- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
DINEs learn component constitutive laws and their interconnection while retaining the algebraic constraints of the resulting physical system. An orthogonal-matrix parameterization represents constant separable Dirac structures, and a trajectory network supports residual-based training with partially observed states. Prediction uses a DAE solver. Experiments examine constraint learning, hidden-state modeling, subsystem composition, and reuse of an isolated spring.
Critical strengths
The mathematical construction directly supports the method: Theorem 1 establishes both validity and coverage of the stated class of Dirac structures. This provides an architectural constraint on power exchange while preserving the component relations needed for composition.
The experiments test capabilities beyond fitting the original system. In particular, the composition and spring-reuse results provide concrete evidence of modularity, with frozen component networks after the specified scale calibration. Evaluation over five seeds and 10,000-step trajectories, compared with 500-step training trajectories, also gives meaningful evidence of temporal generalization.
Critical weaknesses
The evidence for discovering physical structure remains limited to small, noiseless systems with substantial supplied information: element counts, types, observation correspondence, selected linear constitutive forms, and the constraint count. Appendix F provides useful interconnection readouts, but not systematic recovery statistics or sensitivity to noise and the threshold used to interpret coefficients. Theorem 1 establishes representational coverage, not identifiability from partial observations; likewise, failure under an incorrect constraint count in Task 1 does not establish that misspecification will always be detectable.
The large accuracy gaps against the ODE baselines do not fully isolate the contribution of the proposed representation. As Table 7 acknowledges, these baselines receive less component-level prior information; on Tasks 1, 3, and 4, they also use one-step training before evaluation over 10,000 steps. The soft-constraint ablation supports the reliability of the exact parameterization, but its frequent solver failures leave limited evidence for predictive accuracy against a viable alternative with comparable physical assumptions.
Questions for authors
-
Across the five seeds, how accurately are the interconnection subspaces and calibrated constitutive laws recovered, rather than only the observed trajectories? How sensitive are these results to the readout threshold and modest measurement noise?
-
How sensitive are the accuracy gaps on Tasks 1, 3, and 4 to multi-step training of the ODE baselines? Please clarify how much of the advantage can be attributed to the additional component-level priors summarized in Table 7.
-
Since the implemented resistive networks do not enforce dissipativity, do the learned elements satisfy the required power-sign condition over the states visited during composition and reuse? Please distinguish the guaranteed interconnection energy balance from properties of the learned elements.
4 comments
[…] substantial supplied information: element counts, types, observation correspondence, selected linear constitutive forms, […]
Thank you for your comment. We would like to emphasize that it is an advantage of DINEs that they can leverage known information of elements to learn unknown constitutive laws and interconnections.
Linearity of constitutive laws
When the form of a constitutive law is known, we can leverage this knowledge to limit what needs to be learned. For example, momentum is a linear function of velocity, with mass as the coefficient, so only the unknown mass needs to be learned. In the experiments in the main text, we assumed linearity for linear elements. To examine the effect of removing this assumption, we trained the model without assuming linearity for the C-type storage element in Task 4. The learned constitutive law closely matched the true linear relation, and the predictions closely matched the ground truth. These results show that constitutive laws can be learned even when the linearity assumption is relaxed. We have added these results to Appendix F.6 (Figure 8).
Element types and counts
We examined the effects of misspecified element types or counts on mechanical Subsystem A in Task 3, while keeping the observed variables and the number of flow constraints unchanged.
- Reducing the number of resistive elements from two to one, both the training loss and prediction error increased catastrophically compared with the correctly specified model.
- Increasing the number of resistive elements from two to three, the solver failed during initialization or prediction because of the shortage of constraints
- Replacing one resistive element with a C-type storage element, the training loss decreased during training, but the prediction error was much larger than that of the correctly specified model.
These results show that training and prediction can help detect misspecified element types or counts, providing insight into the structure of the target system. We have added these results to Appendix F.6 (Figure 7).
Theorem 1 establishes representational coverage, not identifiability from partial observations; […]
Thank you for your insightful comment. As you point out, Theorem 1 establishes representational coverage, not identifiability from partial observations. The value of our approach is that it finds a system that produces the given trajectories. We have replaced “identify” with “find” in the Abstract and Introduction to avoid further confusion.
[…] failure under an incorrect constraint count in Task 1 does not establish that misspecification will always be detectable.
Thank you for your insightful comment. When the number of flow constraints is unknown, the training loss and prediction accuracy for each candidate value can guide its selection. On Task 1, we compared . The correct value yielded the lowest final training loss, while the other values led to solver failure or poor prediction (see Figure 5 and Table 2(a)). We have removed or revised statements that misspecifying necessarily leads to failure and restricted our conclusions about its effects to the settings examined (Sections 3 and 4.2 and Appendix D.4).
How sensitive are the accuracy gaps on Tasks 1, 3, and 4 to multi-step training of the ODE baselines?
Thank you for your question. On Tasks 1, 3, and 4, we compared prediction performance by increasing the number of integration steps during training from 1 to 10 and 50. Although some metrics improved for the baselines, DINEs achieved lower mean MSE and higher mean VPT over five seeds in every setting. These results confirm that DINEs retain their advantage in prediction accuracy even when the ODE baselines use multi-step training. We have added these results to Section 4.2 and Table 2(a).
