{"id":"b1de32dd-0917-4fe1-b14f-59d2d702d43e","arxiv_id":"2608.12624","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"S-PENNs inject epistemic uncertainty into hard-constrained GENERIC neural networks blockwise, preserving energy conservation and nonnegative entropy production in every sample, and use split conformal prediction to calibrate the intervals.","lead":"The paper attaches small auxiliary 'epinet' networks to each thermodynamic building block of a GENERIC dynamics model, so every sampled uncertainty realization is thermodynamically consistent by construction. It reports well-calibrated prediction intervals at one to three orders of magnitude lower cost than deep ensembles on two ODE systems and a one-dimensional viscoplastic PDE.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dissipation-potential epinet is conditioned on base hidden features that depend on x*, so PICNN convexity in x* is not inherited; the augmented potential need not be convex and second-law admissibility is not guaranteed by construction.","rationale":"The reader's conditionality focuses on noise-free, fully observed data and missing code release; these are practical limitations rather than failures of the paper's internal construction. The load-bearing issue identified here is internal: the central claim is that every sampled S-PENNs realization is thermodynamically admissible 'by construction'. That claim requires the dissipation potential to remain convex in x* after the epinet perturbation. The paper asserts PICNN epinets preserve convexity, but the epinet is conditioned on base hidden features h^~Ξ_ψ(x,x*) that themselves depend on x*. PICNN convexity is guaranteed only for fixed auxiliary inputs; an input-dependent auxiliary input breaks the guarantee. Stop-gradient does not alter the forward map. A concrete realizable choice (h=(x*)^2, z=1, σ=-h) makes \\tilde Ξ_ϑ nonconvex, so the proof of nonnegative entropy production is missing. The reported positive min(Ṡ) in the experiments is an empirical outcome, not a structural certificate. Because the headline contribution is precisely a by-construction structural guarantee, this gap moves the verdict from CONDITIONAL to REJECT unless the construction is amended and a valid convexity proof is provided.","tokens_in":29590,"tokens_out":21459,"duration_ms":219613,"concrete_test":"Fix a scalar base PICNN with convex hidden feature h(x*) = (x*)^2 and set dz=1, z=1. Choose epinet weights so that the learnable output is σ = -h and the prior is zero; this is realizable under Eqs. (10)-(13) because h is an auxiliary input with freely signed weights. Evaluate the augmented raw potential \\tilde Ξ_ϑ(x,x*,z) on x* in [-2,2] and compute its second derivative with respect to x*: it is negative on that interval, so \\tilde Ξ_ϑ is not convex and the Bregman construction in Eq. (18) cannot certify nonnegative entropy production. If the authors add an explicit constraint on the epinet's dependence on h (e.g., monotonicity) and prove convexity, the concern is resolved.","verdict_should_be":"REJECT","load_bearing_attack":"Section 2.3.2 states that PICNNs for the raw dissipation-potential epinet ensure the outputs are convex in x*. This does not follow from the stated architecture. In Eq. (13), the epinet receives \\bar h^~Ξ_ψ(x,x*) = [sg(h^~Ξ_ψ(x,x*)), x, x*], where h^~Ξ_ψ(x,x*) is the base network's hidden feature and is itself a function of x*. A PICNN guarantees convexity in its designated convex input only when all other inputs are held fixed. Here, one auxiliary input varies with x*, so the composition F(x*, h(x,x*), z) need not be convex even though F(·, h, z) is convex for each fixed h. Stop-gradient changes training, not the forward function. A minimal allowed instance: take a base PICNN whose hidden feature is h = (x*)^2, set dz=1, z=1, and let the learnable epinet output be σ = -h. This is realizable in Eqs. (10)-(13) because h enters as an auxiliary input with freely signed weights. Then σ = -(x*)^2, so the augmented raw potential \\tilde Ξ_ϑ = \\tilde Ξ_ψ + σ can be nonconvex. Consequently, the Bregman-type reparameterization in Eq. (18) does not guarantee Ξ_ϑ ≥ 0, and nonnegative entropy production is not guaranteed by construction. The 'second law by construction' claim therefore rests on an unproven composition property.