acceptodds
Under review as a conference paper at ICLR 2027

Stable probabilistic emulation of partially observed dynamical systems

Abstract

Predicting the long-term behavior of partially observed dynamical systems such as the coupled atmosphere-ocean system is central to societal decision-making. Motivated by the success of deep learning emulators in weather forecasting, data-driven climate models represent a promising approach. However, climate emulation poses distinct challenges: models must account for unresolved fast variables introduced by unavoidable coarse-graining, reproduce long-term statistics, and remain stable over arbitrarily long rollouts and under external perturbations. These requirements motivate stochastic data-driven emulators, but modern generative models lack stability guarantees. We introduce forced-dissipative stochastic interpolants, a generative framework for learning stochastic reduced dynamics. Its architecture uses structural constraints inspired by turbulent dynamics, which provide stability by design while retaining the flexibility of stochastic interpolants to represent complex probabilistic dynamics. Building on the theory of Markov-chain stability, we establish a Foster–Lyapunov drift condition and prove geometric ergodicity for our emulator, together with stability under steady external forcing. We evaluate the framework on the two-timescale Lorenz-96 system, a prototype of partially observed multiscale turbulence. The resulting emulator remains stable over long integrations and under finite-amplitude forcing, with accurate prediction of forced responses of both the mean and variance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
emulators, dynamical systems, stochastic interpolants, probabilistic forecasting, climate, stability, statistical responses, Markov chain

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xcazOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a stable probabilistic emulation framework for partially observed dynamical systems. Based on stochastic interpolants, the authors construct a stochastic generative process by learning a drift function to model the transition distribution of dynamical systems. To address the instability caused by accumulated approximation errors during long-term rollouts, the authors introduce forced-dissipative stochastic interpolants, where the drift function is parameterized with a structured decomposition into forcing, antisymmetric, dissipative, and nonlinear dissipative components. This design constrains energy growth while preserving flexibility for modeling complex dynamics. The paper further provides theoretical analysis based on Foster–Lyapunov conditions to establish stability and ergodicity properties of the proposed stochastic process. Experiments on Lorenz–96 systems demonstrate improved long-term statistical consistency and response under external forcing.

Critical strengths
  1. The paper provides a relatively rigorous theoretical analysis. The authors formally define stability conditions and establish stability and ergodicity guarantees of the proposed stochastic process based on Foster–Lyapunov theory.
Critical weaknesses
  1. The paper lacks a clear problem formulation and methodological narrative. It does not sufficiently explain the limitations of existing stochastic interpolant-based emulators and why the proposed drift parameterization can address instability during long-term rollouts, making the motivation and contribution less clear.
  2. The main novelty mainly lies in the structured parameterization of the drift function. However, the authors should better clarify the generality and broader impact of this design beyond the specific stochastic interpolant framework.
Questions for authors

The motivation for introducing the structured drift parameterization needs further clarification. What specific limitations arise when using conventional neural networks to parameterize the drift function? Compared with an unconstrained neural network drift, what theoretical or practical advantages does the proposed parameterization provide?

2 comments
AuthorsOpenReview ↗

Response to Reviewer xcaz

We thank the reviewer for the careful reading and for noting the rigor of the stability and ergodicity analysis. We agree that the presentation can be clearer, and the revised introduction states the motivation more explicitly. We respectfully disagree that the problem formulation is missing from the submission, and summarize it below.

Problem formulation

A generative emulator is a Markov chain: each new state is sampled from a learned transition kernel given the previous state. Climate emulators are run over long times, so the emulator must converge to a unique stationary distribution, which defines its climate (Definition 1). It must also remain stable under an external forcing that drives it out of its training distribution (Section 2.2). Accurate one-step transitions imply neither property, and the existing remedies for unstable rollouts do not guarantee them for probabilistic emulators (Section 1). The paragraph "Problem statement" at the end of Section 2.2 states the problem formally.

The failure of unconstrained neural networks

The introduction of the submission identifies the cause of this failure: small one-step errors accumulate over long rollouts and can lead to finite-time blow-up (Ainslie et al., 2026). A detailed analysis of this failure is a research question of its own, which requires dedicated tools. The submission takes this failure as its starting point and focuses on enforcing stability. The revised introduction adds the following explanation of the failure.

