it does not provide a general identifiability theorem, and the only quantitative evidence is a self-consistency check on data generated by the same model. The significance is therefore moderate: as a perspective piece with a concrete illustration, it is valuable; as a general result, it is incomplete.","major_comments":[{"comment":"The claim in the abstract, introduction, and conclusion that 'when having both the reversible and irreversible (dissipative) terms, the whole energy can be learned' is stated without qualification, but the demonstration only covers the energetic Ehrenfest dissipation of Eq. (23), where the dissipative term contains the Hessian of E and is therefore sensitive to Casimir shifts E -> E + λ |m|^2 because L^T L = |m|^2 I - m m^T. For a generic Casimir-conserving dissipation that is a gradient projected onto the sphere, e.g. dm/dt = m × grad E - γ (I - m_hat m_hat^T) grad E with γ > 0, the shift changes grad E by 2λ m, which lies in the kernel of the projector, so the observed trajectory is identical for all λ and the full energy remains unidentifiable. The paper itself states in Sec. 5 that the Poisson bivector, timestep, quadratic Ansatz, and dissipation mechanism are assumed known. The general conclusion is therefore not established; the authors should either prove an identifiability condition for the class of GENERIC dissipations under consideration or restrict the claim to the specific dissipative structure of Eq. (23).","section":"Sec. 6 and Sec. 5, Eq. (23)"},{"comment":"The numerical experiment is an in-sample self-consistency check: the trajectories are generated by integrating Eq. (23) with a known quadratic energy and known τ, and the learning algorithm fits the same parametric family to those same trajectories. The paper does not report any out-of-sample prediction, noise sensitivity, or cross-validation, so the demonstration supports identifiability of the parameters from perfectly observed data only, not the ability to learn energy from real noisy data. The text should state this explicitly (or add an out-of-sample experiment) so the reader does not infer a stronger claim than the experiment supports.","section":"Sec. 5"}],"minor_comments":[{"comment":"The sentence 'By adding dissipation, the Casimirs now play a role in the dynamics, and the can be learned from the trajectory' contains a typo: 'the' should be 'they'.","section":"Sec. 5, paragraph after Fig. 2"},{"comment":"The caption states both that the energy error 'eventually vanishes as the dissipation becomes significant' and that 'the higher the exact value of the coefficient (stronger dissipation), the higher is the error between the exact and learned values'; these statements appear contradictory, and the text should clarify whether the latter refers only to the error in τ and why that error increases.","section":"Fig. 2 caption"},{"comment":"The entropy S(x) in Eq. (12) is introduced as an interpretation ('We shall interpret ...'), not as a derived quantity; stating explicitly that this is an analogy would help readers distinguish between the thermodynamic formalism and the POD algorithm.","section":"Sec. 3.2.4"},{"comment":"The free-surface-liquid illustration is a summary of previous work [12] with no new simulation; consider shortening it or clearly labeling it as a review so that the novel contribution of the paper is not obscured.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"This is a perspective-style paper that is likely to interest readers of the journal. The main weakness is the gap between the demonstrated rigid-body example and the general conclusions in Sec. 6; if the authors qualify the claims and explicitly state the limitations of the demonstration, the paper would be acceptable. I have no other concerns about novelty or authorship."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The part that matters: the rigid-body example makes a genuine point. With only the reversible Hamiltonian part, the energy is determined only up to Casimirs of the Poisson bracket; adding the energetic Ehrenfest dissipation of Eq. (23) makes a Casimir shift observable and the full energy identifiable. That is a clean, correct observation, and it gives people working on physics-informed learning something concrete to build on. The framing of POD/LLE as gradient dynamics of a Rayleigh-quotient entropy is also faithful to the Brockett/Absil lineage and is a useful way to think about dimensionality reduction, even if it is not new. The code is available, which is good.\n\nThe problems are in the scope of the claim and the strength of the demonstration. Section 6 states that 'when having both the reversible and irreversible (dissipative) terms, the whole energy can be learned.' That is only established for the energetic Ehrenfest dissipation. The term contains the Hessian of E, and the Hessian is exactly what makes a λm² shift visible. For a different Casimir-conserving dissipation, say m_dot = m × ∇E − γ(I − m̂m̂^T)∇E, the shift lies in the kernel of the projector, trajectories are identical for all λ, and the energy remains unidentifiable for every γ > 0. The paper itself notes in Sec. 5 that the Poisson bivector, timestep, and dissipation mechanism are assumed known, so the general conclusion exceeds the evidence.