acceptodds
Under review as a conference paper at ICLR 2027

Interpretable Test-Time Adaptation of Latent Dynamics from Sparse Sensors

Abstract

Data-driven models of spatiotemporal dynamics are trained on a limited range of physical parameters and routinely deployed on regimes outside it, where a few point sensors may be the only measurements available. Under such parameter shift, the learned latent dynamics no longer describe the system, and the full-state data needed for retraining the model does not exist. We propose TTA-SINDy-SHRED for sparse identification of nonlinear dynamics with shallow recurrent decoder networks (SINDy-SHRED), enabling efficient test-time adaptation to unseen pa- rameter regimes. The joint training objective shapes the SINDy-SHRED latent representation so that a mechanistic governing equation captures the dynamics across regimes, with all regime variation concentrated in the coefficient matrix of the latent space physics model. At test time, all network weights stay fixed and only two context variables are inferred. Adaptation requires no backprop- agation, keeps the latent dynamics interpretable, and updates the decoder from point measurements alone. We evaluate on shock-forming advection–diffusion, reaction–diffusion pattern formation, and decaying two-dimensional turbulence. In every held-out regime, adaptation from a short calibration window reduces long-horizon forecast error relative to the unadapted model. We further test trans- ferability to both neural-operator and attention-based architectures paired with learned latent dynamics, and TTA-SINDy-SHRED attains the lowest error across forecast horizons.

Primary area:
learning on time series and dynamical systems
Keywords:
Sparse sensing, AI for science, Latent space modeling, Test-time learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_XvY8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, authors proposes TTA-SINDy-SHRED for adapting latent dynamical models to unseen physical parameter regimes using sparse sensor observations. The method builds on a pretrained SINDy-SHRED model and constructs affine families for the latent dynamics coefficients and decoder output corrections. At test time, two scalar contexts are inferred from a calibration window without backpropagation through network weights. Experiments on Burgers, Gray–Scott, and two-dimensional Navier–Stokes systems report improved forecasting accuracy over the unadapted pooled model.

Critical strengths
  1. Practical adaptation setting. The method requires only sparse target-regime sensor observations, without access to full-state target data. The Gray–Scott experiment involving changes to the diffusion coefficient of an unobserved species is particularly relevant

  2. Adaptation involves a scalar optimization followed by a closed-form regression. The reported context-inference time of approximately 3.4–4.1 seconds suggests practical potential for real-time deployment.

  3. Promising forecasting improvements. Experiment section reports substantial reductions in decoded error across several regimes. The four-arm ablation also provides useful evidence about the different roles of dynamics adaptation and decoder correction.

Critical weaknesses
  1. The contribution over closely related work is insufficiently established. Low-dimensional context adaptation and affine context models have also been studied in Kirchmeyer et al., 2022 and Blanke and Lelarge, 2024. The sparse-observation setting provides a meaningful distinction, but the current comparisons do not establish the additional value of the proposed combination over appropriately adapted alternatives.

  2. The interpretability claim requires stronger support. Equation (3) specifies ridge regression without an explicit sparsity or support-preservation constraint, making it unclear whether the adapted coefficient matrices retain the sparse structure learned during initial training.

  3. The experimental comparisons do not clearly establish which components drive the reported improvements. The proposed method and the baselines differ in several aspects of their architectures and training procedures, making it difficult to attribute the gains specifically to the proposed test-time adaptation mechanism. More controlled comparisons would help clarify the contribution of this mechanism beyond the underlying model and training choices.

  4. Although the experiments demonstrate promising results under controlled distribution shifts, it remains unclear how well the method transfers to more diverse and challenging conditions. Evaluation across a wider range of scenarios would strengthen the evidence for its applicability beyond the settings considered.

References

  1. Kirchmeyer, M., et al., Generalizing to New Physical Systems via Context-Informed Dynamics Model. International Conference on Machine Learning (ICML), 2022.

  2. Blanke, M., and Lelarge, M. Interpretable Meta-Learning of Physical Systems. International Conference on Learning Representations (ICLR), 2024.

Questions for authors
  1. Could the authors further distinguishes the proposed adaptation mechanism from combining the representation with a CoDA/CAMEL-inspired context adaptation method under the same sparse-observation constraints?

