acceptodds
Under review as a conference paper at ICLR 2027

Structural Uncertainty in Physics-Informed Inverse Problems: The Hidden Risk in Scientific AI

12 views

Abstract

Machine learning is increasingly used to infer physical laws from data, yet a model that reproduces observations accurately can still encode a physically non-identifiable coefficient function. This danger is particularly acute in physics-informed machine learning (PIML), where coefficient functions of governing equations—such as Hamiltonians or Lagrangians—are learned by minimizing physics-based predictive losses. We show that this failure can originate in the governing equations themselves: multiple, physically distinct coefficient functions can satisfy the same differential equations on the same observations, a situation that additional data cannot resolve when the non-uniqueness is structural. We introduce a rank-based criterion that diagnoses this structural uncertainty before training and evaluates whether candidate physical constraints restore identifiability. Specifically, we combine the numerical null space of the equation matrix with the DNN tangent space to select the strength of a physical constraint without ground truth, based on the reduction of structural uncertainty. Experiments on Hamiltonian, Lagrangian, diffusion, and wave-kinetic systems demonstrate that predictive accuracy, structural identifiability, and physical-model recovery are distinct quantities.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, physics-informed machine learning, structural uncertainty, structural identifiability, inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ch7AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper examined an interesting question of how to eliminate structural uncertainty when considering PDEs: such issues are problematic as it's unrelated from amount of data or model decisions, and would have to be factored out via training procedures. The authors showed a procedure for linear problems where they were able to evaluate the amount of uncertainty in a given problem, decide on constraints to eliminate them, then train; they then demonstrated on two problems in the main text with some extensions in the appendices.

Critical strengths
  1. Rigorous approach to the problem of how to design a loss function to remove ambiguity from the data. I've never thought too much of this issue in the past, but I think the paper has convinced me that this might be an issue in the future, and might be something to use with different structural preserving NNs. Their approach seems to attack the correct part of the inverse problem which is not discussed in the field.
Critical weaknesses
  1. Originality and scope: while I'm not an expert in inverse problems, I'm very surprised that the field doesn't have similar results proven already as the result is straightforward linear algebra? One would expect that classical uniqueness results would also examine sampled points in the interior too, as those are very relevant scientific problems. I think the main contribution would be the recipe + tangent stuff which brings me to...

  2. Clarity in writing: I had a hard time reading this. For example, Thm 1 is far too long, and one should place the definitions outside the theorem block with potential examples to guide the reader. I do not understand lines 324 to Section 5; which of the step is it trying to solve? These paragraphs seems to me to be the most exciting part of the paper, but it's much too compressed and needs to be expanded drastically. Furthermore, too much details are swept to the appendix that it forces the reader to read it for the experiments.

Questions for authors
  1. See point 1 of weakness regarding originality and scope. Perhaps a wider lit survey is needed.

  2. One major question: is this a bad thing? Using the helio/geocentric example at the beginning, it seems that if we were just trying to predict planetary movement that as long as the ML system were to predict correctly, it really doesn't matter which reference frame was used.

  3. How would this work with something like SINDy or general equation discovery in analytical forms? Wouldn't that be a major win to be able to, say given the planetary data, give you the classical Kepler's law?

  4. I understand for sake of analysis that the limitation to nonlinear dependence makes sense, but isn't this fairly non-restrictive? In PDE land, I feel that most equations are linear in the coefficient feature vectors? Can you exposit more on this?

1 comment
AuthorsOpenReview ↗

Response to Reviewer ch7A

Thank you very much for your encouraging review. As explained in our general response, I apologize for the delay in our response. We are glad that the paper conveyed why this issue may matter for structure-preserving networks. We apologize that our literature survey was insufficient and that parts of the paper were hard to read. Your suggestion to widen the literature survey helped us state the novelty of our work more concretely, and your question on whether structural non-uniqueness is actually harmful led us to add experiments that extract physical quantities from the learned models, which show the effectiveness of the framework much more directly. We are sincerely grateful for both.

