acceptodds
Under review as a conference paper at ICLR 2027

PASL-U: PHYSICS-AWARE SOURCE RECOVERY WITH TEMPERATURE-PARAMETERIZED OPERATORS FOR GUIDED-WAVE DAMAGE LOCALIZATION

Abstract

In guided-wave sensing, sensors record the source waveform that has been altered by propagation through the structure, so recovering the unknown source is an inverse problem that amounts to learning a mapping from measurements back to sources. However, the propagation operator changes with the environment: temperature changes the structure’s stiffness, density, and wave speed, and thus changes the source-to-sensor propagation characteristics. Conventional operator learning cannot separate environmental effects from source effects, and therefore attributes the former to the source. Outside the calibration range, the influence of the environment on propagation gradually increases, causing source recovery to degrade. We propose PASL-U (Physics-Anchored Operator Learning for Source Recovery under Environment Shifts), which preserves the environmental dependence of a physical propagation prior, learns a bounded correction from calibration environments, and explicitly predicts how this correction changes with the environment. The corrected operator is then frozen and used for source recovery in unseen environments. Experiments show that PASL-U maintains lower propagation and source errors under strict scalar and coupled-parameter extrapolation within the tested range. On measured guided waves, it also retains damage-recognition performance in temperature bands outside the propagation-calibration range.

Primary area:
learning on time series and dynamical systems
Keywords:
Operator learning, Inverse source recovery, Environmental extrapolation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_fp66OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a physics-based operator learning framework, PASL-U, for source recovery under environmental changes. This method employs a two-stage learning strategy, integrating physical propagation priors with bounded operator corrections influenced by the environment. Experimental results on synthetic convection-diffusion problems and synthetic and measured guided wave data demonstrate that its performance is superior to multiple baseline methods.

Critical strengths

The paper addresses a practically relevant problem of source recovery under environmental shifts, which is important for improving the robustness of physics-guided inverse modeling.

Critical weaknesses

1.The proposed framework mainly combines physical propagation priors with existing mudoles, such as bounded operator correction and multi-layer perceptrons (MLPs) under environmental conditions. It does not clearly explain the fundamental differences between this method and existing physically-based operator correction methods.

2.The experimental comparison failed to fully distinguish the advantages of the proposed method. There are significant differences in the use of physical priors and reconstruction strategies among the various baselines. Therefore, it is unclear whether the improvement is due to the proposed operator correction mechanism or to the differences in the experimental setup. More controlled comparisons are needed to verify the effectiveness of this design.

3.Tables 10 and 11 report inconsistent results under seemingly identical settings. For example, PASL-U's single-source NMSE at 50°C differs substantially (0.83 vs. 0.313), despite both tables reporting five-seed evaluations at the same noise level. Similar discrepancies appear across other methods and temperatures.

Questions for authors
  1. Beyond combining existing techniques, what are the fundamental methodological innovations of PASL-U compared with existing physics-guided operator correction approaches?

  2. Since the joint single-stage optimization has achieved considerable performance, what are the specific advantages of proposing the two-stage learning strategy?

  3. How does the author distinguish the improvements brought about by the proposed correction mechanism from those resulting from the differences in physical priors and reconstruction strategies among various baselines?

5 comments
AuthorsOpenReview ↗

**W1 / Q1 — Fundamental novelty relative to existing physics-guided operator corrections.**

Existing physics-guided correction methods (Lunz-FAC, Neural Green's Operators, Neumann-NGO, PINO-style) share a common property: they learn a correction at the calibration environment but do not explicitly model how the correction changes with the environment. They therefore learn a correction specific to the calibration environment; under shift they either freeze it or directly extrapolate the network output, with no environment-aware extrapolation mechanism.

PASL-U differs fundamentally in three ways, each tied to equations and experiments.

  1. The correction family and the environment-to-code map are separated; what is extrapolated is a low-dimensional code. We first learn a low-rank correction family (Eq. 2) ΔL(z)=ΔL0+∑k=1Kzk ΔLk\Delta L(z)=\Delta L_0+\sum_{k=1}^{K} z_k\,\Delta L_k, then learn an environment-to-code map z^ϕ(e)\hat z_\phi(e) (Eq. 5). At extrapolation, only this KK-dimensional code z^ϕ(e)\hat z_\phi(e) is extrapolated while the family stays frozen. No existing method has this structure — they either learn the full correction directly or feed the environment into the network, but none has the "low-rank family + environment code" layer, which makes "how the correction varies with the environment" an explicit, extrapolatable, and testable object.
  2. The correction is bounded and anchored, so extrapolation cannot blow up. The code box ∣zk∣≤ak|z_k|\le a_k together with elementwise normalization (Eq. 3) guarantees ∥ΔL(z)∥≤ρ M⊙S\|\Delta L(z)\|\le\rho\,M\odot S, and the reference point is anchored to zero (z^ϕ(e0)=0\hat z_\phi(e_0)=0). Together with the budget θ=pβ<1\theta=p\beta<1, the corrected operator remains invertible at any environment with bounded error (Prop. 1). Existing corrections lack this joint "code box + anchor + budget" guarantee.
  3. The two-stage structure yields a provable error decomposition and an improvement radius. Eq. (7) splits the error into εfam+ce∥z^ϕ−z†∥\varepsilon_{\mathrm{fam}}+c_e\|\hat z_\phi-z^\dagger\|, corresponding to Stage A and Stage B; this yields the improvement radius (Eq. 11*, new in this response) describing when correction stops beating the physics prior. Existing methods have neither this decomposition nor a testable criterion.

Experimentally these benefits appear directly in extrapolation: under strict OOD (μ\mu from ≤1.2\le 1.2 to 2.02.0), coupled two-parameter (to [2.0,2.0][2.0,2.0]), and three/four-parameter settings, PASL-U stays accurate while DeepONet/FNO (no physics structure) and Data-NGO/PINO (no environment-code extrapolation) degrade quickly. PASL-U's advantage is therefore not a "better physics prior" but an explicit environment-to-code extrapolation mechanism that is bounded, decomposable, and testable.

What we will do in the camera-ready version. (i) State the three innovations in the introduction contributions as one sentence: a low-rank correction family with an environment-to-code extrapolation structure, a bounded and anchored correction, and a provable error decomposition with an improvement radius; (ii) use the No Stage B ablation in Table 6 (OOD two-source NMSE 0.3122 vs 0.2758 full) as direct evidence that the environment-code extrapolation layer, rather than a simple combination of existing modules, is the source of the gain.

AuthorsOpenReview ↗

**W2 / Q3 — How to separate the correction mechanism's contribution from baseline setup differences.**

The reviewer worries that baselines differ greatly in physical priors and inversion strategies, so PASL-U's advantage could come from setup differences rather than the mechanism. We explain how the confound is controlled and give direct evidence isolating the mechanism.