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes S-PENNs, a framework for uncertainty quantification in structure-preserving GENERIC neural dynamics. It attaches lightweight epinets to each thermodynamic building block of a pretrained N-GENNs model, using a shared epistemic index to generate joint perturbations, and then reassembles the augmented blocks via the same structure-preserving reparameterizations as the base model. The authors claim every sampled vector field is thermodynamically admissible by construction, thereby preserving energy and nonnegative entropy production; split-conformal calibration provides finite-sample marginal coverage intervals. Numerical experiments on two ODE systems and one PDE system compare S-PENNs with deep ensembles and MC dropout, reporting competitive accuracy at substantially lower computational cost.","tokens_in":29897,"tokens_out":10071,"duration_ms":90831,"significance":"The paper addresses an important gap: UQ for hard-constrained structure-preserving models without breaking the constraints. The block-wise epinet design with a common epistemic index is elegant and general, and the input-independent branch for global tensors is a useful contribution. The numerical study is extensive, with proper scoring rules, calibration curves, and wall-clock comparisons across problems of increasing complexity. If the structural guarantee were fully established, S-PENNs would be a valuable low-cost alternative to deep ensembles in mechanics applications; the paper is clearly written and the ideas are reproducible in principle.","major_comments":[{"comment":"The claim that each epistemic realization is thermodynamically admissible is not established for the dissipation potential. The epinet input in Eq. (13) includes the base hidden feature h^Ξ_ψ(x,x*), which is a function of x*. A PICNN guarantees convexity in x* only for fixed auxiliary inputs; when an auxiliary input itself depends on x*, the composition need not be convex. For example, with d_z=1 and z=1, take the hidden feature h=(x*)^2 and let the learnable epinet output σ=-h, which is realizable through Eq. (10) by setting NN_φ([\\bar h,z])^T z to output -h. Then \\tilde Ξ_ϑ(x,x*,z)=\\tilde Ξ_ψ(x,x*)-(x*)^2, which need not be convex in x*. Consequently, the Bregman-type reparameterization in Eq. (18) does not guarantee Ξ_ϑ≥0 or nonnegative entropy production for each realization. The subsequent paragraph about nonnegative coefficients of the epistemic index does not resolve this, since the x*-dependence through the base features is left unconstrained. This undermines the central \"by construction\" claim stated in the abstract and in Section 2.3.2.","section":"Section 2.3.2, Eqs. (13) and (18)"}],"minor_comments":[{"comment":"The text states that the harmonic-oscillator dataset consists of 100 trajectories, but the trajectory-level split is 60 training + 50 calibration + 40 testing = 150 trajectories; this arithmetic inconsistency should be corrected.","section":"Section 3.1.1"},{"comment":"The conformal guarantee in Eq. (25) is pointwise in j, whereas the empirical coverage curves in Figures 2, 4, and 6 average over forecast entries; the text should clarify that the displayed EC is an aggregate diagnostic and is not itself covered by the finite-sample guarantee.","section":"Section 2.4 and Appendix A"},{"comment":"The paper uses both \"N-GENNs\" and \"N-GINNs\" for the same model family; the terminology should be standardized throughout the text.","section":"Section 2.1 and Reference [23]"},{"comment":"In the viscoplastic example, the notation in Eq. (22) suppresses the dependence of the predictor on boundary histories, which are part of the rollout input; including them in the notation would make the exchangeability argument for the PDE case more transparent.","section":"Section 3.3 and Eq. (22)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and the main numerical results are convincing. The central construction claim, however, rests on an unproven composition property for the dissipation-potential epinet. This is fixable within the manuscript's scope by either constraining the epinet so that convexity in x* is genuinely preserved (e.g., removing x*-dependent auxiliary inputs from the convex path or requiring joint convexity) or by weakening the \"by construction\" language and supplementing it with explicit empirical checks of entropy positivity