Training fits the drift only near the training data. Far from it, the drift is set by the architecture, whose extrapolation is not controlled (Xu et al., 2021). Accumulated errors then push the state out of distribution. Classical stochastic climate models show the same failure, which is addressed with energy-constrained models (Majda and Harlim, 2013).

Our experiments confirm this failure on all three systems (see the general comment above). On two-layer Lorenz-96, the diffusion model with the lowest validation loss is also unstable, so a better fit does not ensure stability.

The role of the forced-dissipative drift

The Foster–Lyapunov condition (Proposition 4.1) is a standard sufficient condition for the stability of a Markov chain (Meyn and Tweedie, 2009). It requires the energy to decrease in expectation whenever it is large, at all states. The forced-dissipative drift b=F+Ax+Dx+S(x)xb = F + Ax + Dx + S(x)x satisfies this condition for any trained weights. The terms AA and SS conserve the energy, which enters only through the bounded forcing FF and the noise and is removed by the dissipation DD at a guaranteed minimum rate (Proposition 3.1). This energy budget gives stability and ergodicity, with and without external forcing (Propositions 4.1 to 4.4).

This structure also matters for accuracy. We compare it with a simpler damped drift with a bounded network (see the general comment above). This drift also satisfies a Foster–Lyapunov condition, but its stationary distribution is inaccurate despite a lower validation loss. In the forced-dissipative drift, the energy-preserving structure of the transfer (A+S)x(A + S)x is imposed everywhere. In the simpler drift, the bounded network can only learn these transfers near the data. Dissipation and bounded forcing alone are thus not sufficient, and the energy-preserving structure is a key physical prior for accurate emulation.

Generality

The Foster–Lyapunov condition involves only the transition kernel, so it applies to any probabilistic emulator. The forced-dissipative constraint acts on the drift of a generative SDE, so it could also be imposed on other drift-based generative models, such as flow matching and diffusion models. We chose stochastic interpolants because their formulation makes the condition easier to establish. In all cases, the constraint requires no knowledge of the modeled equations.

The drift structure follows a design pattern used for deterministic models, such as stable port-Hamiltonian neural networks (Desai et al., 2021; Roth et al., 2025). The submission extends this pattern to probabilistic emulators, where it guarantees a unique stationary distribution, which is the climate of the emulator.

In the revision, the introduction and the end of Section 4 better explain why unconstrained drifts fail, and Appendix D.4 adds the comparison with the damped drift.

References

  • Ainslie et al. (2026). Eigenanalysis framework for autoregressive neural emulators of multi-scale chaotic dynamics.
  • Desai et al. (2021). Port-Hamiltonian neural networks for learning explicit time-dependent dynamical systems.
  • Majda and Harlim (2013). Physics constrained nonlinear regression models for time series.
  • Meyn and Tweedie (2009). Markov chains and stochastic stability.
  • Roth et al. (2025). Stable port-Hamiltonian neural networks.
  • Xu et al. (2021). How neural networks extrapolate: from feedforward to graph neural networks.
AuthorsOpenReview ↗

Follow-up response to Reviewer xcaz

Dear Reviewer,

As the discussion period ends on November 18, we would like to know whether our reply addresses your concerns. In particular, the reply clearly explains the benefits of our forced-dissipative architecture, and the revised introduction states this motivation more explicitly. The new comparison with a simpler damped drift baseline and the new Marshall–Molteni experiments, described in the general comment, also support this motivation.

We would be glad to clarify any remaining point before the deadline, and we would be grateful if you would reconsider your assessment in light of these answers.

Reviewer_8zUJOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces forced-dissipative stochastic interpolants for probabilistic emulation of partially observed dynamical systems. The generative drift combines bounded forcing, energy-preserving components, and a strictly dissipative component. Under the stated regularity and nondegeneracy assumptions, this structure yields a Foster–Lyapunov condition and geometric ergodicity for the autoregressive emulator defined by the continuous-time generative process. The analysis also establishes bounded second moments under bounded time-dependent perturbations and geometric ergodicity under constant perturbations.

Experiments consider fully observed single-layer Lorenz-96 dynamics and partially observed, spatially heterogeneous two-layer Lorenz-96 dynamics. Evaluation includes stationary marginal distributions over long rollouts, responses of the mean and variance to steady and oscillating forcing, and sensitivity to model capacity. The proposed emulator remains stable in the tested configurations and generally reproduces forced responses more accurately than the unconstrained generative baselines.