Originality (W1, Q0). You are right that the result is straightforward linear algebra, and we apologize that the shortened manuscript overstated the novelty of the theorem. We now relate Theorem 1 explicitly to classical identifiability analysis and to uniqueness results for interior data and for learning PDEs, notably He, Zhao & Zhong (2024) and Scholl et al. Our contribution is the framework: the position that structural non-uniqueness must be evaluated before a coefficient function learned with PIML is interpreted as physics, and a routine procedure for doing so across a broad class of PDEs. The procedure is deliberately elementary so that physical scientists who are not specialists in inverse problems can apply it as a standard check. Revision: Abstract; Sections 1 and 2.3; Section 3 ("Relation to classical results").

Clarity (W2). Thank you for the concrete suggestions. We now write the class of PDEs explicitly (Eq. (3)) with examples (Hamilton's equations, the wave-kinetic equation, the diffusion equation). We moved the definitions out of Theorem 1, which now states only the equivalence, and added a paragraph explaining with the wave-kinetic example why compatibility matters. Lines 324–377 are Step 3, which gives a hint for the weight λ of the soft constraint of Step 2. We now open Step 3 with its idea in plain terms: the structurally undetermined directions are those along which the physics loss is flat; adding the constraint should remove this flatness; and we take as a hint the λ beyond which increasing it brings diminishing reduction. Concretely:

  1. from M we obtain the directions the equation cannot see;
  2. from the trained network we obtain the directions in which the network can move;
  3. their overlap, measured by principal angles, is the ambiguity the network can realize;
  4. with a local Gaussian approximation, we track how the variance along high-overlap directions shrinks relative to the others as λ grows.

The main comparison of rules for choosing λ is now in the main text (Fig. 5). Revision: Section 3; Section 4 (Step 3); Fig. 5.

Q1. Is this a bad thing? For predicting the observed quantity it is harmless, which is exactly why the predictive loss cannot detect it. It matters once the learned function is used as physics. In the pendulum example, the unconstrained model reproduces the training relation with small error, yet its effective mass ∂²L/∂θ̇² is negative in 80% of held-out states and the equation of motion derived from it is wrong (median error 5.3 against accelerations of order one). With the constraint, the mass is recovered as 1 and the equation of motion to within 0.01. In the wave-kinetic case, ∂H/∂k_x is a group velocity: H and H + F(I) give the same evolution of I but different velocities. Revision: Section 5.1; Table 10.

Q2. SINDy and equation discovery. The analysis applies upstream of equation discovery. If the library columns are linearly dependent on the data, sparse regression can return any member of the equivalent family, and the null-space analysis shows which constraint or observation resolves it. Once the function is identifiable, symbolic regression can be applied to seek a closed form, in the spirit of your Kepler example. Revision: Section 6.

Q3. Linearity. We agree. The equation needs to be linear only in the values and derivatives of the unknown coefficient function; dependence on the state may be arbitrarily nonlinear. We now write this class explicitly as Eq. (3). It covers Hamiltonian and Lagrangian mechanics, Liouville- and wave-kinetic-type transport equations, Burgers-type equations, and diffusion and advection–diffusion equations with unknown coefficients. Revision: Section 1; Section 3 (Eq. (3)).

Reviewer_cmkLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies structural ambiguity in learning unknown physical functions, emphasizing that agreement with governing equations does not necessarily imply unique recovery. It develops identifiability diagnostics using a consistency map, the null space of a physics operator, and the tangent space induced by a neural network. It further proposes a geometry-based uncertainty ratio for selecting the strength of additional constraints and demonstrates the framework on several physical systems.

Critical strengths

The paper highlights a relevant distinction between fitting observed dynamics and uniquely recovering the underlying physical functions. It organizes equation-induced ambiguity and neural-network representability within a common framework. The numerical examples illustrate how low equation residuals can coexist with poor function recovery and how additional constraints can improve recovery.

Critical weaknesses

The rank conditions and local tangent-space analysis closely correspond to classical identification-matrix and sensitivity-based methods reviewed by Miao et al. (2011), making the theoretical advance beyond these methods unclear. Related formulations already exist in PDE identification from solution data (He, Zhao, and Zhong, 2024) and tangent-space-based neural-network identifiability (Bona-Pellissier, Malgouyres, and Bachoc, 2022), so the differences in setting and assumptions need to be accompanied by a concrete account of the new technical contribution. The main methodological novelty therefore appears to lie in constraint-weight selection, but its effectiveness is insufficiently established because the experiments do not adequately separate the benefit of adding constraints from that of the selection rule, or compare against fixed weights, validation using the same constraint information, and established weight-selection methods.

