acceptodds
Under review as a conference paper at ICLR 2027

Propagating Geometry Uncertainty from a Single Implicit Neural Representation through a Differentiable PDE Solver

Abstract

Implicit neural representations (INRs) such as neural signed distance functions are increasingly used as the geometry for physics simulations that solve partial differential equations (PDEs). However, in these simulations, the reconstructed geometry is usually treated as deterministic even though it is fitted to noisy data (such as point cloud). The effect of these geometric uncertainties on the solution or quantities of interest (QoIs) of a PDE is unclear. In this work, we propagate this uncertainty from the geometry (a trained INR) to PDE solutions. We freeze the INR's nonlinear features and treat only its final layer as free, which makes the boundary an explicit, low-dimensional function of a small set of deformation modes; both inference and propagation are then closed-form within this span. The noisy scan yields a Gaussian posterior over the mode coefficients, which a Shifted Boundary Method (SBM) solver propagates to PDE QoIs faster than Monte Carlo. We validate against two ground truths, exact random-field geometry and a controlled ground truth drawn from the posterior itself, to separate the accuracy of the posterior estimate from the expressive limits of the frozen representation. The result is close to ideal on the controlled test, and against exact geometry the posterior still accounts for most of the variability.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Implicit Neural Representations, Uncertainty Quantification, Low-Dimensional Bayesian Inference

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_t1oaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Geometry uncertainty of defining PDE domain is consdiered in this submission. Some metjhodology to propogate this uncertainty from one image / scan is provided and some empirical evaluation is provided.

Critical strengths

This work addresses a challenging problem - uncertainty in domain geometry and its proposagtion. Methodology is provided and empirical evaluation is given.

Critical weaknesses

The work does not address the issue of what the impact of neglecting this uncertainty actually happens to be - is it a primary cource or secondary source - this remains unanswered as far as i could tell. There is no theoretical analysis provided which for such a rpoblem and for the classes of problem considered would have been valuable - consider early work of Nobile and collaborators. the area of Stoichastic Geometry would have been useful to consider for this work.

Questions for authors

I have no questions for the authors

2 comments
AuthorsOpenReview ↗

geometric vs discretisation error, theory, and stochastic geometry

We thank the reviewer for the review and for the pointers to the work of Nobile and collaborators and to stochastic geometry. We address the three points in turn. The new experiments use the paper's setting (32 modes plus K = 10 discrepancy modes).

Is geometric uncertainty a primary or a secondary source of error?

In our setting it is the primary source, and we now quantify this. With scan noise of 1% of the shape's radius, the geometric uncertainty of the PDE solution is 0.5% at the centre and 3.0–4.4% next to the boundary, so the boundary noise is amplified where boundary-driven quantities are evaluated. This is 10–150 times the discretisation error of the solver we use:

ShapeGeometric QoI uncertainty / QoI (centre, wall)Discretisation error / QoI (centre, wall)Ratio (median over probes)
Circle0.54%, 4.4%0.001%, 0.031%130×
Flower0.52%, 3.0%0.055%, 0.013%10×
Rectangle0.56%, 3.5%0.006%, 0.044%150×

(Discretisation error estimated as the difference between our level-8 solution and a level-10 solution on the true shape.)

Neglecting the geometric uncertainty therefore means reporting the QoI with the dominant error source omitted. A deterministic reconstruction gives a single value with no interval. Our posterior gives an interval that contains the true solution at all 16 probes for every shape (§4.4). On a nonlinear problem, the Bratu equation −Δu = λeᵘ near its fold, the relative geometric uncertainty roughly doubles (to about 1.0% at the centre), so the effect grows when the solution is sensitive to the domain.

Theoretical analysis

We agree that the paper should engage with this theory more closely, and the revision will do so. Our linear propagation is the first-order perturbation approach that Harbrecht, Schneider & Schwab (2008) analyse for the moments of solutions of elliptic problems on randomly perturbed domains. Castrillón-Candás, Nobile & Tempone (2016) prove analytic regularity of such solutions with respect to the parameters of a random domain deformation; this regularity is what makes the QoI smooth in our mode coefficients, and it underlies both our linear approximation and the Monte Carlo or collocation alternatives. We cite both papers, but only in passing; the revision connects our method to them explicitly.

What our setting adds to this framework is the origin of the random domain. In these works the distribution of the deformation is prescribed. In ours it is inferred from a single scan through a trained network, and Appendix A supplies the link the theory needs: an implicit-function bound from last-layer perturbations to boundary displacement, with a remainder quadratic in the perturbation. The revision adds a short proposition bounding the error of the linear law in terms of the second derivative of the QoI with respect to the coefficients and the posterior covariance.

Our experiments behave as this theory predicts. When we inflate the posterior covariance by s², the linear variance stays within 8% of Monte Carlo up to s = 4, while the mean shift relative to the standard deviation grows roughly in proportion to s (0.02–0.21 at s = 1, 0.39–0.56 at s = 4, 1.05–1.44 at s = 16): a bias that is second order in the perturbation, measured against a standard deviation that is first order.