Critical strengths

The main strength is the constructive connection between the generative architecture and the long-term behavior of the resulting autoregressive Markov chain. The paper derives a Foster–Lyapunov condition from explicit constraints on the drift, providing a principled basis for stability under the stated assumptions. Extending the analysis to external perturbations is consequential: an emulator intended for response experiments must remain well behaved when its inputs move beyond the training distribution. The paper also appropriately distinguishes stability of the learned chain from accuracy relative to the physical system.

The evaluation of forced responses is a valuable complement to stationary-distribution matching. The experiments demonstrate that a model can reproduce stationary marginals while producing inaccurate responses or becoming unstable under forcing. Testing both mean and variance responses, including oscillating forcing, provides a more informative assessment than unforced rollouts alone. The capacity study further illustrates that increasing model size or obtaining a lower validation loss does not necessarily ensure reliable autoregressive behavior.

Critical weaknesses

The experiments do not sufficiently isolate the benefit of the full architectural decomposition. In particular, a simpler drift consisting of linear damping plus a uniformly bounded neural residual already satisfies an energy inequality of the same type. Comparing this construction with the proposed skew-symmetric and learned dissipative components would clarify whether those components provide an important accuracy or expressivity advantage. The projected baseline fixes the energy and is consequently poorly suited to forcing experiments, while the convolutional baseline lacks the spatial information needed to represent the heterogeneous forcing directly. Although the parameter-count study is useful, these comparisons leave uncertainty about the advantage over simpler, appropriately constrained alternatives. The evidence is also restricted to variants of Lorenz-96.

Questions for authors

Can you establish a sufficient condition under which the Euler–Maruyama sampler in Eq. (C.6) satisfies a discrete Foster–Lyapunov inequality? If such a condition is unavailable, how do stability and response accuracy change with the number of sampling steps, especially under strong forcing? This would clarify how much of the practical stability can be attributed to the architectural guarantee.

How does the proposed architecture compare with a stochastic interpolant whose drift is simply b(x,s;x0) = -gamma*x + f_theta(x,s;x0), with f_theta uniformly bounded? A comparison using matched parameter counts, input features, and training settings would help identify the additional benefit of the energy-preserving and learned dissipative components.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8zUJ

We thank the reviewer for the positive assessment and the precise questions. They led to three new results, summarized below and detailed in the general comment above.

Comparison with a simpler drift (Question 2)

We trained the suggested drift b=−γx+fθb = -\gamma x + f_\theta with fθ=Btanh⁡(gθ/B)f_\theta = B \tanh(g_\theta / B), with the same size, inputs, dissipation rate, sampler and training as the forced-dissipative interpolant, on two-layer Lorenz-96 with two bounds BB (see the general comment).

ModelParametersValidation lossStationary error
Forced-dissipative interpolant488k0.1070.035
Damped drift, B=40B = 40490k0.0774.5
Damped drift, B=1000B = 1000490k0.073143

The damped drift never diverges and has a lower validation loss. However, its long rollouts leave the attractor and converge to a spurious bounded state. The same failure appears on the Marshall–Molteni model (see the general comment).

The damped drift is the special case A=S=0A = S = 0, D=−γID = -\gamma I of our framework, with a forcing that depends on xx. In the forced-dissipative interpolant, only the forcing FF is uniformly bounded. The operators AA, DD and SS are bounded in operator norm, so the energy transfer (A+S)x(A + S)x and the dissipation DxD x are learned, with a fixed structure valid everywhere. In the damped drift, the learned term fθf_\theta is uniformly bounded. It can imitate transfer and dissipation on the data support, but not outside it, where only the fixed damping −γx-\gamma x grows with the state. This leads to unphysical behavior out of distribution. Therefore, dissipation and bounded forcing alone are not sufficient for a physical emulator, and our results suggest that our energy-preserving structure is a crucial physical prior.

We agree that the projected and convolutional baselines could not isolate this effect.

Euler–Maruyama sampler (Question 1)

We now prove a sufficient condition. Let Λ\Lambda and Θ\Theta bound the operator norms of the symmetric and skew-symmetric parts of A+D+SA + D + S. We define the critical number of steps N⋆:=max⁡(8Λ,8Θ2/γ)N_\star := \max(8\Lambda, 8\Theta^2/\gamma). If the number of sampler steps satisfies N≥N⋆N \ge N_\star, the sampled chain satisfies a discrete Foster–Lyapunov inequality with contraction factor e−γ/2e^{-\gamma/2}. The guarantee therefore holds for the deployed sampler.