  2. Does the family-fitting procedure preserve a fixed sparse support or apply additional thresholding beyond Equation (3)? Can the authors show representative adapted equations, their sparsity, and the relationship between inferred contexts and physical parameters?

  3. Could the authors further clarify whether calibration-window sensitivity be evaluated with a fixed forecast interval to separate the effect of additional observations from changes in prediction difficulty?

I am willing to raise my score if these concerns are satisfactorily addressed with convincing explanations and evidence.

5 comments

We thank the reviewer for the constructive evaluation and for identifying specific questions whose resolution could raise the assessment. We address each weakness and question below, and upload a revised manuscript with the changes marked in red.

  • W1: Contribution over CoDA/CAMEL insufficiently established.

We have substantially strengthened this positioning in the revision. The new Table 1 (Section 2) places TTA-SINDy-SHRED alongside CoDA, CAMEL, and the sparse-sensing baselines along four capability axes: sparse-sensor input, adaptation to unseen regimes, gradient-free adaptation, and interpretable dynamics. TTA-SINDy-SHRED is the only method combining all four.

The key distinction is not merely that CoDA and CAMEL have not been tested on sparse sensors, but that their context-inference procedures require full-state observations. CoDA optimizes a dd-dimensional context by backpropagating through a neural ODE over complete state trajectories; with only p≪np \ll n sensors, this gradient problem is severely ill-conditioned. CAMEL's closed-form solve takes full-state vectors as input, and its identifiability guarantees assume full observability.

TTA-SINDy-SHRED shares CAMEL's affine context structure—a deliberate design choice that enables the closed-form decoder solve and makes gradient-free dynamics inference from sparse measurements tractable. The affine structure sacrifices the expressiveness of CoDA's nonlinear context for the efficiency and identifiability needed when the observation is a small linear projection of the state. This structural connection is now discussed in Section 2.

  • W2: Interpretability claim requires stronger support.

We appreciate the reviewer for this comment. Sparse variable selection (symbolic regression) is enforced where the equations are learned. Therefore, jointly training with the SINDy loss and thresholding (Section 3.1, Appendix A.2) yields a sparse Ξpooled\Xi_\mathrm{pooled}, leading to an interpretable latent space model. Adaptation then moves the model along a single direction Ξ1\Xi_1 through one scalar, so the adapted dynamics remain an explicit latent ODE at every regime and the parameter dependence is captured entirely by Ξ1\Xi_1. We address the concern about Equation (3) in Q2. To support the claim that the context is physically meaningful, the figure below shows that the inferred context places each unseen Reynolds number on the correct side of the training range and preserves their order across seeds.

Context recovery on Navier–Stokes (LINK): We infer the dynamics context from 32 sensors at the four unseen Reynolds numbers and compare it to the true context, for five seeds. Estimates are clipped at ∣c∣≤3|c| \le 3.

  • W3: Experimental comparisons do not clearly establish which components drive improvements.

We acknowledge the concern and note that the experimental design was structured to isolate components as systematically as possible:

(i) The four-arm ablation (Table 2) holds the architecture fixed (TTA-SINDy-SHRED) and varies which contexts are active: no adaptation, decoder only, dynamics only, and full. This isolates the contributions of cdync_\mathrm{dyn} and cdecc_\mathrm{dec} across three PDE systems and multiple parameter regimes.

(ii) The baseline experiments (Table 4, Figures 3 and 14–17) hold the adaptation protocol fixed (same two-scalar mechanism, same decoder family) and vary the architecture and dynamics model: GRU + SINDy ODE (proposed) vs. FNO / DeepONet / Senseiver + ConditionalMLP. This isolates the effect of the jointly trained SINDy dynamics versus post-hoc neural dynamics.

(iii) The sensitivity analysis (Figure 13) varies one hyperparameter at a time (latent dimension, calibration window, sensor noise, lag, number of sensors) to confirm robustness.

Together, (i) shows that dynamics adaptation accounts for most of the long-horizon improvement, (ii) shows that the SINDy dynamics provide a stronger prior than post-hoc alternatives across architectures, and (iii) shows that results are not fragile to hyperparameter choices.

part 1/2

  • W4: Evaluation across a wider range of scenarios.

The three PDE systems already span structurally distinct forms of parameter dependence: direct dependence through the observed field (Burgers), indirect dependence through an unobserved coupled species (Gray–Scott DuD_u sweep), and nonlinear, non-affine dependence requiring a local approximation (Navier–Stokes).