The large accuracy gaps against the ODE baselines do not fully isolate the contribution of the proposed representation. As Table 7 acknowledges, these baselines receive less component-level prior information; […]
[…] Please clarify how much of the advantage can be attributed to the additional component-level priors summarized in Table 7.
Thank you for your comment. The ability to incorporate physical prior knowledge is itself an advantage of DINEs, so we cannot separate the contribution of DINEs from that of this knowledge. Without sufficient prior knowledge, only general methods such as Neural ODEs can be used, and we acknowledge that DINEs would not be useful in that case. DINEs are more flexible than PoDiNNs, which impose the overly restrictive requirement of reducibility to an ODE, and pH-DAE identification, which assumes known interconnection coefficients.
The soft-constraint ablation [...] frequent solver failures leave limited evidence for predictive accuracy against a viable alternative with comparable physical assumptions.
No alternative method uses physical assumptions comparable to those of DINEs. The existing methods compared in Table 1 cannot both learn the interconnection and retain its algebraic constraints in a DAE, and they rely on stronger assumptions. We therefore compared DINEs with DINEs (soft) under comparable constraints. This variant uses the same element models and observations, but replaces the parameterization that exactly satisfies the conditions of Dirac structures with penalties. On Tasks 1 and 2, DINEs (soft) failed during initial-state estimation or prediction for many seeds, whereas DINEs completed prediction for all seeds. These results show that DINEs are superior under the same element models and observations.
[…] do the learned elements satisfy the required power-sign condition over the states visited during composition and reuse? […]
Thank you for your question. We evaluated the learned resistive elements at the states visited during prediction of the composed systems in Tasks 3 and 4. The positive power corresponding to violations of the dissipation condition was less than 0.42% of the RMS of the power of the ground truth. This shows that the magnitude of the violations was sufficiently small. In Task 5, we reused the learned spring without transferring any resistive elements. The dampers in the receiving system use known positive damping coefficients.
Since the implemented resistive networks do not enforce dissipativity, […] Please distinguish the guaranteed interconnection energy balance from properties of the learned elements.
Proposition 1 guarantees energy balance based on power conservation of the interconnection. Dissipativity of a resistive element can be guaranteed by incorporating the power-sign condition into its constitutive relation. Since high prediction accuracy was achieved without this guarantee, we did not include the constraint in the initial manuscript. Thank you for this important comment.
To demonstrate that dissipativity can be guaranteed, we introduced a constraint on the resistive relation of Subsystem B in Task 3 so that the product of flow and effort is nonpositive for every input. Specifically, we define the resistive relation as , where is a neural network with nonnegative outputs. This eliminated violations of the power-sign condition, and both the constitutive relation and the predictions after composition with Subsystem A closely matched the ground truth. These results demonstrate that accurate prediction after composition is possible while guaranteeing dissipativity of the resistive element. We have added these results to Appendix F.4 (Table 12 and Figure 6).
Dear Reviewer xfAz,
Thank you again for your thoughtful review and for recognizing the value of our work. Your comments helped us strengthen the paper through additional experiments and clarifications. We would be grateful if you could take a look at our responses and share your thoughts.
Best regards,
Authors
We sincerely thank you for your time and constructive feedback. We appreciate your assessment that the mathematical construction directly supports the method and that the composition and spring-reuse results provide concrete evidence of modularity. We also thank you for recognizing the meaningful evidence of temporal generalization. Below, we address your comments and explain the revisions prompted by your suggestions.
Thank you for your question.
We evaluated the recovery of interconnection subspaces (i.e., the subspace defined by the Dirac structure) using the normalized subspace error (NSE). A zero NSE indicates that the learned and ground-truth subspaces coincide.
We also evaluated the interconnection coefficients using the root mean squared difference from the ground-truth coefficients (coefficient RMSE).
For the calibrated constitutive laws, we computed the mean squared error (MSE) between their outputs and the ground-truth outputs for the same inputs.
For all evaluated systems in Tasks 1–4, the mean NSE and coefficient RMSE over five seeds were at most on the order of 10−3, indicating that the interconnection subspaces were recovered with high accuracy. The calibrated constitutive laws also generally matched the ground truth over the evaluated input ranges. We have added these results to Table 3 and Appendices F.2 and F.3 (Tables 10 and 11).
Thank you for your question. To identify the presence of coupling as nonzero entries, we applied a threshold to the absolute values of the interconnection coefficients after readout. Precision is the fraction of detected couplings that are present in the ground truth, and recall is the fraction of ground-truth couplings that are detected. We achieved both precision and recall of 1.0 over a wide range of thresholds, although extremely low thresholds led to false positives (see Table 2(b) and Table 10).
Thank you for your question. For Tasks 1 and 2, we trained the models with noisy observations, where we added 1%, 5%, or 10% Gaussian noise and evaluated predictions against the noise-free ground truth.
To prevent overfitting, we reduced the weight λ for the loss for data assimilation for DINEs to 0.1 x of its value in the noise-free experiments.
Under the tested conditions, we confirmed that interconnection recovery was robust to observation noise up to at least 5%. If greater robustness to observation noise is required, we will consider applying Bayesian methods (e.g., B-PINNs). We have added these results to Appendix F.5 (see Tables 14–16).