\n\nSecond, the numerical demonstration is an in-sample oracle test. They generate trajectories from Eq. (23) with a known quadratic E and τ, then fit the same parametric form back using curve_fit, with no noise, no error bars, and no validation on a held-out trajectory or initial condition. That shows identifiability in the ideal case; it does not show that the method works on real data. The trend in Fig. 2—error going down as τ grows—is consistent with the identifiability argument, but it is not a performance claim.\n\nNone of this is fatal. The core observation is correct and the paper is honest about its assumptions. What it needs is a tempering of the generality (or a proof for a class of GENERIC dissipations) and a noisier, out-of-sample experiment. That is a solid revision, not a rewrite.\n\nI would send it to a referee. It is a legitimate contribution to the model-discovery literature, and the Casimir point deserves to be circulated and debated.","headline":"Worth reading for the Casimir-lifting observation, but the paper overgeneralizes from one special dissipation model and its numerical demo is an in-sample fit.","tokens_in":18012,"tokens_out":3668,"would_cite":true,"duration_ms":35911,"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":"When learning the dynamics of a damped rigid body, adding dissipation removes the energy ambiguity that pure Hamiltonian trajectories leave behind.","keywords":["machine learning","thermodynamics","GENERIC","Poisson bracket","Casimir invariants","rigid body","dimensionality reduction","dissipation"],"falsifier":"Simulate a rigid body with the same quadratic energy but a different Casimir-preserving dissipation, for example Rayleigh damping proportional to angular velocity, fit the same quadratic Ansatz to the trajectory, and check whether the learned inverse inertia tensor converges to the exact one as the trajectory lengthens; if it does not, the identifiability result is specific to the energetic Ehrenfest regularisation.","tokens_in":16939,"feed_emoji":"🌀","tokens_out":5082,"duration_ms":48813,"temperature":0.7,"pith_summary":"This paper argues that machine learning of physical dynamics should be understood as a form of thermodynamic reduction: extracting a pattern from data is a dissipative process driven by entropy, while reversible evolution is Poisson (Hamiltonian) propagation within a level of description. Using the GENERIC structure as its template, it shows that when learning a system with a non-canonical Poisson bracket from trajectory data, the energy can be recovered only up to the Casimir invariants of the bracket. If the observed dynamics also contains dissipation of the energetic Ehrenfest type, the Casimir ambiguity disappears and the full energy functional can be learned from the trajectory. The claim is demonstrated on a freely rotating damped rigid body, where kinetic energy is reconstructed from angular-momentum trajectories once $\tau>0$. This matters because real systems typically dissipate, so the dissipative part of data can be used as a resource rather than a nuisance for identifying physics.","feed_headline":"Dissipation lifts the blind spot in physics learning","feed_subtitle":"On a damped rigid body, irreversible dynamics makes the full kinetic energy learnable from angular-momentum data.","key_machinery":"The load-bearing object is the GENERIC evolution equation (General Equation for Non-Equilibrium Reversible-Irreversible Coupling), written as $\\dot x_i = L_{ij}\\,\\partial E/\\partial x_j + \\partial\\Xi/\\partial x^*_i$ evaluated at $x^* = \\partial S/\\partial x$, which splits dynamics into a reversible Hamiltonian part generated by a Poisson bivector $L$ and an irreversible gradient part driven by a dissipation potential $\\Xi$. For the rigid body the Poisson bracket is non-canonical, $\\{F,G\\} = -\\boldsymbol m \\cdot (\\partial F/\\partial \\boldsymbol m \\times \\partial G/\\partial \\boldsymbol m)$, and the Casimir $|\\boldsymbol m|^2$ is conserved by the reversible part. Adding the energetic Ehrenfest regularization, $\\dot{\\boldsymbol m} = \\boldsymbol m \\times \\partial E/\\partial \\boldsymbol m - (\\tau/2)\\, L^T (\\partial^2 E/\\partial \\boldsymbol m \\partial \\boldsymbol m)\\, L\\,(\\partial E/\\partial \\boldsymbol m)$, keeps $|\\boldsymbol m|^2$ fixed while dissipating kinetic energy, which makes the Casimir visible in the trajectory and therefore learnable.","core_discovery":"The paper's central claim is that reversible dynamics alone carries no information about the Casimirs of its Poisson bracket, so any energy learned from a purely Hamiltonian trajectory is determined only up to a shift by a function of those Casimirs (for the rigid body, up to a multiple of $|\\boldsymbol m|^2$). When the same system is observed with dissipation included, the Casimirs enter the dynamics and the full energy becomes identifiable. The authors verify this on the damped rigid body: with $\tau=0$ the learned quadratic energy differs from the exact one by a Casimir shift, while for $\tau>0$ the learned matrix of second derivatives converges to the exact inverse inertia tensor.","pith_inferences":["The identifiability claim is demonstrated for one parametric family of dissipation (the energetic Ehrenfest regularization); a natural extension would test whether other Casimir-preserving dissipation mechanisms, such as linear Rayleigh damping, also lift the degeneracy.","If dissipation lifts the energy ambiguity generally, then deliberately adding controlled dissipation to experimental protocols could become a practical identification technique for the energies of non-canonical systems.","The entropy-driven learning picture suggests that fluctuations in the database carry thermodynamic information about the learning dynamics itself; analysing those fluctuations could supplement deterministic trajectory fitting."],"forward_implications":["For any system with a non-canonical Poisson bracket, purely Hamiltonian trajectory data determine the energy only up to the Casimir invariants of the bracket; generative models that ignore this will return shifted energies.","Adding irreversible dynamics of the type considered in Eq. (23) makes the full energy identifiable, so dissipative data are not a nuisance but a resource for learning physics.","The thermodynamic viewpoint supplies the missing embedding between reduced and detailed manifolds: reduction is entropy-driven and the embedding can be constructed via the maximum-entropy principle, connecting manifold learning methods such as POD and LLE with GENERIC.","Learned reduced models with GENERIC structure can be integrated to predict future states on the reduced manifold and then lifted to the original data space, giving predictions in the rich space where validation is meaningful."],"supporting_citations":[{"why":"Introduces the GENERIC form that the paper uses as the template for splitting reversible Poisson evolution from irreversible gradient dynamics.","marker":"[14, 15, 16, 17]"},{"why":"Defines the energetic Ehrenfest regularization, the specific dissipative term that makes the Casimir learnable in the rigid-body example.","marker":"[38]"},{"why":"The earlier machine-learning study of free-surface fluids that this paper recasts in thermodynamic terms and uses as motivation.","marker":"[12]"},{"why":"Poisson reduction results that give the non-canonical rigid-body bracket and justify using angular momentum as the state space.","marker":"[53, 54, 55]"},{"why":"Supplies locally linear embedding as a competing manifold-recognition method whose projection can also be viewed as gradient dynamics.","marker":"[5]"},{"why":"Supplies the POD method whose thermodynamic interpretation is developed in Section 3.2 as an entropy-driven reduction.","marker":"[42]"}],"fun_headline_variants":["Reversible dynamics hides energy, damping reveals it","Casimirs hidden by reversible dynamics exposed by damping","Dissipation is the key to full energy identification","Only damping reveals the full energy in physics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the observed dissipation has the specific energetic Ehrenfest form of Eq. (23), that the quadratic Ansatz for the energy is correct, and that the Poisson bivector and timestep of the numerical scheme are known; if a real system dissipates through a different Casimir-preserving mechanism, the paper does not establish that the full energy remains identifiable.","fun_headline_variants_meta":{"raw":{"variants":["Reversible dynamics hides energy, damping reveals it","Casimirs hidden by reversible dynamics exposed by damping","Dissipation is the key to full energy identification","Only damping reveals the full energy in physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000828,"raw_usage":{"total_tokens":3540,"prompt_tokens":791,"completion_tokens":2749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":2690}},"tokens_in":407,"tokens_out":2749,"duration_ms":19474,"temperature":1.0,"reasoning_tokens":2690,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:27:26.608775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a rigid body with the same quadratic energy but a different Casimir-preserving dissipation, for example Rayleigh damping proportional to angular velocity, fit the same quadratic Ansatz to the trajectory, and check whether the learned inverse inertia tensor converges to the exact one as the trajectory lengthens; if it does not, the identifiability result is specific to the energetic Ehrenfest regularisation.","supporting_citations":[{"cited_title":"Pavelka, V","cited_arxiv_id":null,"evidence_quote":"Defines the energetic Ehrenfest regularization, the specific dissipative term that makes the Casimir learnable in the rigid-body example."},{"cited_title":"Roweis, L.K","cited_arxiv_id":null,"evidence_quote":"Supplies locally linear embedding as a competing manifold-recognition method whose projection can also be viewed as gradient dynamics."},{"cited_title":"Chatterjee, Current Science78(7) (2000)","cited_arxiv_id":null,"evidence_quote":"Supplies the POD method whose thermodynamic interpretation is developed in Section 3.2 as an entropy-driven reduction."}],"review_version":1}