acceptodds
Under review as a conference paper at ICLR 2027

Dirac-Interconnected Neural Elements: Discovering Modularity in Physical Systems Without Reduction

Abstract

Deep learning has shown remarkable success in the data-driven modeling of dynamical systems. Much of its success is attributed not to the flexibility of neural networks but to inductive biases based on physical prior knowledge, such as energy conservation and symplecticity. However, existing methods do not fully exploit the fact that real-world physical systems are interconnections of components. Some methods require the interconnection to be known a priori, while others assume the system to be reducible to an ordinary differential equation (ODE) and learn only the reduced ODE, discarding the algebraic constraints imposed by the interconnection. Here, we propose Dirac-interconnected neural elements (DINEs), a neural network model that represents a physical system as a differential-algebraic equation (DAE), whose algebraic constraints are given by a Dirac structure in kernel representation. With DINEs, we simultaneously identify from data the interconnection among the components as a Dirac structure and learn the characteristics of the components as neural networks. This allows us to keep the learned subsystems in unreduced form and isolate or compose them to make a new system without retraining. Moreover, DINEs can handle partially observable systems. Experimental results demonstrate these capabilities on physical systems beyond the reach of existing methods.

Primary area:
learning on time series and dynamical systems
Keywords:
deep learning, system identification, Dirac structure, bond graph

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_xfAzOpenReview ↗
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
  1. 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?

  2. 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.

  3. 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

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.

[…] how accurately are the interconnection subspaces and calibrated constitutive laws recovered, […]?

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 10−310^{-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).

How sensitive […] to the readout threshold […]

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).

How sensitive […] to […] modest measurement noise ?

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\lambda for the loss for data assimilation for DINEs to 0.1 x of its value in the noise-free experiments.

  • Presence of coupling: With an appropriate readout threshold, both precision and recall were 1.0 at all tested noise levels
  • Interconnection subspaces and coefficients: Although errors increased with noise, the mean NSE and coefficient RMSE remained at most on the order of 10−310^{-3} at 5% noise.
  • Calibrated constitutive laws: MSE increased with the noise level for many elements.

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).

[…] 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 nfn_f of flow constraints is unknown, the training loss and prediction accuracy for each candidate value can guide its selection. On Task 1, we compared nf=1,…,5n_f=1,\ldots,5. The correct value nf=2n_f=2 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 nfn_f 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 −xNN(x)-xNN(x), where NNNN 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

Reviewer_a1ZbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Dirac-Interconnected Neural Elements, a physical system learning framework based on Dirac structures and differential algebraic equations, which aims to simultaneously identify the constitutive relationships of physical components and their connection structures from observation data without having to reduce the system to ordinary differential equations in advance. The author proposes an accurate parameterization method for separable Dirac structures and combines trajectory neural networks with DAE residual minimization to achieve the modeling of partially observable physical systems. The paper demonstrates the capabilities of constraint recovery, long-term prediction, subsystem combination, and component reuse in five synthetic physical system tasks, and achieves competitive experimental results. Overall, this is a work with clear theoretical motivation and application value, but there are still issues that need to be clarified regarding structural identifiability, the scope of application guaranteed by physical constraints, and the adequacy of experimental comparisons.

Critical strengths

(1) This article has clear and meaningful methodological innovations. Unlike Neural ODEs, Hamiltonian Neural Networks, and traditional port-Hamiltonian models that directly learn dynamic vector fields, DINEs treat the connection relationships of physical systems as learning objects independent of component dynamics. What is particularly commendable is that the paper does not simply add constraint losses to the existing ODE framework, but chooses to retain the algebraic constraints induced by connection relationships at the DAE level.The design of Theorem 1 further enhances the technical value of the method. Through exponential parameterization of orthogonal matrices, the authors transform structural constraints from soft penalties during training into hard constraints satisfied by the model parameterization itself, with clear mathematical explanations.

(2) There is a good correspondence between theoretical design and experimental tasks.This article designed five targeted experiments around the core theoretical advantages of DINEs: Task 1 verifies the learning ability of algebraic constraints, Task 2 examines the modeling ability of some observable systems, Task 3 and Task 4 verify the ability to combine subsystems in the same field and across fields, respectively, and Task 5 demonstrates the isolation and reuse capabilities of learning components. The experimental results echo well with the theoretical motivation. They not only demonstrate the performance advantages of the model in long-term prediction, but also verify the effectiveness of physical structure recovery through the learned connection matrix, giving the core contribution of the paper more systematic experimental support.