across a wider range of cases. The other issues are minor. Given that the core idea is promising, I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is the first epinet-based UQ I know that attaches block-wise perturbations to hard-constrained GENERIC building blocks, and the engineering is genuinely good. The shared epistemic index, the input-independent branch for global tensors, and the split-conformal calibration are real additions over NEON/E-PINNs and the rest. The numerical work is also honest: three test problems, proper scoring rules, wall-clock comparisons against deep ensembles, and an explicit limitations section. The cost reductions look real, not cherry-picked.\n\nThe soft spot is in the headline guarantee. Section 2.3.2 says PICNNs for the dissipation-potential epinet ensure convexity in x*. But the epinet's input vector includes sg(h^{tilde Xi}_psi(x,x*)), and that hidden feature is a function of x*. A PICNN is convex in its designated variable only when the auxiliary inputs are held fixed. With h depending on x*, the composition does not inherit convexity; nonnegative z does not repair this, since nonnegative combinations of functions that are not known to be convex are not known to be convex. The augmented raw potential can therefore be nonconvex, and the projection reparameterization in Eq. (18) no longer yields Xi >= 0. So 'second law by construction' is not established by the argument given. The numerical entropy-production checks are consistent with the claim, but they are empirical evidence, not a guarantee.\n\nI checked whether this is a nit: it is not. The claim that every realization is admissible by construction is load-bearing for the paper's positioning. It can likely be fixed by making the dissipation-potential epinet depend on x* only through a genuinely convex parameterization, or by softening the claim to approximate structure preservation in practice. As written, the theory section overreaches.\n\nMinor issues: the harmonic oscillator data split says 100 trajectories but 60+50+40 = 150; no code or data is shipped; no repeated-seed variability is reported for the stochastic methods. None of these changes my overall read.\n\nWho should read it: people building UQ for structure-preserving SciML. It deserves a serious referee, but I would send it back for major revision on the convexity argument before accepting the structural guarantee.","headline":"A well-engineered epinet-based UQ wrapper for constrained GENERIC dynamics whose central 'second law by construction' claim does not survive the PICNN composition argument as written.","tokens_in":30464,"tokens_out":6062,"would_cite":true,"duration_ms":62650,"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":"S-PENNs attach block-specific epinets to a frozen N-GENNs backbone so that every sampled GENERIC dynamics conserves energy and produces nonnegative entropy.","keywords":["structure-preserving machine learning","epistemic neural networks","GENERIC dynamics","thermodynamic consistency","uncertainty quantification","conformal prediction","N-GENNs","physics-informed machine learning"],"falsifier":"Take the harmonic-oscillator benchmark and corrupt the training states with additive Gaussian noise, for example with standard deviation $0.05$ on $q$ and $p$, then retrain the S-PENNs two-stage procedure. If the median energy rate across 2000 sampled rollouts deviates from zero by more than about $10^{-5}$ (against the $10^{-7}$ scale reported for clean data) or the pointwise minimum entropy-production rate becomes negative on the test set, then the noise-free premise is doing load-bearing work rather than the architecture alone.","tokens_in":1980,"feed_emoji":"⚙️","tokens_out":7938,"duration_ms":115625,"temperature":0.7,"pith_summary":"Uncertainty quantification usually works by randomizing weights or outputs, which can push a physics-constrained model out of the admissible set. S-PENNs instead attach a small epinet to each thermodynamic building block of a pretrained N-GENN—energy, entropy, reversible operator, and dissipation potential—and drive all epinets with one shared epistemic index. Because the perturbed blocks are reassembled through the same skew-symmetry and projection reparameterizations used by N-GENNs, every draw of the index defines a GENERIC vector field that conserves energy and produces nonnegative entropy