Beyond these simulated systems, the revision adds a real-world experiment (see the new Section 4.4 and Table 3 of the manuscript) that widens the evaluation in two directions. First, the cylinder wake from RealPDEBench (Hu et al., 2026) uses experimental PIV measurements instead of simulation, where the flow is turbulent, the data is noisy, and no governing equation is available for any regime. Second, the controlled cylinder data varies Reynolds number and forcing frequency jointly, which tests the multi-parameter extension (Equation 11) directly. Test regimes are grouped by whether Re, ff, or both lie outside the source range, and full adaptation gives the lowest decoded error in every group.

Decoded error vs. forecast horizon (LINK): We reconstruct the cylinder (left) and controlled-cylinder (right) wakes from 32 sensors and compare the four arms over a long forecast horizon.

  • Q1: Could the authors further distinguish the proposed adaptation mechanism from combining a SHRED-like representation with a CoDA/CAMEL-inspired context adaptation method under the same sparse-observation constraints?

In principle, one could train a SHRED encoder, extract its latent trajectories, and apply CoDA- or CAMEL-style context adaptation in latent space. However:

(a) CoDA-style (gradient-based, dd-dimensional context): The context would need to be optimized by backpropagating through a neural ODE using the sensor-space residual as the loss. Because the decoder maps dzd_z-dimensional latents to nn-dimensional fields and only p≪np \ll n sensors are observed, the Jacobian of the sensor residual with respect to the dd-dimensional context is rank-deficient, making the optimization ill-conditioned. Our scalar context with Levenberg–Marquardt avoids this by reducing the search to a single well-conditioned dimension.

(b) CAMEL-style (affine, closed-form): CAMEL's closed-form solve requires evaluating the residual at every spatial location. With sparse sensors, the solve must be projected onto the pp-dimensional observation subspace, breaking the identifiability guarantees that underpin CAMEL's theoretical analysis.

(c) Our approach: TTA-SINDy-SHRED instead co-trains the encoder, decoder, and SINDy dynamics so that the latent space is dynamics-aware, then adapts through the sensor residual using a one-dimensional affine family. The joint training ensures that the SINDy library well-approximates the latent dynamics and that the context coordinate varies smoothly with the physical parameter—structural properties that a post-hoc adaptation of a separately trained SHRED encoder would not inherit.

  • Q2: Does the family-fitting procedure preserve a fixed sparse support? Can the authors show representative adapted equations?

The family fitting (Equation 3) uses ridge regression, which does not enforce sparsity directly. However, the support (sparsity pattern) of Ξ0\Xi_0 and Ξ1\Xi_1 is shaped by the SINDy loss during joint training, which promotes sparsity in Ξpooled\Xi_\mathrm{pooled}. Since Ξ0\Xi_0 approximates the source centroid and Ξ1\Xi_1 captures the first-order parameter dependence, both inherit the sparse structure of the trained model. An additional thresholding step could be applied post-hoc to enforce exact support preservation, though we did not find this necessary in our experiments.

  • Q3: Could calibration-window sensitivity be evaluated with a fixed forecast interval?

The sensitivity study in Appendix C already evaluates with a fixed forecast protocol: all configurations are assessed at the same set of horizons H∈{1,25,50,…,550}H \in \{1, 25, 50, \ldots, 550\} regardless of the calibration window length CC. A shorter CC therefore does not reduce the prediction difficulty—it reduces only the amount of calibration data available for context inference while keeping the forecast task identical. Figure 13 (top-center panel) shows that C=25C = 25 and C=50C = 50 yield higher and more variable error at longer horizons, reflecting the difficulty of estimating cdync_\mathrm{dyn} from limited data, while C=100C = 100 and C=200C = 200 produce similar results, suggesting that the context estimate saturates beyond a sufficient calibration length.

part 2/2

I appreciate authors' detailed responses. I have a few follow-up questions:

  • We would like to further ask about the rank-deficient problem. Observations collected over multiple time steps may yield a full-rank context Jacobian, even when observations at each time step are low-dimensional. Instead, reducing the context to a single scalar does not necessarily guarantee identifiability or well-conditioned optimization. Therefore, it remains unclear why the proposed scalar parameterization is theoretically better suited to sparse sensing.