Miao, H., Xia, X., Perelson, A. S., and Wu, H. (2011). “On Identifiability of Nonlinear ODE Models and Applications in Viral Dynamics.” SIAM Review, 53(1), 3–39.

He, Y., Zhao, H., and Zhong, Y. (2024). “How Much Can One Learn a Partial Differential Equation from Its Solution?” Foundations of Computational Mathematics, 24, 1595–1641.

Bona-Pellissier, J., Malgouyres, F., and Bachoc, F. (2022). “Local Identifiability of Deep ReLU Neural Networks: the Theory.” Advances in Neural Information Processing Systems, 35, 27549–27562.

Questions for authors
  1. What is the specific technical novelty beyond existing identifiability analyses? Please clarify the relationship to identification-matrix and sensitivity-based methods reviewed by Miao et al. (2011), PDE identification studied by He, Zhao, and Zhong (2024), and tangent-space-based neural-network identifiability studied by Bona-Pellissier et al. (2022). In particular, what does the consistency map TT, together with the separate analysis of the physics null space and neural-network tangent space, enable that existing formulations do not?

  2. How can the benefit of adding constraints be separated from the benefit of the proposed weight-selection rule? In both reported weight-selection examples, a fixed choice of λ=1\lambda=1 yields better recovery metrics than the selected value. Please compare the methods using the same candidate constraints across multiple conditions and establish when the proposed rule improves upon simple fixed weights and whether those improvements are reproducible.

  3. How reliable is the relationship between the selection indicator and recovery accuracy, and what is its computational cost? Please explain why the uncertainty-ratio elbow should correspond to a favorable recovery region and assess whether this correspondence persists under changes in noise, data size, network architecture, and numerical-rank thresholds. Please also report the additional cost of Jacobian construction and matrix decompositions and assess whether improvements over simpler selection methods justify that cost.

  4. What practical scientific decisions does the proposed framework improve, and what prior knowledge does it require? The workflow appears to require a specified governing-equation structure, access to the necessary states and derivatives, and scientifically justified candidate constraints, placing substantial demands on the user's understanding of the system. Please provide a realistic use case demonstrating what a domain expert can determine or accomplish with this framework beyond existing identifiability analysis or directly incorporating known constraints, and clarify whether this benefit persists with noisy or partially observed data.

6 comments
AuthorsOpenReview ↗

Response to Reviewer cmkL (1/2)

Thank you very much for your detailed review and for the three references. As explained in our general response, I apologize for the delay in our response. We also apologize for the insufficient literature survey in the submitted version. Working through the references you provided made the novelty of our work considerably more concrete, and the comparisons and robustness tests you suggested allowed us to show the effectiveness of the framework much more clearly than in the submitted version. We are sincerely grateful for both.

Q1. Novelty. We agree that, once a finite representation is fixed, Theorem 1 is an elementary rank condition of the classical type; we now say so explicitly and relate it to the identification-matrix and sensitivity literature, including Miao et al. (2011). As explained in our general response, our contribution is the framework: the position that structural non-uniqueness must be evaluated before a learned coefficient function is interpreted, and a routine procedure for doing so across the class of PDEs in Eq. (3). Comparing with your references clarified what the procedure needs in order to apply to PIML.

  • Miao et al. (2011) address finite parameter vectors in ODE models; we address coefficient functions of PDEs.
  • He, Zhao & Zhong (2024) is the closest work. In their analysis of variable coefficients, the coefficient functions are independent unknowns, which corresponds to T = I in Theorem 1. In PIML, the features are derivatives of one function, and this changes the answer. In the wave-kinetic equation, the pointwise null space g(x,k)∇I reduces to the family H + F(I) once the features are required to come from one Hamiltonian. On our data, the pointwise null space on a 50 × 50 subgrid has dimension 2500, whereas 54 ± 4 near-null directions remain with the compatibility map, and only the additive constant remains under the reflection constraint.
  • Bona-Pellissier et al. (2022) concern identifiability of network parameters; we use the network tangent only to find which coefficient-function directions the network can realize and the equation cannot see.