  1. The inversion strategy is identical. All source-linear methods (PASL-U, Physical-only, Data-NGO, Neumann-NGO, Lunz-FAC) are evaluated with the same group-sparse proximal-gradient solver, regularization (0.05), and iteration count (150). The "different inversion strategy" factor is thus already fixed among source-linear methods; the remaining differences are not in the solver.
  2. The physics prior is also controlled: the cleanest comparison is PASL-U vs Physical-only. They use exactly the same physics prior LphysL_{\mathrm{phys}}, geometry (B,C)(B,C), solver, and data; the only difference is that PASL-U additionally learns a bounded correction ΔL\Delta L. The gain over Physical-only can therefore only come from this mechanism: since two-stage and joint single-stage accuracy are comparable (Part 2), the only remaining difference is the bounded correction ΔL\Delta L. The data confirm it: PASL-U is strictly better in the far-extrapolation regime (two-source NMSE 0.4059 vs 0.4683 at μ=2.0\mu=2.0) and overtakes again on the three/four-dimensional extension of the coupled two-parameter setting.
  3. Operator error is more direct evidence than source error. Source NMSE goes through inversion and can be affected by noise and conditioning; operator relative error ∥G^−G⋆∥/∥G⋆∥\|\hat G-G^\star\|/\|G^\star\| directly measures how far the corrected propagation operator is from the true operator, before inversion. PASL-U has the lowest operator error in every environment (0.023–0.076 vs 0.069–0.161 for Physical-only in the two-parameter table C.5), showing the mechanism itself makes the operator more accurate — evidence independent of inversion settings. For completeness, the single-parameter operator errors (a different setting from the two-parameter C.5 numbers quoted above) are:
µPhysical-only operator errorPASL-U operator error
1.40.14950.1234
1.60.16170.1357
1.80.16850.1504
2.00.17410.1657
  1. Controlled experiments further isolate causality. We ran two injection experiments: (i) fixing the task and interpolating the inversion operator from the true operator to the physics prior, source NMSE rises monotonically with operator mismatch (μ=1.6\mu=1.6: 0.4215 → 0.4611; μ=2.0\mu=2.0: 0.3891 → 0.4249), proving "operator mismatch → source error" is causal, not coincidental; (ii) injecting a full-rank, non-smooth step residual, PASL-U does not collapse and is no worse than Physical-only (0.166 vs 0.174 at μ=2.0\mu=2.0). Both experiments fix the task and solver and change only the operator/correction, pointing directly at the mechanism.
EnvironmentλOperator errorTwo-source NMSE
µ=1.60 (true operator)0.0000.4215
µ=1.60.50.0460.4412
µ=1.61 (physics prior)0.0930.4611
µ=2.00 (true operator)0.0000.3891
µ=2.00.50.0440.4068
µ=2.01 (physics prior)0.0870.4249
  1. The shared-background guided-wave experiment follows the same logic. All methods subtract the same healthy background and use the same inversion solver, comparing only their propagation corrections; after decoupling the background, PASL-U remains lowest across all three temperature bands (single-source 0.0277 / 0.0565 / 0.0990).

In summary, the setup-difference confound is controlled: the solver is unified, the physics prior is aligned through PASL-U vs Physical-only, and the background is aligned through the shared protocol; the operator error, fixed-task injections, and shared-background experiment all point to the same conclusion — PASL-U's gain comes from the propagation correction mechanism itself, not from setup differences.

AuthorsOpenReview ↗

**Q2 — What advantage does the two-stage strategy have, given that joint single-stage optimization already achieves comparable accuracy?**

The core advantage of the two-stage strategy over joint single-stage training is a drastic reduction in recalibration cost, not an improvement in accuracy; they are comparable in accuracy. Stage A learns the shared correction modality, which captures the systematic deviation of the operator along environment-independent dimensions. Once learned, this modality is frozen and reused directly for new environments without any retraining, incurring no recalibration cost. Stage B learns only a lightweight environment-to-code map (412 parameters) that converts environmental information into a low-dimensional code driving the correction modality to adapt to the current environment. At each recalibration, only these 412 parameters plus the per-environment free code (54 parameters), totaling 466, need to be refitted.

ComponentParameter countDuring recalibration
Stage A shared correction modality912,247reused, not relearned
Stage A per-environment free code54refitted
Stage B environment→code map412refitted
Environment-dependent total466refitted
By contrast, joint single-stage training couples the modality and the map together, so recalibration requires relearning the entire roughly 912,000-parameter modality and map from scratch, about 2000 times more parameters and nearly twice the runtime. Whenever the sensor array, temperature range, or field conditions change, the joint single-stage approach must rerun the full training pipeline, which is costly and hard to maintain in engineering deployment.
As shown in the table below, the two-stage scheme updates only 466 parameters at recalibration, taking around 22.6 s, whereas joint single-stage must update 912,659 parameters, taking around 44.6 s. The two-stage approach compresses the recalibration parameter count to 0.05% of the joint scheme and cuts the runtime nearly in half.
SchemeRecalibration contentParameter countTime (measured)
Two-stagerefit code + map only466~22.6s
Joint single-stagerelearn modality + map912,659~44.6s
Reviewer_XCp6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a new method for learning a mapping from measurements to the source term in a PDE that has generated them. In particular, it focuses on the problem of handling the fact that the propagation operator changes with the environmental conditions, which can cause much variability. In particular, PASL-U learns an environment-dependent correction to an approximate physical propagation operator, which is subsequently used for source reconstruction. The work then applies this method to three different problems: two synthetic and one real dataset. The authors demostrates the effectiveness of the proposed method on all three problems, comparing it to some baseline methods.

Critical strengths
  • The authors have clearly stated the contributiuons of the paper and the novelty of the proposed method, including some theoretical analysis.
  • The operator learning framework is physically-informed, rather than a pure inverse problem approach, accounting for the specific environmental variations.
Critical weaknesses
  • The authors have dedicated lots of space to the presentation of the methodology and the discussion of the results on the synthetic datasets only, leaving the real dataset analysis for a very short section. Therefore, the discussion of the results and the limitations and conclusions are not fully comprehensive. In particular, the methodology is nicely and carefully presented, but overall the results section is not as comprehensive as it could be.
  • The results in Table 3 show a relatively high uncertanty, raising concerns about the consistency and robustness of the proposed method.
Questions for authors
  • In the 4.1 section, you adopt 64 interior sensors for the synthetic dataset. What would be the effect of using more or less sensors? How does this choice is transferred to the real dataset?
  • As mentioned above, the discussion and presentation of the results on the real dataset is quite brief and would benefit from a more comprehensive analysis. I wonder if it could be possible to summarise the main findinds of the synthetic datasets in the main manuscript and leaving the details in the appendix, so that the main manuscript is more focused on the real dataset analysis? Furthermore, I would be interested in the high variability of the results in Table 3, could you provide more details on the reason why this is the case?
4 comments
AuthorsOpenReview ↗

**W1 / Q2 — Real-dataset analysis and the proposed restructure.**

We agree that the real-dataset section is brief relative to the synthetic experiments. Rather than only promising an expansion, we give the core real-dataset results here; the expanded section will contain exactly these tables.