by construction. Split conformal prediction then rescales the ensemble variance into intervals with finite-sample marginal coverage. On a harmonic oscillator coupled to a heat bath, an idealized chemical motor, and a one-dimensional viscoplastic model, the method attains near-ensemble accuracy at one to three orders of magnitude lower cost.","feed_headline":"Every uncertainty sample obeys the 1st and 2nd laws","feed_subtitle":"Lightweight epinets keep every UQ sample thermodynamically valid at a fraction of ensemble cost.","key_machinery":"The machinery is the S-PENN block-epinet assembly. Each constrained building block of the N-GENNs backbone is augmented with an epinet conditioned on stop-gradient base features and the same epistemic index $z$; for the global tensor $B$, an input-independent linear epinet $B_\\vartheta(z)=B_\\psi+\\sum_{n=1}^{d_z} z_n \\phi_n^B$ keeps the perturbation state-independent. The same two reparameterizations as in N-GENNs then close the admissible class: the reversible operator is sandwiched by skew-symmetric matrices $A^i_\\vartheta(z)=B^i_\\vartheta(z)-(B^i_\\vartheta(z))^\\top$ built from the entropy gradient, which kills $L_\\vartheta DS_\\vartheta$, and the dissipation potential is projected with $P_{E_\\vartheta}(x,z)=I-DE_\\vartheta DE_\\vartheta^\\top/\\|DE_\\vartheta\\|^2$, which kills the energy dependence and preserves convexity when the epistemic-index distribution has nonnegative support. One draw of $z$ therefore yields one admissible GENERIC vector field $g_\\vartheta(x,z)=L_\\vartheta(x,z)DE_\\vartheta(x,z)+D_{x^*}\\Xi_\\vartheta(x,x^*,z)|_{x^*=DS_\\vartheta(x,z)}$.","core_discovery":"The central discovery is that hard constraints need not be re-derived for each uncertainty sample: the admissible class itself can be made into a sample space. Treating the deterministic N-GENNs as frozen base networks, S-PENNs adds block-specific epinets $\\sigma_\\varphi(\\cdot,z)$ to $E_\\psi$, $S_\\psi$, $\\tilde{L}_\\psi$, and $\\tilde{\\Xi}_\\psi$, with a dedicated input-independent linear epinet for the global skew-symmetry tensor $B_\\psi$. Reassembling through $L_\\vartheta(x,z)=Q_{S_\\vartheta}(x,z)^\\top \\tilde{L}_\\vartheta(x,z)Q_{S_\\vartheta}(x,z)$ and the energy-orthogonal projection reparameterization of the dissipation potential enforces both GENERIC degeneracy conditions for each fixed $z$, while a partially input-convex network keeps the dissipation potential convex. Every sampled rollout therefore obeys $\\dot{E}=0$ and $\\dot{S}\\ge 0$ at the trajectory level, and the shared index couples the block perturbations so that uncertainty propagates coherently through the coupled dynamics.","pith_inferences":["A direct extension not pursued in the paper: the same idea applies to symplectic, metriplectic, and polyconvex constitutive networks, as long as each epinet respects the closure property of its block; a natural testbed would be hyperelasticity with convexity enforced by input-convex epinets.","The paper fixes the epistemic-index distribution to nonnegative support to preserve convexity, so prior choice is coupled to admissibility rather than being a free Bayesian prior; one could instead parameterize the dissipation epinet with a nonnegative output map, allowing signed epistemic indices and a wider family of priors.","The trajectory-level exchangeability assumption used for conformal calibration will be violated when a system is monitored online under drifting regimes, and an adaptive or online conformal variant would be needed to keep coverage guarantees in that setting.","Because the base network is frozen, S-PENNs uncertainty covers only the residual between the admissible class and the data, not the full model-class uncertainty; if the base N-GENN is trained on one parameter regime, intervals will remain narrow even where extrapolation fails, which is testable by training on one regime and checking coverage in another."],"forward_implications":["A single trained S-PENN can produce thousands of physically admissible rollouts from one frozen backbone, so downstream design, control, and reliability computations can average over stochastic dynamics without filtering out invalid samples.","Split conformal calibration makes interval coverage a finite-sample guarantee under trajectory-level exchangeability, independent of how well the predictive distribution