To measure how much of the practical stability the guarantee explains, we estimated Λ\Lambda and Θ\Theta on the trained models. The two-layer Lorenz-96 model has N⋆=233N_\star = 233, and our experiments use N=1600N = 1600. We then varied NN under the strongest forcing of the paper, 10σ10\sigma, with 4096 ensemble members and the same weights.

NN10501004001600
N≥N⋆N \ge N_\starnononoyesyes
Stable trajectories99.2%99.98%100%100%100%
Mean response error0.170.180.180.180.18
Variance response error0.470.480.490.490.49

Instability occurs only for N<N⋆N < N_\star, where the theory gives no guarantee. The response errors do not appear to depend on NN. Without forcing, the forced-dissipative interpolant never diverged even with 10 steps, while the diffusion model diverged for N≤50N \le 50 (see the general comment).

Beyond Lorenz-96: quasi-geostrophic atmospheric model

We added the quasi-geostrophic model of Marshall and Molteni (1993), a turbulent model of the atmosphere with 513 spectral variables that reproduces atmospheric blocking. The emulators learn the daily dynamics of its 100 leading principal components. We trained forced-dissipative and unconstrained interpolants, and ran each model once for 100,000 days with N=1600≥N⋆N = 1600 \ge N_\star. Results are reported in the general comment.

Our results show that the unconstrained stochastic interpolant is unstable on this system. The damped drift baseline is stable, but converges to an unphysical stationary distribution. In contrast, our forced-dissipative architecture is stable and is also found to accurately represent the response to small perturbations. These results show that our method benefits from a strong physical prior enabling stability, with an explicit condition, and accuracy. It is thus a promising data-driven method to learn challenging large-scale climate dynamics.

The revised paper adds the comparison with the damped drift (Appendix D.4), the study of the number of sampler steps (Appendix D.2) and the Marshall–Molteni experiments (Appendices C.4 and D.3). Section 4 now states the guarantees for the discrete sampler, and Appendix F.4 gives the proof.

AuthorsOpenReview ↗

Follow-up response to Reviewer 8zUJ

Dear Reviewer,

As the discussion period ends on November 18, we would like to know whether our reply and the revised paper answer your questions. Your review led to three results: the comparison with the damped drift baseline, the stability guarantee for the Euler–Maruyama sampler, and an additional experiment on a quasi-geostrophic atmospheric model.

We would be glad to answer any remaining question before the deadline.

Thanks to the authors for analysing the feedback given. I will increase the score!

Reviewer_ErSMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes forced-dissipative stochastic interpolants for probabilistic emulation of partially observed dynamical systems. The method parameterizes the generative drift using bounded forcing, energy-preserving operators, and dissipative components. Under regularity and nondegeneracy assumptions, the authors establish Foster–Lyapunov drift conditions and geometric ergodicity, including stability under steady external forcing and bounded energy under time-varying perturbations.

The method is evaluated on single-layer and two-layer Lorenz–96 systems, comparing long-term stationary statistics and responses to external perturbations against unconstrained stochastic interpolants, projected variants, convolutional models, and conditional diffusion models.

Critical strengths

Well-motivated stability formulation.\textbf{Well-motivated stability formulation.} The paper addresses an important challenge in probabilistic emulation: maintaining stable autoregressive rollouts while reproducing long-term statistics and responses to external perturbations. Evaluating forced responses is particularly relevant to physical modeling.

Explicit theoretical guarantees.\textbf{Explicit theoretical guarantees.} The forced-dissipative parameterization yields a Foster–Lyapunov condition and geometric ergodicity under stated assumptions. The extension to bounded time-varying and steady forcing provides a useful theoretical characterization.

Experiments on stability. \textbf{Experiments on stability. } The authors examine long rollouts, perturbation responses, and model capacity. These experiments provide evidence that the proposed structural constraints can improve robustness over unconstrained generative models in the tested settings.

Critical weaknesses

Insufficient empirical breadth for the physical sciences application track.\textbf{Insufficient empirical breadth for the physical sciences application track.} The experiments are restricted to two configurations of the Lorenz–96 system: a fully observed single-layer setting and a partially observed two-layer setting. Although Lorenz–96 is an established benchmark, both experiments belong to the same synthetic dynamical-system family. No evaluation is provided on another physical system, such as nonlinear PDE dynamics, more complex turbulent flows, or realistic reduced climate models. Consequently, the evidence is insufficient to establish the method's broader applicability to physical sciences, particularly given the paper's motivation in climate emulation.