Per-class damage identification. PASL-U classifies four damage classes (D04/D12/D16/D24) on the measured OGW2 data. Per-class recall, mean ± sample SD over five seeds:

Temperature bandD04D12D16D24
45–50 °C0.990±0.0221.000±0.0000.910±0.1190.970±0.027
50–55 °C0.590±0.4641.000±0.0000.280±0.3270.380±0.520
55–60 °C0.190±0.2940.724±0.3870.000±0.0000.448±0.514

At 45–50 °C all four classes are recovered with high recall and low variance. In the far bands D04/D16/D24 degrade while D12 dominates, and D16 is never recovered at 55–60 °C — this is the far-OOD information limit, not a failure of source recovery (W2 gives the mechanism).

Deployment cost (measured OGW2, 66 reciprocal paths, 194 intact calibration records), five-seed mean ± SD.

MethodTraining (s)Recovery (s)Total (s)Peak GPU (GB)
Physical-only0.00 ± 0.002.53 ± 0.0821.03 ± 0.820.016 ± 0.000
PASL-U298.37 ± 11.993.76 ± 0.66321.29 ± 13.650.224 ± 0.000
Data-NGO640.78 ± 60.252.46 ± 0.07661.49 ± 60.980.148 ± 0.000
Neumann NGO35.08 ± 0.453.72 ± 0.1257.15 ± 0.940.435 ± 0.000
PINO-style4.15 ± 2.122.55 ± 0.0725.06 ± 1.730.018 ± 0.000
Lunz-FAC284.14 ± 7.953.55 ± 0.09306.65 ± 8.751.375 ± 0.000
DeepONet68.94 ± 7.270.062 ± 0.00389.44 ± 7.820.088 ± 0.000
FNO137.89 ± 7.640.086 ± 0.002158.79 ± 8.690.103 ± 0.000

Stability and robustness of the real-data results. Two points. (i) The source-recovery residual stays flat across the three bands (0.977 / 0.975 / 0.974), so the recovered source is stable; the instability sits in the downstream readout, whose cause is analyzed in W2 (Part 2). (ii) The near-chance far-band result is robust to the data partition: across three data splits (20260903 / 20260904 / 20260905), PASL-U's 55–60 °C balanced accuracy is 0.329 / 0.283 / 0.341, so it is not an artifact of one partition.

Sensor-count transfer. The synthetic study is anchored at 64 interior sensors because the real OGW2 uses 66 reciprocal paths (about 64 measurement channels); the full sensor-count sweep, showing that more sensors monotonically improve source recovery, is given in Q1 (Part 3).

We will move the synthetic per-environment details and the grid/sensor study to the appendix, and make the real-dataset analysis above (per-class identification, deployment cost, stability, and the sensor-count transfer) the focus of the main manuscript.

AuthorsOpenReview ↗

**W2 / Q2 — The Table 3 variability.**

The large variance in the far band does not originate from PASL-U itself; rather, the classifier head cannot effectively discriminate between classes in that temperature range, causing some seeds to collapse to a single majority class. A temperature-conditioned (FiLM) head eliminates this collapse, but the improvement is bounded by source-feature quality.

Reporting and data splits. Every number is a balanced accuracy reported as mean ± sample standard deviation over five training seeds (20260903–20260907), with no seed omitted. The linear head is evaluated on three data splits (20260903, 20260904, 20260905), each with five seeds; the temperature-conditioned (FiLM) head is evaluated on one data split (20260905) with five seeds — the same split as the third linear split, so the two heads are directly comparable. Across the three splits the linear-head conclusion is unchanged (PASL-U at 55–60 °C: 0.329 / 0.283 / 0.341), so the near-chance far-band result is not a partition artifact.

The variance is a readout collapse, not noise. The large standard deviation is concentrated in the far band and comes from two of five seeds collapsing to chance (balanced accuracy 0.25) with per-class recall [0, 1, 0, 0] — the linear classifier predicts a single majority class. Meanwhile the recovery residual stays flat (0.977 / 0.975 / 0.974), so the instability is in the classifier head, not in the recovered source.

A temperature-conditioned head removes the collapse. We retrained the same recovered features with a temperature-conditioned (FiLM) linear head whose decision boundaries adapt to temperature. Balanced accuracy (mean ± SD over five seeds, split 20260905):

Method45–50 linear45–50 FiLM50–55 linear50–55 FiLM55–60 linear55–60 FiLM
Physical-only0.950±0.0320.945±0.1230.385±0.0540.753±0.2360.250±0.0000.560±0.217
PASL-U0.968±0.0340.885±0.0350.562±0.3050.605±0.1330.340±0.1010.471±0.123
Data-NGO0.535±0.0260.570±0.1750.250±0.0000.300±0.1120.250±0.0000.250±0.000
Neumann NGO0.392±0.0410.365±0.0640.253±0.0060.255±0.0110.243±0.0070.252±0.005
PINO-style0.340±0.0290.305±0.0260.250±0.0000.250±0.0000.250±0.0000.250±0.000
Lunz-FAC0.740±0.2220.628±0.1840.303±0.0460.318±0.0560.264±0.0210.260±0.016
DeepONet0.2500.2500.2500.2500.2500.250
FNO0.2500.2500.2500.2500.2500.250

The FiLM table is a diagnostic for the variance, not a re-ranking of methods. The reported protocol is the linear head: there PASL-U beats Physical-only in all three bands (0.968 vs 0.950, 0.562 vs 0.385, 0.340 vs 0.250), and on source recovery itself PASL-U is lowest in every band (single-source NMSE 0.028 / 0.057 / 0.099 vs 0.029 / 0.062 / 0.103). Physical-only's larger FiLM gain is a shortcut: without operator correction its features keep the uncorrected temperature drift, which the temperature-conditioned head exploits, giving a large but unstable boost (SD 0.236 / 0.217); PASL-U has already removed that drift inside the operator, so its gain is smaller but stable — the informative quantity is the variance, where PASL-U is the most stable method. The 45–50 °C drop (0.968→0.885) is the expected robustness–accuracy tradeoff: the linear head is already near-saturated there, and the extra FiLM parameters only overfit the small development set.