is specified.","The cost gap measured on the three benchmarks—one to three orders of magnitude less wall time than deep ensembles—should grow with model dimension, since epinet sampling avoids training complete constrained models.","The same block-epinet recipe extends to other hard- or soft-constrained models, including frozen or partially frozen pretrained models, whenever the epinet is chosen from the same admissible function class as the block it perturbs.","The input-independent epinet branch gives a construction for global unknown parameters in inverse problems, with parameter estimates comparable to a Hamiltonian Monte Carlo Bayesian physics-informed neural network on the Kraichnan–Orszag and Korteweg–de Vries benchmarks."],"supporting_citations":[{"why":"Supplies the frozen N-GENNs backbone, its structure-preserving reparameterizations, and the three benchmark problems used throughout.","marker":"[23]"},{"why":"Defines the epinet and shared epistemic index construction that S-PENNs generalizes to constrained building blocks.","marker":"[47]"},{"why":"Gives the split conformal rank argument and finite-sample marginal coverage guarantee used for calibration.","marker":"[52]"},{"why":"Defines the deep-ensemble baseline whose cost and predictive accuracy S-PENNs is compared against.","marker":"[28]"},{"why":"Introduces the convex dissipation-potential parameterization that the dissipation epinet inherits.","marker":"[24]"},{"why":"Provides the Bayesian physics-informed neural network baseline used in the inverse-problem benchmarks in Appendix D.","marker":"[30]"}],"fun_headline_variants":["Physics-safe UQ, orders of magnitude cheaper","Every sample obeys the laws, without the cost","Thermodynamics-compliant uncertainty, fast","Structure-preserving UQ: physics guaranteed"],"cache_read_input_tokens":32512,"weakest_assumption_plain":"The framework assumes noise-free, fully observed state trajectories and a frozen, well-specified pretrained base model; if the data are noisy or the base dynamics are misspecified, the thermodynamic constraints are imposed on a contaminated vector field and the learned dynamics and its uncertainty can be biased.","fun_headline_variants_meta":{"raw":{"variants":["Physics-safe UQ, orders of magnitude cheaper","Every sample obeys the laws, without the cost","Thermodynamics-compliant uncertainty, fast","Structure-preserving UQ: physics guaranteed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2443,"prompt_tokens":1077,"completion_tokens":1366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":1316}},"tokens_in":693,"tokens_out":1366,"duration_ms":12630,"temperature":1.0,"reasoning_tokens":1316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:03:59.195579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the harmonic-oscillator benchmark and corrupt the training states with additive Gaussian noise, for example with standard deviation $0.05$ on $q$ and $p$, then retrain the S-PENNs two-stage procedure. If the median energy rate across 2000 sampled rollouts deviates from zero by more than about $10^{-5}$ (against the $10^{-7}$ scale reported for clean data) or the pointwise minimum entropy-production rate becomes negative on the test set, then the noise-free premise is doing load-bearing work rather than the architecture alone.","supporting_citations":[{"cited_title":"Nonlinear GENERIC Informed Neural Networks (N-GINNs): learning GENERIC dynamics with non-quadratic dissipation potentials","cited_arxiv_id":"2605.09058","evidence_quote":"Supplies the frozen N-GENNs backbone, its structure-preserving reparameterizations, and the three benchmark problems used throughout."},{"cited_title":"Epistemic neural networks.Advances in Neural Information Processing Systems, 36:2795–2823, 2023","cited_arxiv_id":null,"evidence_quote":"Defines the epinet and shared epistemic index construction that S-PENNs generalizes to constrained building blocks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the convex dissipation-potential parameterization that the dissipation epinet inherits."},{"cited_title":"B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data.Journal of Computational Physics, 425:109913, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian physics-informed neural network baseline used in the inverse-problem benchmarks in Appendix D."}],"review_version":1}