Critical weaknesses

(1) Lack of adequate identifiability analysis for accurate recovery of connected structures. There is a lack of adequate identifiability analysis for the precise recovery of connection structures.Theorem 1 proves that for any given 𝑛𝑓 separable Dirac structure of , corresponding parameters exist Ω can represent it. However, this does not directly mean that the jointly optimized model can uniquely identify the true connection relationship from the observation data alone.This issue is particularly important under partially observable conditions. The model simultaneously learns the connection matrix, constitutive relations of each component, and unobservable states, so different parameter combinations may produce the same or approximately the same observable trajectories.The author has discussed gauge freedom in Section 3 and Appendix D.3 and proposed scale calibration methods, but these discussions are not enough to establish general identifiability conclusions.

(2) The expression “guaranteing physical rationality” needs to be more cautious. Proposition 1 establishes the energy balance relationship of the model through Eq. (5), but this result does not mean that the model can unconditionally guarantee energy dissipation. The authors point out in Proposition 1 that the energy dissipation properties can only be further guaranteed if the dissipative elements additionally satisfy the corresponding sign constraints. However, Section 3 and Appendix D.1 clearly state that this article does not impose this constraint during the training process. Therefore, although DINEs can guarantee the power conservation of the connection structure, it cannot guarantee that the learned dissipative components always meet the physical dissipation requirements.

Questions for authors

(1) Given component type, component quantity, and 𝑛𝑓 , what observation conditions can ensure that the connection structure is identifiable in some equivalent sense?

(2) In addition to the non-uniqueness of port variable scaling and matrix row space representation, is it possible to have models with different connectivity relationships but the same observed behavior?

(3) Does the input stimulus need to meet some sufficient stimulus condition to reliably distinguish between different structures?

(4) How robust is the model under noisy conditions?

(5) To what extent does subsystem combination and reuse rely on additional observational information?

4 comments

Thank you for your valuable time and thoughtful feedback. We appreciate your recognition of the clear and meaningful methodological innovations of our work, particularly that DINEs treat the connection relationships of physical systems as learning objects independent of component dynamics and retain the algebraic constraints induced by connection relationships at the DAE level. We also thank you for recognizing the good correspondence between theoretical design and experimental tasks. Below, we address your comments and questions.

[...] this does not directly mean that the jointly optimized model can uniquely identify the true connection relationship from the observation data alone.

Given component type, component quantity, and 𝑛𝑓, what observation conditions can ensure that the connection structure is identifiable in some equivalent sense?

[...] is it possible to have models with different connectivity relationships but the same observed behavior?

Does the input stimulus need to meet some sufficient stimulus condition to reliably distinguish between different structures?

Thank you for your insightful comment. We sincerely acknowledge that our approach cannot guarantee unique identification of the true connection relationship. It is well known that linear systems with different parameter combinations produce the same observable trajectories and so do their combinations. We set some elements to be nonlinear in experiments and identified interconnections, but to the best of our knowledge, there exists no general theorem of the identifiability of nonlinear systems. Hence, this is the common limitation of nonlinear system identification. At least, we can say that our approach finds a single possible candidate system that produces the given trajectories. We have replaced “identify” with “find” in the Abstract and Introduction to avoid further confusion.

The expression “guaranteing physical rationality” needs to be more cautious. [...] although DINEs can guarantee the power conservation of the connection structure, it cannot guarantee that the learned dissipative components always meet the physical dissipation requirements.

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 −xNN(x)-xNN(x), where NNNN is a neural network with nonnegative outputs. This completely eliminated the very small violations of the power-sign condition observed without the constraint, 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 (see Table 12 and Figure 6).

How robust is the model under noisy conditions?

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\lambda for the loss for data assimilation for DINEs to 0.1 x of its value in the noise-free experiments.

Results in Table 13 demonstrate DINEs’ robustness to noise. While its robustness is consistent, a lower weight λ\lambda for the loss for data assimilation is better for a larger noise. Its automatic adjustment and combination with Bayesian methods (e.g., B-PINNs) are left for future work.

To what extent does subsystem combination and reuse rely on additional observational information?

Thank you for your question. Composing subsystems requires fixing the scales and offsets of the port variables being connected, which may require additional observations. In our experiments, none of the subsystems contains transformers or gyrators.

In Tasks 3 and 5, flows (velocities) and efforts (forces) are assumed to be observed, so reducing the interconnection matrices uniquely determines the scales (see Appendix D.3), allowing zero-shot composition. In Task 4, observations of both flows (volume flow) and efforts (pressures) are available for the hydraulic subsystem, so its scale can be fixed. For the mechanical subsystem, observations are available only for flows, leaving the scale and offset of the effort unknown. The ratio of the transformer used for composition is also unknown. Determining these three unknown coefficients therefore requires at least two observations of the effort at the mechanical boundary port and one observation of the composed system. In our experiments, we used four observations of each type to reduce error. These calibration procedures are described in Appendix D.3, and the observations used in the experiments are described in Appendix E.5.