AuthorsOpenReview ↗

**Response to Q1.**

You ask what happens with more or fewer interior sensors, and how the 64-sensor choice transfers to the real dataset. The short answer: more sensors monotonically improve source recovery, because denser observation better constrains the inverse problem — the learned operator itself barely changes — and the 64-sensor setting matches the real dataset's roughly 64 measurement channels. The grid size is limited only by the compute of a single RTX 3080, not by the method. Here is the evidence. We ran a controlled grid-and-sensor study (three seeds 20260903–20260905, fixed physical geometry, nested sensor layouts S25 ⊂ S64 ⊂ S144) over three grids. Results are mean ± sample SD over the three seeds:

GridSensorsSource NMSEOperator rel. errorTraining (s)
21×21250.5223±0.00430.0314±0.002143.21±0.98
21×21640.4231±0.00310.0400±0.001947.34±0.67
21×211440.3434±0.00180.0426±0.002456.26±0.12
25×25250.5243±0.00470.0292±0.0046454.63±0.72
25×25640.4083±0.00490.0389±0.0066470.44±10.96
25×251440.3714±0.00620.0431±0.0067543.48±24.85
29×29250.5192±0.00460.0223±0.0013639.8±15.9
29×29640.3932±0.00250.0315±0.0018638.9±2.5
29×291440.3123±0.00070.0363±0.0035665.5±29.4
Two things follow from the table. First, more sensors improve source recovery monotonically: at 21×21, source NMSE goes 0.522 (25 sensors) → 0.423 (64) → 0.343 (144), while the operator relative error stays roughly flat (0.031–0.043). So the gain comes from the denser observation, not from a better operator. Second, the sensor count and the number of directly observed source positions co-vary in the nested layout (5 / 9 / 25), so the improvement is the combined effect of sensor density and source coverage; we state this explicitly in the revision.
On the transfer to the real dataset: OGW2 uses 66 reciprocal paths from 12 transducers, i.e., roughly 64 measurement channels, which is also why we anchored the synthetic study at 64 sensors. The real dataset therefore sits near the middle of the synthetic sweep, where source recovery is already well-constrained. The cost scaling with the grid (43–666 s training, 1.0–4.4 GB GPU) further indicates that finer discretization is mainly limited by the available compute; the 29×29 setting reported here is simply the practical limit of the single RTX 3080 we used, not a limitation of the method itself. We have added this clarification to the manuscript so that readers do not infer that the method is restricted to 29×29.

I thank the authors for their detailed responses and for providing additional experimental results. The rebuttal addresses several of my questions. Although the proposed restructuring and expansion of the real-data analysis would improve the manuscript, these changes are not yet reflected in the current version of the manuscript. Overall, I acknowledge that the rebuttal has clarified several aspects of the work and addressed some of my original concerns. Nevertheless, I would be inclined to keep my original assessment.

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

The authors consider source recovery when the forward operator depends on an environmental parameter and must be used outside its calibration range. They write the operator as an approximate physical operator plus a bounded correction, built from learned modes with weights predicted from the environment, and keep the corrected operator fixed to recover the source. They derive error bounds that grow linearly with the distance from the calibration range and evaluate the method on simulated advection-diffusion and on simulated and measured guided waves.

Critical strengths

Correcting an approximate physical operator, instead of learning the inverse map directly, is a reasonable way to retain physical structure.

Critical weaknesses
  • While operators that change with the environment are a real concern, it is unclear when a correction learned in a few calibration environments can be extrapolated beyond them at all.

  • The theory does not show what the method achieves under extrapolation. The bound grows linearly with distance because the predicted and the ideal corrections are both assumed to be Lipschitz in the environment, and it does not indicate when the learned correction stops improving on the uncorrected operator.

  • While the authors present their method as a general framework, they report no runtimes, memory figures or scaling results. Training requires a solve with the corrected operator for every calibration environment and frequency at each of thousands of updates, yet the only discretized problem is a 21x21 grid, which begs the question of whether this is applicable to realistic settings.

  • The evidence do not support the claims:

    • On advection-diffusion, source error therefore does not track operator error, so this benchmark does not test the premise that operator mismatch is amplified during inversion.
    • On simulated guided waves, the source error at the two farther test environments is worse than returning a zero source. Tables 10 and 11 disagree on these values.
Questions for authors
  • Could the authors state conditions on the physics under which the smoothness assumption (i.e., ideal correction varies smoothly with the environment) holds, and show how the method behaves when the discrepancy lies outside the correction family?

  • How do the cost and the admissible correction scale with the discretization and with the number of environmental parameters?

8 comments
AuthorsOpenReview ↗

**W1 — When can a correction learned on a few calibration environments extrapolate beyond them?**

We thank the reviewer for this important question. Our answer: PASL-U does not extrapolate the full operator directly from data; it evaluates an exact physics prior and extrapolates only a small residual correction anchored to that prior. Whether extrapolation works is therefore determined by two testable conditions.

We write L^(e)=Lphys(e)+ΔL(z^ϕ(e))\widehat L(e)=L_{\mathrm{phys}}(e)+\Delta L(\hat z_\phi(e)). The physics prior Lphys(e)L_{\mathrm{phys}}(e) is analytic and can be evaluated at any environment, including outside the calibration range, so it needs no extrapolation; only the learned correction code z^ϕ(e)\hat z_\phi(e) does. Two conditions.

First, invertibility θ=pβ<1\theta=p\beta<1. p=∥Lphys(e)−1∥2p=\|L_{\mathrm{phys}}(e)^{-1}\|_2 is computed from the prior, and β=ρ∥M⊙S∥F\beta=\rho\|M\odot S\|_F from our budget hyperparameters; neither uses test data. When θ<1\theta<1, the corrected operator stays invertible and bounded at any environment, so extrapolation cannot make it suddenly ill-conditioned.

Second, residual smoothness. The code extrapolates because its smoothness is inherited from the physics: smooth PDE coefficients ⇒\Rightarrow smooth true operator ⇒\Rightarrow the same-form prior is smooth ⇒\Rightarrow their difference (the residual) is smooth ⇒\Rightarrow the ideal code is smooth. A smooth function extrapolates from finitely many points with error growing linearly in distance — the Lipschitz bound, a consequence of coefficient smoothness and the same-form prior, not an extra assumption.

Two experiments support these conditions.

First, the prior absorbs the dominant trend. If it did not, we would still be extrapolating "the whole operator". Table C.5 (appendix) gives the propagation-operator relative error under coupled two-parameter settings:

EnvironmentPASL-UPhysical-onlyPINO-styleData-NGONeumann-NGO
[1.6,1.0]0.04030.12510.24450.42010.2568
[2.0,1.0]0.07030.12240.34770.85170.6559
[1.0,1.6]0.03260.06890.15720.34680.0446
[1.0,2.0]0.06130.08410.22320.54910.0445
[1.4,1.4]0.02300.13100.17520.38860.1322
[1.6,1.6]0.04350.14590.44360.66070.2983
[1.8,1.8]0.06070.15420.75101.05210.5123
[2.0,2.0]0.07640.16130.59651.52770.7787

The prior's error is 0.069–0.161, so it already carries the dominant environment variation; PASL-U pushes it to 0.023–0.076. The residual to extrapolate is thus an order of magnitude smaller than the full operator's variation — the premise of "extrapolating a small residual". This ties to the second condition: in Table C.5 the PASL-U column grows slowly with distance, and Figure 3 shows only the code-prediction error grows while the family representation error stays roughly constant.