  • Another issue is about the claim of joint training yields a physically meaningful context. From my understanding, the affine family is non-identifiable under reparameterization while joint training alone does not guarantee monotonic or smooth dependence on physical parameters. The distinction between predictive adaptation and physical parameter identification remains unclear as well.

Reviewer_fzXjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This submission addresses the problem of adapting a forecasting neural network for partially observed dynamical systems with varying physical parameters. The authors propose the TTA-SINDy-SHRED neural architecture, building on prior work and incorporating SINDy in latent space to reconstruct and forecast system states from short sequences of sparse observations under physical regimes that differ from the training dataset. This architecture is trained and adapted by sequentially optimizing its components, leveraging the linear structure in the context parameters. The approach is evaluated across several physical systems under varying parameter regimes, outperforming variants where individual components are ablated. Additionally, the authors present sensitivity studies and explore generalization to alternative architectures.

Critical strengths

Designing adaptation strategies to mitigate data scarcity in physical systems is an important and timely problem.

The work is well-positioned relative to recent literature in the deep learning and dynamical systems communities.

Experiments cover a variety of physical systems, including the Navier-Stokes equations.

Critical weaknesses

Overall, the methodology and model development are fairly complex. I have concerns about the problem motivation and the empirical evaluation.

1. Method positioning

1.2 The motivation for the problem setting is limited. The paper lacks a concrete real-world application justifying why full physical states xx would be available during training but only partial observations are available at prediction time. This is a major conceptual gap that the authors do not address.

1.2 The addressed problem is closely related to data assimilation. However, data assimilation is barely discussed in the related work section and is omitted thereafter.

1.3 Interpretability is highlighted as a key property in the introduction, yet it is neither analyzed nor demonstrated in the experimental results.

2. Experimental evaluation

2.1 Following point 1.1, standard data assimilation baselines (e.g., 3D-Var with an approximate model of the dynamics) should be included. Comparing the proposed method almost exclusively against its own ablations weakens the empirical evaluation.

2.2 In Figure 1(a), the wavefront for the Burgers equation barely propagates over time, suggesting the prediction horizon is very short relative to the natural timescale of the dynamics. This raises concerns regarding the evaluation protocol. Furthermore, reporting primarily aggregate error metrics with minimal visualization of actual physical state predictions makes it difficult to assess qualitative performance improvements.

3. Clarity of the method exposition

3.1 The exposition of the method is difficult to follow. The proposed approach relies heavily on the underlying SINDy-SHRED framework, which is described only briefly.

3.2 The methodology is presented as a loose sequence of modeling decisions lacking a clear overarching structure. Key choices, such as opting for a scalar context variable, the specific latent space formulation (1), and the encoder design, lack sufficient justification and explicit underlying assumptions.

3.3 While extending the adaptation mechanism to other architectures is a valuable direction, it remains unclear how the training and adaptation schemes transfer to complex models like Fourier Neural Operators (FNOs), where dependence on the context parameter is inherently nonlinear.

Questions for authors

1. In real-world physical systems where models are imperfect representations of reality, varying dynamics can rarely be reduced to a single, well-identified physical parameter. How should the number of context parameters be determined in practice?

2. What is the primary motivation for making the decoder context-dependent?

3. What motivates using a polynomial candidate library in the latent space, and does this choice yield actionable interpretability?

5 comments

We thank the reviewer for the detailed evaluation and for recognizing the timeliness of the adaptation problem and the breadth of the experimental systems. We address each numbered concern below, and upload a revised manuscript with the changes marked in red.

  • 1.1: Motivation for the problem setting.

The setting—full-state data available during offline training but only sparse sensors at deployment—is the standard operating regime for sensor-network monitoring of physical systems. Ocean monitoring relies on sparse buoy networks, turbulence diagnostics use a handful of pressure taps, and reactor monitoring samples only a small fraction of the spatial domain (Introduction). Full spatial states come from high-fidelity simulations or controlled experiments that are feasible during model development but infeasible to repeat whenever the operating condition changes. The method is designed for exactly this asymmetry: train once on simulation data, deploy with sparse sensors under shifted parameters. Such paradigm is pervasive across physical, biological, and environmental sciences, from infrastructure monitoring to process control to environmental observation networks, making the problem setting practically motivated well beyond.