Revision: Section 3 ("Relation to classical results", "Why compatibility matters"); Appendix G; Appendix L.5.

Q2. Constraint versus selection rule. Thank you for this important point. We ran the comparisons you suggested with our original code (Fig. 5; Tables 8–9). Fixed-weight rules use the raw residuals, because normalization factors chosen by hand amount to tuning λ in advance.

  • Validation on the physics residual alone is unreliable in both systems (cosine 0.50 and 0.72).
  • In the wave-kinetic case, gradient-norm balancing (0.64), a fixed weight λ = 1 (−0.33; trained additionally) and fixed-weight validation including the constraint residual (0.68) also fail; with hand-chosen rescaling of the residuals, the fixed-weight rules would give about 0.87.
  • The NT rule selected λ = 10^-3 in every wave-kinetic slice (0.845) and never selected the collapse region; it is invariant to residual rescaling, because rescaling only shifts R_NT(λ) along the λ axis.
  • In the Lagrangian case, whose objective has no normalization factors, the fixed-weight rules and gradient-norm balancing were more accurate (0.9997 against 0.992).

Across all conditions, the NT rule was the only rule that did not fail (lowest cosine 0.845, against 0.68 or below for every other rule in at least one case). Revision: Section 4 (Step 3); Section 5.2; Fig. 5; Appendix L.1.

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer cmkL (2/2)

Q3. Reliability and cost. Across all combinations of evaluation grid, tangent threshold and eigenvalue floor, the mean cosine at the selected weight stays within 0.992–0.9997 (Lagrangian) and 0.80–0.87 (wave-kinetic), and changing the numerical-rank threshold of M (10^-10 to 10^-6) leaves the selection unchanged (Appendix L.3). With noise σ = 0.1, the rule gives cosine 0.98. Using 1/2 or 1/5 of the data on the same trajectories left the selected weight (10^-3) and the recovery (0.99) unchanged, and so did a wider network (width 128; 0.997) and, in the wave-kinetic case, a 2-50-10-1 network (0.858, against a post-hoc optimum of 0.873) (Table 12). With a narrow network (width 32), the selected weight lies near the onset of the high-recovery regime, where recovery is lower (0.89). The elbow marks the onset of the reduction in structural uncertainty, not an optimal weight. As for cost, R_NT(λ) for every λ is obtained at one unconstrained solution from one Jacobian and two SVDs, so no model has to be trained per candidate λ. Revision: Section 6; Appendix L.2–L.4 and L.6 (Tables 10–12).

Q4. Practical use and prior knowledge. The framework assumes the form of the governing equation and candidate constraints from domain knowledge; it does not discover constraints. Within that setting, in the wave-kinetic case it tells the physicist before training three things: H is undetermined up to F(I); the reflection symmetry removes this ambiguity under a determinant condition verified on the data (all 49,000 blocks); and recovery will be ill-conditioned in certain regions. The pendulum example shows the error that such a check prevents: without the constraint, a model with small prediction error has a negative effective mass in most states. The practical situations that motivate the work are described in our general response. Noisy and partially observed data are an important next step. Revision: Section 5.1; Table 10.

Thank you for the response and the additional experiments. The response clarifies that Theorem 1 is a classical rank condition and explains the role of compatibility through examples. Computing the selection criterion from one unconstrained solution, without retraining for each weight, is also an advantage. I still have three concerns.

  1. Checking identifiability before interpreting a physical function is already standard practice in identifiability analysis. What is new about the procedure for constructing the compatibility map, and what can the resulting analysis tell us that existing methods cannot? Were the 2,500 null-space dimensions and the 54±454\pm4 near-null directions counted using the same criterion? Please distinguish exact non-identifiability from threshold-dependent ill-conditioning, and clarify whether the claim that reflection leaves only an additive constant is proved or checked numerically.