Second, when the conditions hold, extrapolation works. We verify with source-recovery error, not operator error alone. The rerun 5-seed Table 1 (scalar strict OOD, two-source NMSE @30 dB):

Methodµ=1.4µ=1.6µ=1.8µ=2.0
Physical-only0.5142 ± 0.00000.5056 ± 0.00000.4898 ± 0.00000.4683 ± 0.0000
PASL-U0.4760 ± 0.00060.4544 ± 0.00100.4310 ± 0.00180.4061 ± 0.0029
Data-NGO0.5277 ± 0.03210.5591 ± 0.05920.6091 ± 0.06620.6528 ± 0.0694
Neumann NGO0.4745 ± 0.00360.4485 ± 0.01600.4355 ± 0.03850.4800 ± 0.0794
Lunz-FAC0.5159 ± 0.00070.5078 ± 0.00090.4925 ± 0.00140.4717 ± 0.0018
PINO-style0.6694 ± 0.01850.6786 ± 0.02730.6541 ± 0.03940.7331 ± 0.0108
DeepONet0.8829 ± 0.00350.8993 ± 0.00290.9078 ± 0.00320.9027 ± 0.0041
FNO1.0010 ± 0.00431.0055 ± 0.00271.0119 ± 0.00321.0151 ± 0.0044

At µ=1.8, 2.0, PASL-U attains both the lowest operator error and the lowest two-source NMSE, with nearly identical five seeds — extrapolation stability is genuinely preserved where θ<1\theta<1 and residual smoothness hold.

Therefore: when the physics prior is invertible, θ=pβ<1\theta=p\beta<1, and the true residual is smooth in the environment, a correction learned on a few calibration environments can be extrapolated. θ\theta is computed directly from the prior and hyperparameters, and smoothness is estimated by finite differences on the calibration residual and validated on held-out environments — both testable.\n\nWhat if the assumptions break. Even if the residual is not smooth, or the mismatch falls outside the learned correction family, the correction remains bounded elementwise by ∣ΔL(z)∣≤ρM⊙S|\Delta L(z)|\le\rho M\odot S and anchored at zero at the reference, so the method degrades gracefully back to the physics prior rather than diverging: extrapolation fails softly, not catastrophically.

AuthorsOpenReview ↗

**W2 — Why the bound grows linearly, and when does improvement stop?**

You observed that in the advection-diffusion experiment source error does not follow operator error, and question whether this benchmark actually tests the premise that "operator mismatch is amplified in inversion". We agree the observation is correct, but it has two expected, explainable causes, and we add a fixed-task separation control that directly proves the causal relation.

Cause 1: the near field is dominated by the noise floor, the far field by operator mismatch. Source-recovery error is governed by two terms — noise and operator mismatch (the latter amplified by the inversion condition number). When mismatch is small, noise dominates and source error is insensitive to operator error; only once mismatch grows beyond the noise floor does source error start to follow operator error. Hence in the near field (µ=1.4, 1.6) all methods' operator errors are small and source NMSE stays near the 0.45–0.48 noise floor, with ordering set by noise/solver — exactly the "not following" you observed.

Cause 2: the task itself changes with µ. The source-recovery task is defined by the true operator G⋆(μ)G^\star(\mu) (with noise). As µ changes, the true operator's condition number and the source–sensor coupling change, so even with an exactly correct operator, the degree to which 30 dB noise is amplified and the recoverable baseline accuracy differ across µ. This is directly visible in the data:

Physical-onlyµ=1.4µ=1.6µ=1.8µ=2.0
Operator rel. error0.08530.09270.09130.0871
Two-source NMSE0.51420.50560.48980.4683

Physical-only's operator error barely changes with µ, yet its source NMSE falls monotonically from 0.5142 to 0.4683 — a drop not explainable by operator error (which is flat), only by the task becoming easier. This is the conditioning of the inverse problem, not the operator: with the true operator (λ=0 below) the Oracle NMSE drops from 0.4215 at µ=1.6 to 0.3891 at µ=2.0, so the baseline difficulty of the task itself decreases as µ grows. The cross-environment comparison is a tug-of-war between a task that gets easier (good) and an operator mismatch that grows (bad); in the near field the easier task wins, and in the far field the mismatch wins.

Conditioning-controlled ablation. To directly answer whether the far-field "following" is merely coincidence, we ran a strictly controlled ablation that varies only the operator while the task, noise, and solver stay fixed: choose µ=1.6 and µ=2.0, keep the true operator, source, and noise unchanged, and interpolate the inversion operator G^\widehat G from the true operator to the physics prior,

G^(λ)=G⋆+λ(Gphys−G⋆),λ∈{0,0.5,1},\widehat G(\lambda)=G^\star+\lambda(G_{\mathrm{phys}}-G^\star),\quad \lambda\in\{0,0.5,1\},

so operator mismatch grows monotonically from 0 to the prior's mismatch (0.087–0.093) while the task is fully fixed. Results (two-source NMSE, 150 iterations, regularization 0.05):

µ=1.6Operator errorTwo-source NMSE
λ=0 (true operator)0.0000.4215
λ=0.50.0460.4412
λ=1.0 (physics prior)0.0930.4611
µ=2.0Operator errorTwo-source NMSE
λ=0 (true operator)0.0000.3891
λ=0.50.0440.4068
λ=1.0 (physics prior)0.0870.4249

Under the same task, noise, and solver, source NMSE rises monotonically with operator mismatch (µ=1.6: 0.4215→0.4611; µ=2.0: 0.3891→0.4249). This directly proves the causal direction "operator mismatch → source error", not a coincidence of cross-environment comparison — the only thing that changes is the mismatch magnitude. This control turns the qualitative "noise-dominated near, operator-dominated far" explanation into a separable causal test.

The premise holds in the far field, and that is where it should be tested. "Operator mismatch is amplified in inversion" is a far-field phenomenon: it dominates source error only when mismatch exceeds the noise floor and the environment-dependent condition number amplifies it. The far-field data do exactly this:

Two-source NMSEµ=1.4µ=1.6µ=1.8µ=2.0
PASL-U0.47600.45440.43090.4059
Neumann NGO0.47450.44850.43550.4800
Physical-only0.51420.50560.48980.4683
Data-NGO0.52770.55910.60910.6528

At µ=1.8, 2.0, PASL-U with the lowest operator error (0.050, 0.066) also attains the lowest two-source NMSE (0.4309, 0.4059), while methods with larger operator error (Neumann NGO 0.22–0.42, Data-NGO 0.67–1.01) get correspondingly worse or rising source NMSE — here source error strictly follows operator error. Single-source NMSE is fully consistent.