The new real-world experiment (see the new Section 4.4 and Table 3 of the manuscript) is a direct instance of this setting. RealPDEBench (Hu et al., 2026) was built to evaluate learned models on experimental measurements instead of simulation alone, which reflects that deployment on real measured flows is a recognized practical setting. Full PIV fields of a cylinder wake are used for offline training on the source regimes. At deployment the model receives only 32 point sensors at an unseen Reynolds number, or at an unseen combination of Reynolds number and forcing frequency, with no governing equation available. See our response to Reviewer QBtz (W1) for more details.

  • 1.2: Data assimilation is barely discussed. and 2.1: Standard data assimilation baselines (e.g., 3D-Var) should be included.

We respectfully note that classical data assimilation methods (e.g., 3D-Var, EnKF) address a different task: they assimilate observations into a known forward model to estimate the current state, assuming access to an approximate dynamical model of the target regime. Our method instead adapts a learned reduced-order model to an unseen regime from sparse sensors alone, without access to a target-regime forward model and without retraining. The two settings share the sparse-observation constraint but differ in what is assumed known (forward model vs. pretrained encoder–decoder).

For context-conditioned dynamics-adaptation methods (CoDA, CAMEL), which are structurally closer to our approach, a head-to-head comparison is precluded by the full-state observation requirement discussed in our response to Reviewer QBtz (W2) and formalized in the new Table 1.

  • 1.3: Interpretability is not analyzed or demonstrated.

We demonstrate interpretability in two ways. First, the adapted model is an explicit equation at every regime. Ξ(c)=Ξ0+c Ξ1\Xi(c) = \Xi_0 + c\,\Xi_1 gives a latent ODE whose coefficients are affine in one scalar, and Ξ1\Xi_1 identifies which latent terms the physical parameter acts on and with what sign. Second, the inferred context is physically meaningful. On Navier–Stokes, the context inferred from 32 sensors places every unseen Reynolds number on the correct side of the training range and preserves their order across all five seeds. Though the estimates overshoot the true context in magnitude, they consistently follow the physical parameter, so the single adapted scalar has a direct physical reading. Neither property holds for the neural baselines, where the context enters a ConditionalMLP and its effect on the dynamics cannot be read directly.

Context recovery on Navier–Stokes (LINK): We infer the dynamics context from 32 sensors at the four unseen Reynolds numbers (800, 900, 1900, 2000) and compare it to the true context, for five seeds. Estimates are clipped at ∣c∣≤3|c| \le 3.

part 1/3

  • 2.2: Burgers wavefront barely propagates, suggesting a short prediction horizon.

The Burgers evaluation uses a forecast horizon of H=300H = 300 time steps, over which the shock travels approximately one-third of the spatial domain (Figure 1a). The apparent slow propagation in the figure reflects the long evaluation window, not a short horizon: the time axis runs downward across 300 steps, and the shock visibly sharpens and advances between the low-viscosity (ν=0.02\nu = 0.02) and high-viscosity (ν=0.10\nu = 0.10) regimes. Decoded error curves (Figure 1b) confirm that the error is evaluated continuously up to H=300H = 300, with dynamics adaptation providing sustained improvement over the full range.

  • 3.1: Exposition of the method is difficult to follow.

We appreciate this feedback and have worked to improve clarity in the revision. The method has three distinct stages—(i) pooled training across source regimes (Section 3.1), (ii) affine family fitting for dynamics and decoder (Equations 2–4), and (iii) test-time context inference from sparse sensors (Section 3.2)—and we have tightened the transitions between them.

  • 3.2: Key modeling choices lack justification.

The scalar context is justified by the experimental design: each simulated system varies a single physical parameter, so a one-dimensional family Ξ(c)=Ξ0+cΞ1\Xi(c) = \Xi_0 + c \Xi_1 is the natural first-order approximation (Section 3.1). The multi-parameter extension to kk context dimensions is discussed in Section 4.4 (Equation 11), requires no architectural change, and is validated on the controlled cylinder, where Re and ff vary jointly (Table 3).

The latent-space formulation (Equation 1) factorizes the sparse-sensor prediction into an encoder (sensor →\to latent), a SINDy dynamics model (latent →\to latent), and a decoder (latent →\to full state). This factorization is inherited from SINDy-SHRED (Gao et al., 2026b) and is what makes the dynamics interpretable and the affine family construction possible.