  2. I agree that normalization changes the effective weight, but normalization chosen without the ground truth is still a reasonable baseline. The reported results, about 0.87 for a rescaled fixed weight and 0.845 for the proposed method, suggest that the rule avoids failures caused by residual scaling; they do not show that it works better than a simple method with appropriate normalization. A comparison using normalization fixed in advance, without access to the ground truth, is needed. Also, why did the selected weight in the wave-kinetic example change from λ=10\lambda=10 in the submission to 10−310^{-3} in the response? Please explain any changes to the objective or normalization and how the two weights correspond.

  3. The negative-effective-mass example shows why prediction error alone is not enough to assess a learned physical function. Howver why is the proposed procedure preferable to imposing known constraints directly or checking the relevant physical quantities after fitting? For computational cost, please report runtime and memory usage, rather than just the number of Jacobians and SVDs. The noise and subsampling tests are useful, but they do not address missing states or derivatives. Please state what must be observed and where the method no longer applies.

The new experiments answer some of my questions, but I am not yet convinced that the paper makes a sufficient technical contribution beyond classical identifiability analysis.

Reviewer_Af16OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper argues that accurate predictions do not imply recovery of ground-truth physical laws. One reason for this is that the same law can be formulated in different but equivalent ways. The main theoretical contribution is a discrete identifiability criterion, which checks whether equation-preserving perturbations can actually arise from an admissible coefficient function, rather than treating its values and derivatives as independent unknowns (Theorem 1). The proposed workflow diagnoses these ambiguities, introduces physical constraints to remove them, and selects their penalty strength using the overlap between the equation’s null space and the neural network’s local tangent space.

While this is an important question to consider, it is not a new one. There is a vast literature on structural identifiabllty of dynamical systems. There is also related literature of "sloppy models" that considers parameter values in a learned model that can vary over orders of magnitude while having no noticable impact on the underlying equations. The authors do not do a good job situating their work relative this - the PIML aspect is somewhat novel but, technically doesn't add much novelty.

The main rank criteria the authors derive is actually somewhat weaker than what the abstract promises. Theorem 1 establishes an exact identifiability criterion, but does not justify the weighted uncertainty ratio in Eq. (5) or its elbow-based selection rule. The proxy in (34) is evaluated at the unconstrained solution for all constraint strengths, and no theoretical connection to coefficient recovery is established. These should be presented as heuristic diagnostics, with sensitivity analyses supporting their reliability.

I think there's also a technical error. After (5) teh authors state that "in the exact noiseless limit, γi∈0,1\gamma_i\in {0,1}...". Consider M=[0 1]M= [0 ~1], PN=diag(1,0)P_N = \mathbf{diag}(1,0), and $Q_T= \frac{1}{\sqrt{2}}\mathbf{1} $. This gives γ=12\gamma = \frac{1}{2}. I think the issue is that QTTPNQTQ_T^TP_NQ_T is not necessarily a projection matrix. It's not clear to me how much this effects the paper - if it is just interpretive or has consequences for the full pipeline.

Critical strengths

As I said above - the question is relevant (even if not new).

Critical weaknesses

My main concern is the gap between Theorem 1’s discrete identifiability criterion and the proposed null–tangent uncertainty ratio and constraint-strength selection procedure: neither has an established theoretical connection to coefficient recovery, and the incorrect noiseless-limit claim following Eq. (5) overstates the ratio’s interpretation. The experimental support is limited. Teh pendulum example omits an available identifying relation, while the wave-kinetic example retains substantial recovery error and evaluates held-out PDE residuals rather than forward prediction. Consequently, the paper demonstrates a potentially useful diagnostic workflow, but does not yet establish its reliability across realistic inverse problems.

Questions for authors

How do you justify the binary-weight claim following (5), given that exact, noiseless subspaces can have intermediate principal angles, and how would correcting it affect your interpretation of the uncertainty ratio?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Af16 (1/2)

Thank you very much for your careful review. As explained in our general response, I apologize for the delay in our response.