AuthorsOpenReview ↗

**W3 — Runtime, memory, and scaling.**

We thank the reviewer for the focus on practicality, and we understand the core concern is whether the method can actually be deployed. To make the cost transparent we split it into a one-time offline training and online inference at deployment, and give comparison data for all baselines, including real measurements.

The concern that "every calibration environment, frequency, and update requires solving a correction operator" applies to the offline training stage only, and there it is one torch.linalg.solve batched over (environment × frequency) on GPU, not a serial per-environment, per-frequency solve; what actually needs learning and updating is only the low-rank correction (rank K=6), with O(K⋅N)O(K\cdot N) parameters, far smaller than the full operator. Training is done once; at deployment the operator is frozen, and for any test environment e⋆e^\star the operator G^(e⋆)=G^(e⋆,z^ϕ(e⋆))\widehat G(e^\star)=\widehat G(e^\star,\hat z_\phi(e^\star)) is assembled once from "physics prior + predicted correction code"; source recovery is then a fixed number of group-sparse proximal-gradient iterations (forward/adjoint matrix products) — no further learning and no solving.

Table 1: full deployment cost on measured OGW2 data (real measurements, 66 pitch–catch paths, 194 intact calibration records), five-seed means.

MethodTraining (s)Recovery (s)Total (s)Peak GPU (GB)
Physical-only0.02.220.90.02
PASL-U300.43.4322.80.22
Data-NGO574.32.2595.10.15
Neumann NGO35.63.557.70.44
PINO-style4.32.325.20.02
Lunz-FAC281.23.3303.61.38
DeepONet66.40.187.00.09
FNO136.60.1158.30.10

This is the full deployment pipeline on real data: PASL-U uses 194 real intact records for offline training (about 5 min, 0.22 GB peak GPU), then performs source-feature recovery and damage classification on all real damaged records; the recovery stage over all 644 evaluation records takes 3.4 s (about 5 ms/record), and classification is a linear probe with negligible cost.

Table 2: training and deployment cost on the synthetic benchmark 21×21, 64 sensors.

MethodTraining (s)Total (s)Inference (ms/sample)Peak GPU (MiB)
Physical-only0.005.210.99546.6
PASL-U46.5853.141.0381033.0
Data-NGO15.3420.970.87658.3
Neumann NGO24.1533.4662.102866.8
Lunz-FAC191.77298.7127.700117.9
PINO-style2.3310.411.7366401.7
DeepONet209.83210.460.14753.1
FNO412.96413.590.159145.8

PASL-U's training is ~47 s and deployment inference ~1.04 ms/sample, nearly identical to the physics prior; the deployment overhead of the low-rank correction is negligible.

Table 3: training and deployment cost on synthetic guided waves.

MethodTraining (s)Inference (ms/sample)Peak GPU (GiB)
Physical-only0.00003.9386 ± 0.70750.0120
PASL-U54.04743.8822 ± 0.73570.0780
Data-NGO16.87164.5305 ± 0.26760.0313
Neumann NGO19.69794.1840 ± 0.89740.5746
Lunz-FAC116.180922.4842 ± 4.42730.1130
PINO-style25.40993.9716 ± 0.82480.0281
DeepONet2.25490.3821 ± 0.04780.0575
FNO2.97340.4396 ± 0.05380.1259

The reviewer also notes we only reported 21×21. We therefore scale the discretization further :

GridInterior DOFsSensorsTraining (s)Peak GPU (GB)
21×213612543.21.03
21×213616447.31.14
21×2136114456.31.49
25×2552925454.62.39
25×2552964470.42.50
25×25529144543.52.59
29×2972925639.84.38
29×2972964638.94.07
29×29729144665.54.01

Training cost generally increases with grid size and sensor count while staying within a single GPU (peak 1.03–4.38 GB), so the method is not limited to 21×21. The cost also scales with the number of calibration environments, not with the environment dimension itself

Param dimCalibration envsTraining (s)Peak GPU (GB)
1 (scalar OOD)947.31.14
2 (two-parameter)16246.81.94
3 (three-parameter)128873.05.25
4 (four-parameter)64643.46.89
The correction capacity is rank-K and independent of the environment dimension ded_e; only the Stage-B map grows as K⋅deK\cdot d_e, so the total is O(K⋅N+K⋅de)O(K\cdot N+K\cdot d_e). Training is a one-time offline cost; at deployment the operator is frozen and inference is about 1.04 ms/sample, essentially the same as the physics prior (0.995 ms/sample), so the low-rank correction adds no meaningful online cost for edge deployment.
Reviewer_6KBYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PASL-U, a physics-anchored operator learning framework for source recovery under environmental shifts. The method combines an environment-dependent physical propagation operator with bounded, learnable corrections. A two-stage training strategy first identifies shared correction modes and then learns an environment-to-code mapping for extrapolation. The corrected propagation operator is subsequently used for group-sparse source reconstruction. The authors also provide conditional error bounds connecting correction approximation, environmental extrapolation, and source recovery. Experiments on advection–diffusion systems and synthetic and measured guided-wave data demonstrate improved robustness under environmental shifts compared with several neural operator and physics-based baselines.

Critical strengths
  1. The proposed framework provides a reasonable approach to handling environmental shifts by preserving the environmental dependence of an approximate physical operator while learning bounded discrepancy corrections. Unlike direct inverse mapping approaches, PASL-U maintains source linearity and explicitly models how propagation changes with environmental parameters. The combination of operator correction and source reconstruction is practically meaningful.

  2. The experiments cover single-parameter and coupled-parameter extrapolation, synthetic guided-wave source recovery, and measured guided-wave damage recognition. PASL-U generally achieves lower propagation-operator errors and more stable source recovery under the tested environmental shifts. The component ablations also demonstrate the benefits of physical anchoring and environment-dependent correction prediction.

Critical weaknesses
  1. The ablation study in Table 6 shows that joint single-stage training achieves essentially the same performance as the proposed two-stage procedure. This raises questions about the necessity of the two-stage design beyond interpretability and error decomposition.

  2. The theoretical guarantees are conditional on operator invertibility, bounded perturbations, and Lipschitz assumptions. Their applicability to strongly varying or poorly conditioned physical systems remains unclear.

  3. The experiments primarily consider one- and two-dimensional environmental parameter spaces. Generalization to more complicated environmental changes remains insufficiently validated.

  4. The computational overhead of learning correction modes, performing two-stage optimization, and constructing corrected propagation operators is not sufficiently compared with the baselines.

Questions for authors
  1. Table 6 shows that joint single-stage training performs comparably to the proposed two-stage approach. What practical advantages justify the additional training separation? Does it improve stability, computational efficiency, or generalization in other settings?

  2. Could the authors clarify the proof of Proposition 2? Specifically, how does the Rademacher-complexity argument establish the supremum error bound in Eq. (22), rather than a bound on expected generalization error?