The encoder design (GRU) is a standard choice for processing temporal sensor sequences in the SHRED family (Williams et al., 2024). Our baseline experiments show that the adaptation mechanism works across encoder architectures (GRU, spectral convolution, branch MLP, cross-attention), so the specific encoder is not critical to the method's effectiveness.

  • 3.3: How does adaptation transfer to nonlinear architectures like FNO?

The adaptation mechanism does not require the dynamics to be linear in the context. For the neural baselines, we replace the SINDy ODE with a ConditionalMLP gθ(zt,c)g_\theta(\mathbf z_t, c) (Appendix D.3, Equation 15) where the scalar context cc is concatenated with the latent state. The context enters the dynamics nonlinearly through the MLP's hidden layers, yet the test-time inference still optimizes a single scalar via Levenberg–Marquardt on the sensor residual (Equation 17). The decoder family remains affine and its context is solved in closed form regardless of architecture. Figures 15–17 confirm that full adaptation improves the baselines, with the relative contributions of cdync_\mathrm{dyn} and cdecc_\mathrm{dec} varying by architecture.

part 2/3

  • Q1: How should the number of context parameters be determined in practice?

In the simulated systems studied, a single physical parameter varies at a time, so a one-dimensional context suffices. In practice, the number of context dimensions kk should match the number of independent axes of variation that the practitioner expects at deployment. The multi-parameter extension (Equation 11) is algebraically straightforward; the practical constraint is that Stage 1 inference becomes a kk-dimensional optimization rather than a scalar one, and the source training set must span the multi-parameter space. We validate it on the controlled cylinder (Section 4.4, Table 3), where Reynolds number and forcing frequency vary jointly and full adaptation gives the lowest decoded error in every group.

  • Q2: What motivates making the decoder context-dependent?

The dynamics context cdync_\mathrm{dyn} corrects the latent trajectory, but the decoder was trained on source-regime data and may map the adapted trajectory to a slightly biased reconstruction. The decoder context cdecc_\mathrm{dec} applies a closed-form affine correction to the output layer that compensates for this mismatch. The four-arm ablation (Table 2) demonstrates the complementary roles: decoder-only adaptation helps most at short horizons (before the trajectory drifts), while dynamics-only adaptation dominates at long horizons. Full adaptation combines both benefits.

  • Q3: What motivates the polynomial library in latent space, and does it yield actionable interpretability?

The polynomial library Θ(z)\Theta(\mathbf{z}) is inherited from the SINDy framework (Brunton et al., 2016) and is co-trained with the encoder and decoder so that the latent dynamics are well approximated by a sparse polynomial ODE. This is what distinguishes TTA-SINDy-SHRED from the neural baselines: the baselines learn unconstrained latent spaces where a SINDy library would be unreliable, so they use a neural dynamics model instead (Appendix D.3).

The interpretability is actionable as adaptation changes the coefficients, not the actual network weights. A practitioner can read which latent couplings shift with the parameter and in which direction. The inferred context itself tracks the physical parameter (see our response to 1.3). Moreover, the interpretable polynomial structure serves a dual role: it allows practitioners to identify which latent couplings shift with the parameter, and—because joint training ensures the dynamics are faithfully captured—it enables reliable adaptation over longer rollouts. Our baseline comparison (Table 4) shows that replacing the SINDy library with a more expressive ConditionalMLP could yield less stable long-horizon forecasts despite sharing the same scalar context and adaptation protocol. The interpretability and the adaptation quality are linked through the quality of the learned dynamics.

part 3/3

Reviewer_QBtzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TTA-SINDy-SHRED for parameter shift testing without retrain the network. The method constructs the affine families of latent dynamics and decoder, and use sparse sensor observation to approximate the scalar contexts. Experiments in Burgers, Gray-Scott and Navier-Stokes equations show that this method can decrease the prediction error, where the source of long term improvement comes from the latent dynamics adaptations.

Critical strengths

The method is light weighted. It is able to improve the performance in parameter shift scenarios, without training.