As summarized in our general response, we agree that the technical increment of Theorem 1 is small, and we apologize that the shortened manuscript overstated its novelty and did not situate the work properly. Your pointers to structural identifiability and sloppy models were very helpful. They made it clear that our contribution lies in the framework for PIML and in the position that structural non-uniqueness must be assessed before a learned coefficient function is interpreted, rather than in the theorem itself, and they allowed us to state this novelty concretely. Your suggestion to support Step 3 with sensitivity analyses also led to additional experiments that show the effectiveness and the limits of the framework much more clearly. We are sincerely grateful for both. We now state explicitly that, with T = I, Theorem 1 is the classical rank condition of local identifiability analysis, and we relate it to this literature. Revision: Abstract; Sections 1 and 2.3; Section 3 ("Relation to classical results").

Theorem 1 and Step 3. We agree that Theorem 1 supports only Steps 1–2. Step 3 is a heuristic aid inspired by Theorem 1: it monitors how firmly the structurally undetermined directions are pinned down as λ grows, and gives a hint for the order of magnitude of λ. It does not estimate recovery accuracy, and we now state this explicitly at the beginning of Step 3. Following your suggestion, we added sensitivity analyses: across all combinations of evaluation grid, tangent threshold and eigenvalue floor, the selected λ varies between 10^-3 and 10^2 (Lagrangian) and between 10^-4 and 10^-2 (wave-kinetic), and the mean cosine at the selected weight stays within 0.992–0.9997 and 0.80–0.87, respectively; the rank threshold of M has no effect. The selection is also stable under changes in data size and network architecture, except that with a narrow network the selected weight lies near the onset of the high-recovery regime, where recovery is lower. Regarding the evaluation at the unconstrained solution, we also evaluated the indicator at each constrained solution. Averaged over runs, it selects comparable weights on noiseless data, so estimating the structural directions from the unconstrained network is not an artifact of that choice. Individual runs, however, vary over several decades, because each point of the curve comes from a separately trained network, so averaging over runs is required; with noise, the averaged curve selected too small a weight. Evaluating all weights at one unconstrained solution gives smooth curves and stable selections from a single run, and we keep it in the main text. Revision: Section 4 (Step 3); Section 6; Fig. 5; Appendix L.3, L.4 and L.7 (Tables 11–13).

(continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer Af16 (2/2)

The statement after Eq. (5). Thank you for the counterexample. You are right, and we apologize for the error. K = Q_T^T P_N Q_T is a compression of the null projector, not a projector, so γ_i takes intermediate values even for exact, noiseless subspaces. Our own Fig. 8(b) (Fig. 7(b) in the submitted version) shows this, and our additional runs confirm it: 27.6 of 91.6 tangent directions are exact intersections in the Lagrangian case and 29.3 of 93.3 in the wave-kinetic case, with a strong dependence on the tangent threshold. The computations do not rely on the incorrect statement: R_NT uses the continuous weights, and the hard-threshold variant selects the same λ* in both systems. The correction also connects Step 3 to your point on sloppy models: directions with γ_i = 1 are structurally non-identifiable, whereas directions with 0 < γ_i < 1 are identifiable but weakly constrained. On a personal note, the work on sloppy models is a line of research we have long followed and admired, and we are genuinely glad that your comment showed us how closely it relates to our study. We are very pleased to be able to discuss this connection in the revised paper. Revision: Section 4 (Step 3); Appendix G; Table 11.

Experiments. The pendulum is a controlled test in which the omitted relation is exactly the constraint that Step 2 restores. The non-uniqueness of Lagrangians is classical (Douglas, 1941; Currie & Saletan, 1966), and the expanded Euler–Lagrange equation is linear in (∂_q L, ∂_q̇q̇ L, ∂_qq̇ L), so it lies within the class of Eq. (3). Extracting physics from the learned L also shows why the constraint matters: without it, the effective mass is negative in 80% of held-out states and the derived equation of motion is wrong, even though the prediction error is small. In the wave-kinetic case, we agree that the remaining recovery error is substantial; it reflects ill-conditioned local blocks, with condition numbers up to 2.8 × 10^4 (Appendix F). We relabeled the metric as the held-out PDE residual; forward prediction of I cannot distinguish H from H + F(I). Revision: Section 3; Section 5.1; Fig. 4 caption; Table 6; Table 10.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers (1/3)

Before our response, I would like to apologize sincerely for the delay. During the discussion period, I fell ill with a heavy cold and, at the same time, faced an important personal decision of a kind that comes only rarely in life, which required time and attention. I am truly sorry that this has shortened the time available for discussion. I have tried to make up for it by answering every point as carefully as I could, and I would be very grateful if you could still take a look at our response.