  3. Can the method handle higher-dimensional environmental parameters or strongly nonlinear environmental dependencies? How does the complexity of the correction representation scale with environmental dimensionality?

  4. In the synthetic guided-wave experiments, source NMSE increases significantly at higher temperatures. What causes this degradation, and how can the authors distinguish errors due to propagation extrapolation from those caused by inverse-problem ill-conditioning?

  5. The measured damage-recognition results have relatively large standard deviations, and performance approaches chance level in the highest temperature band. Can the authors provide additional statistical analysis or independent acquisition tests to support the claimed robustness?

  6. How do the total training time, inference cost, and memory requirements compare with those of the baseline approaches? Can the authors provide computational-budget-matched comparisons?

6 comments
AuthorsOpenReview ↗

**W1 / Q1 advantage of the two-stage strategy over joint single-stage training.**

The core advantage of the two-stage strategy over joint single-stage training is a drastic reduction in recalibration cost, not an improvement in accuracy; they are comparable in accuracy. Stage A learns the shared correction modality, which captures the systematic deviation of the operator along environment-independent dimensions. Once learned, this modality is frozen and reused directly for new environments without any retraining, incurring no recalibration cost. Stage B learns only a lightweight environment-to-code map (412 parameters) that converts environmental information into a low-dimensional code driving the correction modality to adapt to the current environment. At each recalibration, only these 412 parameters plus the per-environment free code (54 parameters), totaling 466, need to be refitted.

ComponentParameter countDuring recalibration
Stage A shared correction modality912,247reused, not relearned
Stage A per-environment free code54refitted
Stage B environment→code map412refitted
Environment-dependent total466refitted
By contrast, joint single-stage training couples the modality and the map together, so recalibration requires relearning the entire roughly 912,000-parameter modality and map from scratch, about 2000 times more parameters and nearly twice the runtime. Whenever the sensor array, temperature range, or field conditions change, the joint single-stage approach must rerun the full training pipeline, which is costly and hard to maintain in engineering deployment.
As shown in the table below, the two-stage scheme updates only 466 parameters at recalibration, taking around 22.6 s, whereas joint single-stage must update 912,659 parameters, taking around 44.6 s. The two-stage approach compresses the recalibration parameter count to 0.05% of the joint scheme and cuts the runtime nearly in half.
SchemeRecalibration contentParameter countTime (measured)
Two-stagerefit code + map only466~22.6s
Joint single-stagerelearn modality + map912,659~44.6s

The 22.6 s versus 44.6 s is the wall-clock saving on our small 21×21 benchmark, so the absolute numbers look modest only because the grid is small; the architectural point is what matters. The two stages decouple the expensive, environment-independent part from the cheap, environment-sensitive part: Stage A is the shared correction modality (the physical substrate), built once and reused unchanged, while only the 412-parameter environment-to-code map and the 54-parameter per-environment code are re-fitted at each deployment. In a real system such as fleet-scale wing monitoring the domain has tens of thousands of grid nodes and Stage A may be an operator distilled offline on a server from large multi-channel records; an edge node cannot backpropagate through a ~912k-parameter modality, but re-fitting 466 parameters is a milliseconds-scale fit that runs on the sensor itself. The value of two stages is therefore not the 22 s, but the separation of the shared physical substrate from the environment-sensitive code: the part that must change when the environment drifts is exactly the part that stays cheap to re-fit.

AuthorsOpenReview ↗

**W2 / Q2 Real-system validity of the assumptions, and the corrected Proposition 2.**

The reviewer asks whether the three premises: operator invertibility, bounded correction, and smooth environment dependence hold in real systems, and what happens for ill-conditioned or strongly varying systems. We give concrete examples and explain how the algorithm handles each case.

  1. Operator invertibility holds naturally, and we can actively preserve it. A diffusion–reaction operator with positive diffusion is positive-definite and invertible — a physical property, not an added assumption; the guided-wave operator is well-conditioned away from resonance. If an environment becomes ill-conditioned (large pp), we shrink the correction budget ρ\rho so θ=pβ<1\theta=p\beta < 1 is restored; the correction becomes more conservative and stays close to the physics prior, but never diverges. Ill-conditioning therefore does not invalidate the theory; it automatically shrinks the correction.

  2. Boundedness is enforced by construction. The correction is parameterized inside a box (tanh + code box + elementwise normalization), so it has an elementwise bound ∣ΔL(z)∣≤ρM⊙S|\Delta L(z)|\le\rho M\odot S at every environment, regardless of the system. This is a constructive guarantee, not an assumption.

  3. Smooth environment dependence is physical. Material stiffness, density, and wave speed vary smoothly and differentiably with temperature which is also the physical basis of temperature-polynomial background compensation in guided-wave monitoring. In advection–diffusion the coefficients are smooth functions of the environment parameters. This smoothness is estimated from calibration data and validated on held-out environments.

  • Strongly varying but smooth: a large smoothness constant only increases the slope of the extrapolation error (it grows linearly in distance); the method stays valid with a shorter effective range.
  • Truly non-smooth (phase transitions, fracture): this is outside the method's applicability, but because the correction is anchored at zero and bounded, the method degrades smoothly back to the physics prior instead of diverging.

Corrected Proposition 2. We thank the reviewer for identifying the gap: the Rademacher bound controls the average error, not the supremum over environments, and the original one-line Lipschitz transfer leaves a JJ-independent diameter term. We correct it with the standard covering argument below. Notation: E⊂Rde\mathcal{E}\subset\mathbb{R}^{d_e} compact, e1,…,eJe_1,\dots,e_J i.i.d. from PP; z^ϕ\hat z_\phi is LϕL_\phi-Lipschitz and z†z^\dagger is L†L_\dagger-Lipschitz; ℓ(e)=∣z^ϕ(e)−z†(e)∣2≤Mℓ\ell(e)=\\|\hat z_\phi(e)-z^\dagger(e)\\|_2\le M_\ell; RJ(Φ)\mathcal{R}_J(\Phi) is the empirical Rademacher complexity; L^J(ϕ)=1J∑j=1Jℓ(ej)\widehat L_J(\phi)=\frac{1}{J}\sum_{j=1}^{J}\ell(e_j).