Critical weaknesses
  1. Lack of practical applications. The evaluation is limited to simple 2D PDEs, such as Burgers, Gray-Scott and NS equations. For the parameter-shift setting, the paper only vary one physical parameter at a time. These cases are too ideal and do not support the practical value of this method.
  2. Direct comparisons with existing adaptation methods are missing. The comparisons with FNO, DeepONet and Sensiever mainly shows the methods applicability accorded different model architecture. But the paper does not compare the proposed method with existing dynamics adaptation methods. It remains unclear how much advantage the proposed method offers over existing methods.
Questions for authors

None.

3 comments

We thank the reviewer for recognizing the adaptation mechanism and its effectiveness under parameter shift. We address both critical weaknesses below, and have uploaded a revised manuscript with the changes marked in red.

  • W1: Lack of practical applications.

We agree that a real-world application is important, so the revised manuscript now includes an experiment with real-world data (see the new Section 4.4 and Table 3 of the manuscript). In this added experiment, we use particle image velocimetry (PIV) measurements of the wake behind a circular cylinder in a water tunnel, from RealPDEBench (Hu et al., 2026). This matches the deployment paradigm our framework is designed for: full velocity fields are available offline for a set of source Reynolds numbers, but at deployment the model observes only 32 sparse point sensors in the wake at an unseen Reynolds number, and no governing equation is available for any regime.

This reflects a common practical setting in real-world applications, where dense, full-field data can be collected only for a limited set of conditions (e.g., through laboratory experiments, simulations, or intensive measurement campaigns), while the system in operation is observed only through a few sensors, as in ocean monitoring from buoys, wind-farm flow estimation from met masts, and patient-specific blood-flow modeling from sparse clinical measurements. We further consider a controlled cylinder driven at frequency ff, so that Re\mathrm{Re} and ff vary jointly. Here we extend the context to one coordinate per parameter (Equation 11) with no architectural change. In both settings, adaptation from the sparse sensors alone sustains accurate long-horizon forecasts, whereas the frozen model drifts (Table 3).

Decoded error vs. forecast horizon (LINK): We reconstruct the cylinder (left) and controlled-cylinder (right) wakes from 32 sensors and compare the four arms over a long forecast horizon.

  • W2: Direct comparisons with existing adaptation methods are missing.

We agree that positioning against existing dynamics-adaptation methods is important and have strengthened this in the revision (see updated Section 2 and Table 1 of the manuscript).

CoDA (Kirchmeyer et al., 2022) and CAMEL (Blanke & Lelarge, 2024) are the closest methods to TTA-SINDy-SHRED, but both require full-state trajectories for context inference at test time. Neither method can be applied in our setting without first solving the full-state reconstruction problem they were not designed for. For further details, CoDA applies gradient-based optimization of a dd-dimensional context through a neural ODE, and CAMEL uses a closed-form solve under an affine context structure. CoDA backpropagates through the neural ODE over complete state vectors; with only p≪np \ll n point sensors, this high-dimensional gradient problem becomes severely ill-conditioned. CAMEL's closed-form solve likewise takes full-state observations as input, and its identifiability guarantees (Blanke & Lelarge, 2024, Theorem 1) assume full observability.

Therefore, in TTA-SINDy-SHRED, we're studying a fundamentally different problem setting that only solves adaptation from sparse sensors alone. The updated Table 1 makes this distinction explicit along four capability axes (sparse-sensor input, adaptation to unseen regimes, gradient-free adaptation, interpretable dynamics). TTA-SINDy-SHRED is the only method combining all four.

Back to the reviewer's point, we believe the baseline comparison could be improved indeed. Therefore, our baseline experiments address the complementary question of whether the two-scalar adaptation mechanism transfers across architectures under the same sparse-observation constraints. The four baselines (FNO, FNO-masked, DeepONet, Senseiver) each retain the identical adaptation protocol but replace the encoder and dynamics model. Full adaptation improves all architectures (Appendix D, Figures 14-17), confirming that the mechanism is not architecture-specific. The between-method comparison further shows that TTA-SINDy-SHRED's jointly trained SINDy dynamics provide a stronger prior for long-range latent rollout than the post-hoc ConditionalMLP used by the baselines (Table 4).

Our result is notable because the ConditionalMLP has greater representational capacity than the SINDy family, yet performs worse at longer horizons. The jointly trained SINDy structure constrains the adapted dynamics to a low-dimensional manifold of physically coherent trajectories, an inductive bias whose value grows with the forecast horizon. This finding suggests that for test-time adaptation of dynamical systems, investing in the fidelity of the learned dynamics may matter more than expanding model capacity, a principle that may inform future work in this direction.