We thank all reviewers for their careful reading and constructive comments. The prior work you pointed out allowed us to state the novelty of our work much more concretely, and the additional experiments you suggested allowed us to demonstrate the effectiveness of the proposed framework much more clearly. We are sincerely grateful for both. Below we summarize how the reviews have changed the presentation of our contribution and what we revised. Changes in the revised PDF are shown in red.

Novelty. We agree that Theorem 1 is elementary and that its technical increment over classical identifiability analysis is small. In shortening the manuscript to the page limit, our presentation came to overstate its novelty, and our discussion of related work was insufficient; we apologize for both. Our contribution is not the theorem but (i) the position that, when PIML is used for physical understanding, the structural non-uniqueness of a learned coefficient function must be evaluated before it is interpreted as physics, and (ii) a routine procedure for this evaluation that applies, without problem-specific analysis, to the broad class of PDEs linear in the values and derivatives of the unknown coefficient function, now stated explicitly as Eq. (3). The procedure is deliberately elementary, so that researchers who are not specialists in inverse problems can use it as a standard check. Step 3 is a heuristic aid inspired by Theorem 1, which gives a hint for the order of magnitude of a soft-constraint weight; it is not a theoretical contribution.

Motivation. This work is motivated by physical-science fields in which inverse problems have rarely been treated explicitly, and in which estimated parameters are sometimes reported as physical properties without examining whether they are identifiable. Several fields have found their own structural degeneracies separately. In instrumented indentation, materials with distinct elastoplastic properties produce indistinguishable load–displacement curves (Tho et al., 2004; Chen et al., 2007), yet several parameters are still extracted from a single curve. In neutron reflectometry, the loss of phase information makes families of density profiles indistinguishable, and the spin-model parameters of α-RuCl3 remain debated a decade after the first proposals (Möller et al., 2025). In time-delay cosmography, relaxing the assumed form of the lens mass profile widens the uncertainty on H0 from about 2.4% to about 5% (Birrer et al., 2020). In each case, the degeneracy was identified by problem-specific analysis, often after the methods were already in use.

(continued in the next comment)

AuthorsOpenReview ↗

General Response to All Reviewers (2/3)

Summary of revisions (changes in red in the revised PDF):

  • Abstract and Section 1: contribution restated as a position and a routine procedure; motivating examples; the class of admissible equations (all reviewers).
  • Section 2.3: sloppy models, identifiability in PINN-based parameter estimation, and learned Lagrangians (Af16).
  • Section 3: the PDE class written explicitly (Eq. (3)), with definitions moved out of Theorem 1, which now states only the equivalence (ch7A); a new paragraph relating Theorem 1 to classical identifiability analysis, uniqueness results for learning PDEs (Scholl et al.; He, Zhao & Zhong), the inverse problem of the calculus of variations, and characteristics (Af16, cmkL, ch7A); a paragraph showing how compatibility reduces the pointwise ambiguity to H + F(I) (cmkL).
  • Section 4 (Step 3): presented as a heuristic aid inspired by Theorem 1, with its idea stated in plain terms; statement on the principal-angle weights corrected; invariance to residual normalization; curvature computed on training residuals (Af16, cmkL, ch7A).
  • Sections 5–6: physical quantities extracted from the learned Lagrangian (Section 5.1); the wave-kinetic weight λ now acts on the raw residuals (Section 5.2); new Fig. 5 comparing rules for choosing λ; Discussion rewritten (Af16, cmkL, ch7A).
  • Appendices: statement on γ_i corrected and relation to network-parameter identifiability (Appendix G); metric relabeled as held-out PDE residual (Table 6); new Appendix L with the additional experiments (Tables 8–13).

The results of the additional experiments are summarized in a separate comment below.

AuthorsOpenReview ↗

General Response to All Reviewers (3/3)

Results of the Additional Experiments

In the wave-kinetic case, the loss is now written as 10^8 (L_PDE + λ L_sym), so that λ weights the raw residuals; the overall factor does not change the balance between the two terms, and the selected weight is λ* = 10^-3 on this scale.