Limited methodological novelty beyond established stability principles.\textbf{Limited methodological novelty beyond established stability principles.} Probabilistic emulation using conditional generative models is already well established. The main methodological contribution is imposing a forced-dissipative drift structure within stochastic interpolants. However, using bounded forcing and dissipative dynamics to derive Lyapunov stability conditions is a classical approach. While the extension to stochastic interpolant-based autoregressive emulators is technically meaningful, the paper does not sufficiently distinguish its contribution from existing stable dynamical models and structure-preserving generative approaches. Comparisons against simpler dissipative stochastic transition models would help establish whether the proposed architecture offers benefits beyond enforcing a known sufficient condition for stability.

Gap between theoretical guarantees and numerical implementation.\textbf{Gap between theoretical guarantees and numerical implementation.} The geometric ergodicity result requires a nonvanishing noise schedule (Assumption 3), whereas the implemented stochastic interpolant uses (\sigma_s=1-s), which vanishes at the terminal time. The authors argue that Euler–Maruyama discretization introduces a positive noise floor, but this does not directly establish that the continuous-time assumptions hold or that the discretized sampler satisfies the claimed Foster–Lyapunov condition. Since the numerical sampler is what is actually deployed, a discrete-time stability analysis or a systematic discretization-sensitivity study would substantially strengthen the central claim.

Incomplete and partially unmatched baseline comparisons.\textbf{Incomplete and partially unmatched baseline comparisons.} Although several baselines are evaluated, the comparisons do not fully isolate the benefit of the proposed structure. For example, Table 1 reports approximately 488k parameters for the proposed method versus 98k for diffusion in the two-layer forcing experiments. The capacity sweep partly addresses this concern, but the main response comparison still uses differently sized models. More importantly, the paper lacks comparisons against simpler stable probabilistic autoregressive models or alternative dissipative stochastic architectures. Without these, it is difficult to determine whether the observed improvements require the proposed stochastic interpolant formulation.

Limited characterization of probabilistic and physical fidelity.\textbf{Limited characterization of probabilistic and physical fidelity.} Most quantitative evaluations focus on marginal stationary distributions and forced responses of the mean and variance. These are useful but insufficient to characterize the full learned stochastic dynamics. Temporal correlations, joint spatial statistics, spectral properties, and conditional distribution quality are not systematically evaluated. Furthermore, the oscillating-forcing experiment exhibits deteriorating variance accuracy. This raises questions about whether the proposed method achieves faithful probabilistic emulation beyond bounded trajectories and selected low-order statistics.

Questions for authors
  1. Can the authors provide experiments on at least one additional physical dynamical system beyond Lorenz–96? In particular, how does the approach perform on higher-dimensional or spatially structured dynamics, where the assumptions and computational cost of the proposed parameterization may become more restrictive?

  2. What advantage does the proposed stochastic interpolant architecture provide over a simpler probabilistic autoregressive model with bounded forcing, dissipative drift, and stochastic noise? Perhaps a controlled comparison would clarify whether the improvement comes from the stochastic interpolant framework or primarily from the imposed dissipative structure?

  3. Can the authors establish an equivalent Foster–Lyapunov condition for the actual Euler–Maruyama discretization, including the terminally vanishing noise schedule?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ErSM

We thank the reviewer for the detailed review. Additional results can be found in the general comment above.

Additional physical system (Question 1)

We added the three-level quasi-geostrophic model of Marshall and Molteni (1993). It is a large and turbulent model of the atmosphere, with 513 spectral variables against 36 for Lorenz-96. It is much closer to a climate model and produces atmospheric blocking. In rollouts of 100,000 days, the forced-dissipative interpolant stays stable, and its mean responses are as accurate as those of the baselines. The unconstrained interpolant diverges in most runs. The details are in the revised Appendices, and we can move them to the main text if the reviewer finds it useful.

These results show that our method benefits from a strong physical prior enabling stability and accuracy. It is thus a promising data-driven method to learn challenging large-scale climate dynamics. The computational cost is mitigated using a reduced space, and by analyzing and reducing the number of integration steps (see the general comment).