Thank you for the additional experiments and clarifications. The real-world results partially address my first concern. However, I still have concerns about the experimental evidence supporting the contribution. The baseline comparison changes the encoder, dynamics model, and training procedure simultaneously, making it difficult to determine whether the improvement comes from the proposed adaptation method. In addition, divergent seeds are excluded from the reported errors, leaving only one seed for the Gray–Scott DuD_u experiments. The new real-world experiments also lack comparisons with alternative adaptation methods. Given the limited methodological novelty and the remaining experimental concerns, I will maintain my score.

We respectfully believe the experimental concern conflates two complementary experiments that together isolate the adaptation contribution.

Adaptation ablation (Table 2). Within TTA-SINDy-SHRED, the four-arm comparison (No-adapt, Decoder only, Dynamics only, Full) holds the architecture, dynamics model, and training procedure identical and varies only which contexts are inferred at test time. This directly isolates the contribution of the proposed adaptation mechanism: full adaptation reduces decoded error over the unadapted model in every system and regime.

Architecture transfer (Table 4, Figure 3). The baseline experiment does the opposite: it holds the two-scalar adaptation mechanism fixed and replaces the encoder and dynamics model across four architectures (FNO, FNO-masked, DeepONet, Senseiver). As stated in Section 4.5, this tests whether the adaptation gains transfer across representation-learning paradigms, not whether a single architecture dominates. All five methods share the same decoder structure and (cdyn,cdec)(c_{\mathrm{dyn}}, c_{\mathrm{dec}}) adaptation; the finding that TTA-SINDy-SHRED outperforms at longer horizons reflects the quality of the jointly trained SINDy dynamics prior, not a confound in the adaptation.

Why alternative adaptation methods cannot be compared under sparse sensing. The reviewer asks for comparisons with other adaptation methods on the real-world data. Existing context-conditioned adaptation methods (CoDA, CAMEL) require full-state trajectories for context inference (Table 1). This is not an incidental implementation detail but a structural requirement: multi-dimensional context inference from partial observations faces an ill-conditioning problem formalized by Stuart (2010) and Ghattas & Willcox (2021), as we discuss in our reply to Reviewer XvY8. Standard test-time adaptation methods (e.g., entropy minimization; Wang et al., 2020; Sun et al., 2020) likewise assume access to the model's full predictive distribution over the observed domain, which is unavailable when the sensor mask covers ~1% of the field. No existing method in Table 1 combines sparse-sensor input, regime adaptation, gradient-free inference, and actionable dynamics; this gap is precisely the contribution.

Seed divergence. The Gray-Scott DuD_u sweep is a hard setting by design: the parameter acts on the unobserved species and reaches the sensors only through the nonlinear coupling UV2UV^2, so rollout divergence is more frequent. Excluding seeds whose rollouts diverge (i.e., produce unbounded latent states) is standard practice in long-horizon evaluations of learned dynamics: the divergence reflects the stability of the trained model at that seed, not a failure of the adaptation, and the exclusion criterion is objective and applied uniformly across all four arms. Supplying the true context in place of the inferred one also diverges at a comparable rate (by H=500H=500, 12 vs. 11 diverged regime-seed pairs for DuD_u), and for the excluded seed even the unadapted model diverges, confirming that the instability originates in the learned dynamics rather than context inference. We acknowledge this leaves a single seed for DuD_u; the adaptation mechanism's effectiveness is corroborated across all other system/sweep combinations with multiple seeds (Burgers; Gray-Scott DvD_v; Navier-Stokes; both cylinder-wake groups). We chose to include the DuD_u sweep precisely because it stress-tests the method on indirect parameter dependence; omitting it would have hidden this boundary case rather than reported it.

References:

Stuart, A.M., 2010. Inverse problems: a Bayesian perspective. Acta numerica, 19, pp.451-559.

Ghattas, O. and Willcox, K., 2021. Learning physics-based models from data: perspectives from inverse problems and model reduction. Acta Numerica, 30, pp.445-554.

Wang, D., Shelhamer, E., Liu, S., Olshausen, B. and Darrell, T., 2020. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726.

Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A. and Hardt, M., 2020, November. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning (pp. 9229-9248). PMLR.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.