(1) Rules for choosing λ (Reviewers cmkL, Af16; Fig. 5, Tables 8–9). We compared the null–tangent (NT) rule with validation on the physics residual alone, validation including the constraint residual with a fixed weight, a fixed weight λ = 1, and gradient-norm balancing (GNB; Wang et al., 2021). Fixed-weight rules use the raw residuals, because normalization factors chosen by hand amount to tuning λ in advance.

  • Validation on the physics residual alone is unreliable (cosine 0.50 in the wave-kinetic case and 0.72 on noiseless Lagrangian data), as claimed in the paper.
  • In the wave-kinetic case, GNB (0.64), a fixed weight λ = 1 (−0.33; these models were trained additionally) and fixed-weight validation (0.68) also fail. If the residuals are first rescaled by hand-chosen factors, the fixed-weight rules give about 0.87.
  • The NT rule selected λ = 10^-3 in all ten wave-kinetic slices (0.845; post-hoc optimum 0.876) and never selected the collapse region.
  • In the Lagrangian case, whose objective has no normalization factors, the fixed-weight rules and GNB were more accurate than the NT rule (0.9997 against 0.992 without noise and 0.980 with noise σ = 0.1).

Across all conditions, the NT rule was the only rule that did not fail: its lowest cosine similarity was 0.845, whereas every other rule fell to 0.68 or below in at least one case.

(2) Physical quantities (Reviewers ch7A, Af16, cmkL; Table 10). Without the constraint, the pendulum model with small prediction error has a negative effective mass ∂²L/∂θ̇² in 80% of held-out states, and the equation of motion derived from it has a median error of 5.3 against accelerations of order one. With the constraint, the mass is recovered as 1.00 and the equation of motion to within 0.006. At the NT-selected λ, the squared small-oscillation frequency is 0.99 without noise and 1.10 with σ = 0.1 (true value 1).

(3) Sensitivity (Reviewers Af16, cmkL; Appendix L.3, Table 11). Across all 20 combinations of evaluation grid (6×6 to 12×12), tangent threshold (10^-2 to 10^-4) and eigenvalue floor (10^-12 to 10^-4), the selected λ ranges from 10^-3 to 10^2 (Lagrangian) and from 10^-4 to 10^-2 (wave-kinetic), and the mean cosine at the selected weight stays within 0.992–0.9997 and 0.80–0.87, respectively. Changing the numerical-rank threshold of M (10^-10 to 10^-6) leaves the selection unchanged. Intermediate γ_i are common (27.6 of 91.6 and 29.3 of 93.3 tangent directions are exact intersections), and their count depends strongly on the tangent threshold.

(4) Data size and architecture (Reviewer cmkL; Appendix L.4, Table 12). Using 1/2 or 1/5 of the data on the same trajectories left the selected weight (10^-3) and the recovery (0.99) unchanged. A wider network (width 128) gave 0.997, and a 2-50-10-1 network in the wave-kinetic case gave 0.858 (post-hoc optimum 0.873; physics-residual and fixed-weight validation 0.50 and 0.54). With a narrow network (width 32), the selected weight lies near the onset of the high-recovery regime, where recovery is lower (0.89).

(5) Compatibility on the wave-kinetic data (Reviewers cmkL, ch7A; Appendix L.5). On a 50 × 50 subgrid, the pointwise null space has dimension 2500. With the compatibility map, 54 ± 4 near-null directions remain, and they are functions of I, as predicted by the family H + F(I). The reflection constraint leaves only the additive constant.

(6) Cost (Reviewer cmkL; Appendix L.6). R_NT(λ) for every λ is obtained at one unconstrained solution from one Jacobian and two SVDs, so no model has to be trained per candidate λ.

(7) Evaluation at the constrained solutions (Reviewer Af16; Appendix L.7, Table 13). Evaluating the indicator at each constrained solution instead requires averaging over runs, which vary over several decades. It gave comparable weights without noise but weights in the collapse region with noise. The evaluation at the unconstrained solution is thus at least as reliable and stable from a single run.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.