Simpler dissipative stochastic models (Question 2)

We trained an interpolant with a simpler stable drift b=−γx+fθb = -\gamma x + f_\theta, where fθf_\theta is a bounded network, matched in size and training to the forced-dissipative interpolant. It satisfies an energy bound and never diverges. However, despite a lower validation loss, its long rollouts converge to a spurious bounded state, and hence are inaccurate.

The benefit of our architecture therefore does not come from stochastic interpolants, or dissipation. As explained in the general comment, our results show that the advantages of our architecture, including stability and accuracy, come from the strong physical prior imposed by our structure. In contrast, unconstrained stochastic interpolants can produce unstable and inaccurate emulators.

Novelty with respect to classical Lyapunov arguments

We agree that the Foster–Lyapunov framework is classical. To our knowledge, it has not been used to guarantee the stability of a learned generative emulator. The contribution is to build this condition into the drift of a stochastic interpolant, so that it holds for any trained weights. Existing stable models are either deterministic, or constrain the energy to a fixed level through projection or energy-consistent coefficients. A fixed energy level is incompatible with external forcing, as the projected baseline shows.

Assumption 3 and the discrete sampler (Question 3)

We now prove a Foster–Lyapunov condition for the Euler–Maruyama sampler with σs=1−s\sigma_s = 1-s. It holds when the number of steps exceeds a threshold set by the norms of the linear operators. Geometric ergodicity then follows without the noise lower bound of the original Assumption 3. We checked that this assumption is satisfied in our experiments. This new result gives a key practical guarantee for the stability of our forced-dissipative emulator.

Model sizes

The diffusion model of Table 1 is the best one of the capacity study. The diffusion model of 490k parameters, matched in size to the forced-dissipative interpolant, was found to diverge in long rollouts without forcing, although it has the lowest validation loss of all diffusion models. Its responses are therefore not defined. Figure 3 compares the two architectures at matched sizes from 5k to 490k parameters, and the forced-dissipative interpolant is stable at every size. Section 5.2 now states this explicitly.

Statistics beyond marginals

On the two-layer Lorenz-96 system, we compared long rollouts with the truth on the autocorrelation function, the covariance matrix, the correlation against the distance between sites, and the energy spectrum over the Fourier modes of the ring. Below are the average errors. The sampling error is that of a true trajectory of the same length.

ErrorSampling errorForced-dissipativeDiffusionDamped drift
Autocorrelation0.0020.0040.0040.79 to 0.95
Covariance0.0840.1030.09513 or more
Correlation0.0050.0040.0040.15 to 0.27
Spectrum0.0150.0560.0181.0 to 7.7

The forced-dissipative interpolant and the diffusion model are close to the sampling error, showing that they accurately capture the physical correlations. Note that the damped drift baseline fails on every statistic, and that the diffusion emulator is unstable under external perturbations. Therefore, our forced-dissipative model is the only architecture that is stable and accurate for all evaluation metrics.

We agree that the variance response error grows over time. However, no other emulator of our experiments reproduces the variance response consistently, since the baselines either diverge, or produce unphysically large statistics.

The revision adds the Marshall–Molteni experiments, the damped drift, and the new statistics. Section 4 now states the guarantees for the discrete sampler.

Reviewer_ErSMOpenReview ↗

Re: Response to Reviewer ErSM

Thank you for the revision and additional experiments. The new quasi-geostrophic experiment, damped-drift baseline, statistics beyond marginals, and discrete-sampler stability analysis meaningfully strengthen the paper. However, I do not think the revision now fully resolves several central concerns:

  1. Physical fidelity: The additional QG experiment is conducted in a 100-dimensional EOF-reduced space, and its linear response error (0.41) remains substantially above the sampling error (0.05). Moreover, the deterioration in variance-response accuracy under time-varying forcing remains insufficiently characterized.

  2. Methodological novelty: The damped-drift comparison is useful and establishes the importance of the proposed parameterization over simple fixed damping. However, it does not fully clarify the methodological distinction from existing energy-structured and structure-preserving generative models. A more precise discussion of what is fundamentally enabled by the proposed architecture is needed.

  3. Probabilistic fidelity and theoretical guarantees: The additional stationary statistics are valuable but do not fully characterize conditional stochastic dynamics. Furthermore, the numerical verification of the sampler condition relies on operator norms estimated over finitely many visited states, whereas the theoretical guarantee requires global bounds.