Thank you for your detailed responses and additional experiments. The new results on noise robustness and dissipativity have addressed my previous concerns. Your clarification of the additional observations required for subsystem composition and reuse is also helpful.

Regarding identifiability, I understand that your method aims to find a system consistent with the observed trajectories rather than guarantee unique recovery of the true interconnection. Although this remains a theoretical limitation, I appreciate your clarification and the corresponding revisions to the manuscript.

Overall, your responses and additional experiments have substantially addressed my concerns and strengthened the paper.

Dear Reviewer a1Zb,

Thank you for carefully considering our responses and additional experiments. We appreciate your understanding of our aim to find a system consistent with the observed trajectories.

Your thoughtful comments helped us strengthen the evaluation of noise robustness and dissipativity and clarify the additional observations required for subsystem composition and reuse. We are encouraged that these additions and clarifications have substantially addressed your concerns.

Thank you again for your time and constructive feedback, which have helped us improve our work.

Authors

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

The proposed Dirac-interconnected method for learning modular physical systems jointly learns component constitutive relations and interconnection constraints within a differential-algebraic model. The framework learns the trajectory and dynamics simultaneously with partial observations. Experiments demonstrate constraint identification, subsystem composition, and component reuse on five simulated tasks.

Critical strengths

The idea of separating component laws from interconnection constraints is interesting and offers a useful approach to modular system identification. The proposed parameterization enforces the conditions of separable Dirac structures by construction, which automatically satisfies the hidden mathematical structure behind the data.

Critical weaknesses
  1. The presentation is difficult to understand, the illustration in Fig. 1 may be better with interaction with mathematical formula. In the main text, extensive formalism precedes a concrete explanation of the learning pipeline. The roles of the trajectory network, component networks, and DAE solver could be distinguished more clearly.

  2. The numerical experiments focus on simulated data from known physical models instead of real data. Although the results show the advantages of the proposed method over the classical methods such as Neural ODE, they are limited in practical applications.

Questions for authors
  1. How robust is the method to noisy measurements and misspecified component types or counts?

  2. Can the authors demonstrate the method on measured physical data?

2 comments

Thank you for your time and valuable feedback on our paper. We appreciate your recognition of the value of separating component laws from interconnection constraints. We have addressed your concerns below.

[…] the illustration in Fig. 1 may be better with interaction with mathematical formula […]

Thank you for this suggestion. We have added flow and effort variables at the element and boundary ports in Figure 1 and shown that the Dirac structure Dint\mathcal{D}_{int} connecting the two systems imposes fA+fB=0f_A+f_B=0 and eA−eB=0e_A-e_B=0

[…] formalism precedes a concrete explanation of the learning pipeline. […] The roles of the trajectory network, component networks, and DAE solver could be distinguished more clearly.

Sorry for the lack of clarity.

  • The outputs of the trajectory network represent the states of storage elements (e.g., the momentum of a mass, the displacement of a spring), the flows of resistive elements (e.g., the velocity of a damper), the conjugate variables of source ports (e.g., the current of a voltage source).
  • The component networks (i.e., element networks) define the energy functions of storage elements and the constitutive relations of resistive elements (e.g., the relation between velocity and force in a damper).

We jointly optimize both networks and the interconnection parameters by minimizing the sum of the loss for data assimilation and the DAE residual in Eq. (5). For prediction, we discard the trajectory network and use a DAE solver to simulate the model defined by the learned component networks and interconnection. We have moved the description of the learning pipeline before the mathematical details of the interconnection parameterization in Section 3.

The numerical experiments focus on simulated data [...] instead of real data. [...] they are limited in practical applications.

Can the authors demonstrate the method on measured physical data?

Thank you for your suggestion. Most extensions of HNNs listed in Table 1, such as PoDiNNs, SymODENs, and CHNNs, were evaluated only on simulated data. We consider simulated data more appropriate for a fair comparison. From this perspective, we evaluated the methods on simulated data with noisy measurements. Evaluation on measured physical data is left for future work.

How robust is the method to noisy measurements [...]

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\lambda for the loss for data assimilation for DINEs to 0.1 x of its value in the noise-free experiments.

Results in Table 13 demonstrate DINEs’ robustness to noise.  Its robustness is consistent, and a lower weight λ\lambda for the loss for data assimilation is better for a larger noise. Its automatic adjustment and combination with Bayesian methods (e.g., B-PINNs) are left for future work.

How robust is the method to [...] misspecified component types or counts?

Thank you for your question. We examined the effects of misspecified component 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).

Dear Reviewer XsTe,

Thank you again for your time and thoughtful review. We believe that our responses and additional experiments address your concerns. We would be grateful if you could take a look at our responses and share your thoughts. Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.