Corrected statement. With probability at least 1−δ1-\delta, the average future code error obeys the standard Rademacher bound (Eq. 22a):

Ee∼P[ℓ(e)]≤L^J(ϕ)+2RJ(Φ)+Mℓlog⁡(1/δ)2J,\mathbb{E}_{e\sim P}[\ell(e)] \le \widehat L_J(\phi)+2\mathcal{R}_J(\Phi)+M_\ell\sqrt{\frac{\log(1/\delta)}{2J}},

and the worst-case code error obeys the covering bound (Eq. 22b):

sup⁡e∈Eℓ(e);≤;max⁡1≤j≤Jℓ(ej)+(Lϕ+L†),rJ,\sup_{e\in\mathcal E}\ell(e) \\;\le\\; \max_{1\le j\le J}\ell(e_j)+(L_\phi+L_\dagger)\\,r_J,

where rJ=sup⁡e∈Emin⁡j∣e−ej∣2r_J=\sup_{e\in\mathcal E}\min_j\\|e-e_j\\|_2 is the covering radius of the observed sample and rJ→0r_J\to 0 almost surely as J→∞J\to\infty (compact E\mathcal E, PP with full support). The ε\varepsilon-net form is the same statement: for any ε\varepsilon-net Eε\mathcal E_\varepsilon of size N(E,ε)≤(diam(E)/ε)de\mathcal N(\mathcal E,\varepsilon)\le(\mathrm{diam}(\mathcal E)/\varepsilon)^{d_e}, sup⁡eℓ(e)≤max⁡e~∈Eεℓ(e~)+(Lϕ+L†)ε\sup_e\ell(e)\le\max_{\tilde e\in\mathcal E_\varepsilon}\ell(\tilde e)+(L_\phi+L_\dagger)\varepsilon.

Proof sketch. Step 1 (average, unchanged): McDiarmid, symmetrization, and Talagrand's contraction lemma give Eq. (22a). Step 2 (covering transfer, the corrected step): for each ee let e(j)e_{(j)} be its nearest sample point; Lipschitz continuity gives ℓ(e)≤ℓ(e(j))+(Lϕ+L†)∣e−e(j)∣2≤max⁡jℓ(ej)+(Lϕ+L†)rJ\ell(e)\le\ell(e_{(j)})+(L_\phi+L_\dagger)\\|e-e_{(j)}\\|_2\le\max_j\ell(e_j)+(L_\phi+L_\dagger)r_J, and maximizing over ee gives Eq. (22b). The original proof instead used the coarse transfer sup⁡eℓ(e)≤E[ℓ]+(Lϕ+L†)diam(E)\sup_e\ell(e)\le\mathbb{E}[\ell]+(L_\phi+L_\dagger)\mathrm{diam}(\mathcal E), which is valid but leaves the JJ-independent constant diam(E)\mathrm{diam}(\mathcal E); the corrected step replaces it by the sample's own covering radius rJr_J, which vanishes as J→∞J\to\infty, with the covering number entering only through the ε\varepsilon-net bound. Choosing ε=ε(J)→0\varepsilon=\varepsilon(J)\to 0 therefore makes the worst-case excess over the average vanish, unlike the fixed diameter term.

This preserves the downstream use: the pointwise extrapolation bound follows by evaluating Eq. (22b) at the reference environment ebe_b and applying the Lipschitz relation Eq. (8). We will fix the statement and proof of Prop. 2 in the revision.

AuthorsOpenReview ↗

**W3 / Q3 — Higher-dimensional and nonlinear environments.**

The method's structure does not change with the environment dimension; therefore, when extended to three and four environment parameters, PASL-U still outperforms Physical-only, while the direct-inversion baselines (DeepONet/FNO) fail completely. Strongly nonlinear but smooth dependence is likewise covered, because the environment-to-code map is a neural network and the theory requires only Lipschitz smoothness. Discontinuous dependence (e.g., phase transitions) lies outside the guarantees, but the bounded, anchored correction then degrades gracefully back to the physics prior. The main experiments are one- and two-dimensional; we extended to three and four environment parameters with the same "physics prior + bounded correction + environment-code extrapolation" framework, changing only the environment dimension and the number of calibration environments (3D: 128; 4D: 64), with extrapolation taken at 1.2/1.4/1.6 as before. Table 1: 3-parameter and 4-parameter experiments

SettingPASL-UPhysical-onlyDeepONetFNO
3-parameter (diffusion, advection, reaction; calib [0.8,1.2]³, 128 envs)0.36340.3924~2.3–2.6~2.3–2.6
4-parameter (diffusion, x-adv, y-adv, reaction; calib [0.8,1.2]⁴, 64 envs)0.39370.4064~2.3–2.6~2.3–2.6

Results (single-source NMSE @30 dB, three-seed mean ± SD): in three parameters PASL-U overtakes Physical-only (0.3634 vs 0.3924), and in four parameters it does so again (0.3937 vs 0.4064), while the direct-inversion baselines (DeepONet/FNO) fall to 2.3–2.6 and fail. The four-parameter setting (diffusion, x-advection, y-advection, reaction; calibration [0.8,1.2]⁴) and the three-parameter setting (diffusion, advection, reaction; calibration [0.8,1.2]³) are reported as formal tables in the revision.

Why the 4D gain is small but that is the point: with only 64 calibration environments in a four-dimensional box the coverage is sparse, and the physical prior itself is already decent (0.4064) because the coefficients only move within [0.8,1.2]. In this sparse regime the bounded, anchored correction is deliberately conservative — the active budget shrink keeps the correction close to the prior instead of fitting the sparse samples aggressively — so PASL-U gains a small but reliable margin (0.3937 vs 0.4064) rather than chasing a larger gain and risking collapse. The direct-inversion baselines show what happens without that anchored budget: DeepONet and FNO fall to 2.3–2.6. A small, safe gain under sparse high-dimensional coverage is therefore evidence that the safety mechanism is doing its job, not a sign of a curse of dimensionality.

Why higher dimension works: the method's structure does not change with ded_e. The physics prior Lphys(e)L_{\mathrm{phys}}(e) is still analytically evaluable, the correction family is still rank-K and bounded, and what is extrapolated is still the small environment-to-code map. What grows with ded_e is only the Stage-B input dimension (linear, a few hundred parameters) and the number of calibration environments needed to cover the space (2D: 16, 3D: 128, 4D: 64). Higher-dimensional prediction is therefore the same mechanism running on a larger environment space, not a different method.

On complexity scaling: the admissible correction is rank-K, ΔL(z)=ΔL0+∑k=1KzkΔLk\Delta L(z)=\Delta L_0+\sum_{k=1}^{K} z_k\Delta L_k, with a KK-dimensional code box; its capacity (rank K) is independent of ded_e. Only the Stage-B map grows, as K⋅deK\cdot d_e (negligible versus the K⋅NK\cdot N of the correction modality), so the total parameter count is O(K⋅N+K⋅de)O(K\cdot N+K\cdot d_e) — the dominant term scales linearly with discretization and is independent of the environment dimension.

On strongly nonlinear dependence: the environment-to-code map is a neural network, so it can represent nonlinear dependencies, and the theory only requires Lipschitz smoothness, not linearity. Strongly nonlinear but smooth dependence is therefore covered; a larger Lipschitz constant Lϕ+L†L_\phi+L_\dagger simply shortens the improvement radius (Equation (11)). Discontinuous dependence (e.g., phase transitions) is outside the guarantees, but the bounded, anchored correction then degrades gracefully toward the physics prior. We will state this boundary explicitly in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.