I therefore consider these concerns only partially addressed and keep the rating unchanged.

AuthorsOpenReview ↗

Follow-up response to reviewer ErSM

We thank the reviewer for reading the revision carefully and for this follow-up. We answer the three points below.

Physical fidelity

We would like to clarify that the reported number of 0.41 is averaged over lead days 1 to 15. The response decays over this period, so the late leads weigh heavily in the relative error. Over the first days, the relative error of the mean response is much smaller: it is 0.08 for the first day.

Importantly, the unconstrained interpolant reaches the same accuracy when selecting the stable trajectories (before going unstable), so the forced-dissipative structure does not lose accuracy. This indicates that the gap to the truth comes from learning the dynamics in a reduced space of principal components, which is the standard setting for emulating such a model and is equally hard for both architectures. The two architectures differ in stability: the unconstrained interpolant diverges in long rollouts, while the forced-dissipative interpolant remains stable.

We agree that the variance error grows over time under the oscillating forcing. As in the quasi-geostrophic experiment, the emulator sees only part of the state, here the slow variables of the two-layer Lorenz-96 system. The variance response depends on the energy exchanged with the unobserved fast variables, and it is not clear that a reduced-order model of the slow variables alone can track it exactly (Majda and Qi, 2018). Since both architectures reach the same accuracy in the quasi-geostrophic experiment, this error likely comes from the missing information rather than from the forced-dissipative structure. Our method is compatible with the remedies of this literature, such as conditioning the drift on several past states to represent the memory of the fast variables (Falasca and Zanna 2026). We leave this aspect for future work.

Methodological novelty

The architecture enables a probabilistic emulator that is geometrically ergodic for any trained weights, and whose energy responds to a forcing while staying bounded. To the best of our knowledge, our related work section covers the existing structured methods, and our experiments compare against the representative ones. The projected interpolant stands for generative models with a fixed energy, and the damped drift for the simplest dissipative generative model. For completeness, the table below summarizes the differences with the related works cited in the submission.

Model classProbabilisticStationary statistics guaranteedEnergy responds to forcing
Structured deterministic models: port-Hamiltonian, metriplectic, energy-conserving closuresnonoyes
Structured generative models: energy-consistent interpolants, metriplectic flow matching, projectionyesnono, the energy is fixed
Forced-dissipative interpolantyesyesyes, and it stays bounded

Structured deterministic models guarantee bounded trajectories, but this does not ensure that long-run statistics are well-defined. Structured generative models fix the energy, so they cannot respond to a forcing. To our knowledge, the forced-dissipative interpolant is the only stochastic model that has both properties required by forced-response studies. If the reviewer has other works in mind, we would be glad to discuss them.

Conditional dynamics and global bounds

We would like to clarify that two of the new statistics describe the conditional dynamics. The autocorrelation function is a two-time statistic of the stochastic emulator. The forced responses are statistics of ensembles that start from the same states, so they depend directly on the accuracy of the conditional distribution.

The reviewer is right that the theorem requires global bounds Λ\Lambda and Θ\Theta on the operator norms. These bounds can be derived from the network architecture, as follows. The operators AA, DD and AiA_i are linear functions of the output of a neural network whose last hidden layer is RMS-normalized, so this output has a bounded norm for every input. When the weights ϕi\phi_i of the nonlinear term are also bounded, for instance as ϕi=Btanh⁡(gi/B)\phi_i = B \tanh(g_i / B), the triangle inequality gives explicit bounds on Λ\Lambda and Θ\Theta as products of norms of the weight matrices. Assumption 1 then holds, and the critical number of steps N⋆N_\star of Assumption 3 is explicit and can be computed from the weights before sampling. This bound is conservative, and the values measured on the data give only a heuristic estimate of the number of steps.

As a separate remark, the condition on the number of steps can be removed using a linearly implicit Euler sampler, which treats the linear part of the drift implicitly. Each step then solves a linear system, which trades computational cost for unconditional stability: the Foster–Lyapunov condition holds for every number of steps NN, without any bound on Λ\Lambda or Θ\Theta. We can elaborate on this method if judged interesting.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comments: additional results

We thank the reviewers for their careful reading. We address here three points raised by several reviewers, with new results. These changes are included in the revised submission, written in blue.

Comparison with a simpler damped drift

Following Reviewers 8zUJ and ErSM, we trained a stochastic interpolant with the damped drift b(x,s;x0)=−γx+fθ(x,s;x0)b(x, s; x_0) = -\gamma x + f_\theta(x, s; x_0), where fθ=Btanh⁡(gθ/B)f_\theta = B \tanh(g_\theta / B) and gθg_\theta is an MLP. It satisfies an energy bound of the same type as Proposition 3.1. On two-layer Lorenz-96, it shares the size, inputs, dissipation rate, sampler and training of the forced-dissipative interpolant (Appendix D.4).

ModelParametersValidation lossStationary error
Forced-dissipative interpolant488k0.1070.035
Damped drift, B=40B = 40490k0.0774.5
Damped drift, B=1000B = 1000490k0.073143

The damped drift never diverges. However, despite a lower validation loss, its long rollouts leave the attractor and converge to a spurious bounded state.

This can be interpreted with the drift structure. In the forced-dissipative interpolant, only the forcing FF is uniformly bounded. The operators AA, DD and SS are bounded in operator norm, so the energy transfer (A+S)x(A + S)x and the dissipation DxD x are learned and grow with the state. In the damped drift, the learned term fθf_\theta is uniformly bounded. It can imitate transfer and dissipation on the data support, but not outside it, where only the fixed damping −γx-\gamma x grows with the state. This leads to unphysical behavior out of distribution. Therefore, dissipation and bounded forcing alone are not sufficient for a physical emulator, and our results suggest that our energy-preserving structure is crucial.

Stability of the discrete sampler

We now prove a Foster–Lyapunov condition for the Euler–Maruyama sampler used in practice, with NN steps. Write the linear part of the drift as M:=A+D+SM := A + D + S, and let Λ\Lambda and Θ\Theta bound the operator norms of its symmetric and skew-symmetric parts. We call critical number of steps N⋆:=max⁡(8Λ,8Θ2/γ).N_\star := \max\big(8\Lambda, 8\Theta^2 / \gamma\big).

We show the following result. If N≥N⋆N \ge N_\star, then the kernel PNP_N of the sampled emulator satisfies PNV≤aV+cP_N V \le a V + c with V(x)=1+∥x∥2V(x) = 1 + \Vert x \Vert^2 and a≤e−γ/2a \le e^{-\gamma/2} (Proposition 4.1). A sharper version in Appendix E.1, with a larger N⋆N_\star, gives a contraction factor as close as desired to the continuous-time factor e−γe^{-\gamma}. The proofs are in Appendix F.4.

This also removes the need for the noise lower bound of the original Assumption 3.

We checked that this integration stability condition is satisfied for all the experiments involving the forced-dissipative interpolant in the submission. As an ablation, we also varied NN from 10 to 1600 with the same weights. The stationary error of the forced-dissipative interpolant stays at or below 0.05 on both systems, while the diffusion model diverges on Lorenz-96 below 100 steps.

Additional experiment: quasi-geostrophic atmospheric model

We added the three-level quasi-geostrophic model of Marshall and Molteni (1993). It models the Northern Hemisphere atmosphere with realistic orography. It is a large and turbulent system, much closer to a climate model than Lorenz-96. It reproduces realistic climate phenomena such as atmospheric blocking. The emulators learn the daily dynamics of its 100 leading principal components. The details are in Appendices C.4 and D.3, and we can move a summary to the main text if the reviewers find it useful.

The interpolants are trained with several noise levels σ\sigma, and all use 1600 sampler steps. Each model is run once for 100,000 days, the length of the test trajectory. The stationary error is averaged over the 100 components. The table gives its mean over models trained with different noise levels σ\sigma.

ModelParametersStationary error
Forced-dissipative500k0.086
Forced-dissipative50k0.12
Damped drift500k24
Unconstrained500kunstable
Unconstrained50kunstable

The forced-dissipative interpolant stays accurate, while the damped drift often reaches a spurious stationary distribution.

We also measured the mean response to small impulses along each of the 100 components, with 10,000 members. The forced-dissipative interpolant reproduces the linear response with an accuracy close to that of the unconstrained interpolant, and none of its members diverges.

On this larger and more complex system, the forced-dissipative interpolant thus reproduces the stationary statistics and the mean linear response as faithfully as the unconstrained interpolant. It is also the only interpolant that remains stable and accurate at every model size.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.