Stochastic geometry

Stochastic geometry studies random sets and random shapes with distributions specified a priori, for example through random closed sets or random field models of a boundary (Chiu et al., 2013). Our posterior is such a random domain, but one conditioned on data: it is supported on a low-dimensional family of shapes derived from one trained network and weighted by the measurement model of a scan. The revision will place our construction in this context, and will note that prior models from stochastic geometry could replace our Gaussian prior on the mode coefficients.

Changes in the revision

  • Related work: the connection to perturbation and regularity theory for random domains, and to stochastic geometry.
  • A proposition bounding the error of linear propagation, with the amplitude sweep as its empirical counterpart.
  • The comparison of geometric uncertainty with discretisation error.
Reviewer_t1oaOpenReview ↗

response

thank you for responding to the comments made at present it is difficult to move from current score given strength of response effect to contribution.

Reviewer_c5mMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose a computationally efficient framework for quantifying how geometric uncertainties, arising for instance from reconstructing geometries from noisy point clouds, propagate through PDE-based simulations and affect quantities of interest (QoIs). To achieve this, they train an Implicit Neural Representation (INR) of the geometry and introduce geometric perturbations by modifying only the network's final linear layer, while keeping the remaining parameters fixed. They then construct a low-dimensional basis of possible boundary deformations and infer a Gaussian distribution over the corresponding coefficients from the noisy observations. This uncertainty is subsequently propagated through the PDE using a first-order approximation and adjoint sensitivities, requiring only one forward and one adjoint solve instead of an expensive Monte Carlo procedure.

The authors evaluate their framework on three simple two-dimensional geometries (a circle, a flower, and a rectangle) governed by the Poisson equation. They compare their predictions against Monte Carlo simulations, demonstrating close agreement in the estimated uncertainty distributions, with computational speedups of approximately 300–350×. However, the agreement is strongest when the geometric variations are restricted to those representable by the chosen INR deformation basis.

Critical strengths
  1. The authors develop a mathematically sound framework that connects perturbations of an Implicit Neural Representation (INR) to physical boundary deformations and propagates the resulting geometric uncertainty to PDE quantities of interest.

  2. The proposed approach achieves substantial computational speedups (approximately 300–350×) compared to traditional Monte Carlo simulations, while maintaining good agreement with Monte Carlo uncertainty estimates within the assumed deformation model. This demonstrates the computational efficiency of the framework and its potential for uncertainty quantification in PDE-based simulations.

Critical weaknesses
  1. The geometric perturbations considered in this paper are restricted to Gaussian random-field perturbations of the radial boundary or deformations generated through the INR itself. However, the authors do not investigate whether these perturbations are representative of the geometric uncertainties encountered in practical engineering applications, particularly when reconstructing geometries from noisy or incomplete point clouds. In such settings, point-cloud sampling density may vary significantly with local curvature and geometric complexity, potentially introducing errors and reconstruction biases that are not captured by the proposed experiments. The absence of validation under more realistic geometric uncertainty conditions raises questions about the applicability of the framework to industrial scale problems.

  2. The proposed framework restricts geometric perturbations to variations of the INR's final linear layer, with a Gaussian prior imposed on the resulting deformation coefficients. However, this restricted representation already struggles to reproduce the independently generated geometric perturbations considered in the experiments. For instance, the retained deformation modes recover only 66% of the geometric displacement for the flower, compared to 90% for the circle, resulting in degraded uncertainty coverage for the QoI (83% versus the nominal 90%). Given that these limitations arise for relatively simple two-dimensional geometries and a linear Poisson equation, it is unclear whether the framework would remain sufficiently expressive and accurate for more complex three-dimensional geometries and nonlinear PDEs, where geometric perturbations and their effects on the solution may be substantially more challenging to capture.

Questions for authors
  1. The experiments consider geometric perturbations generated from a prescribed Gaussian random field perturbations of the radial boundary or from the INR's own deformation model. How representative are these perturbations of the geometric uncertainties encountered in practical applications, particularly with noisy, incomplete, or nonuniformly sampled point clouds?

  2. All PDE experiments are conducted on the linear Poisson equation with small geometric perturbations. How would the proposed first-order uncertainty propagation behave for nonlinear PDEs, such as the Navier–Stokes equations, where relatively small geometric changes can lead to significant nonlinear changes in the solution? Can the authors provide evidence that the computational advantages persist in such settings without compromising accuracy?

  3. The proposed framework restricts geometric perturbations to variations of the INR's final linear layer, while keeping the remaining network parameters fixed. Although this simplifies uncertainty propagation, it also substantially limits the space of admissible geometric deformations. Could the authors investigate, through an ablation study, whether allowing the entire INR (or several nonlinear layers) to vary would improve the representation of geometric uncertainty? Such an experiment would help quantify the trade-off between computational efficiency and representational expressiveness introduced by the proposed restriction.

5 comments
AuthorsOpenReview ↗

boundary vs QoI coverage, and realistic scans

(1/3)

We sincerely thank the reviewer for the time spent on our paper and for a detailed and constructive review. We are glad that the reviewer found the framework mathematically sound and the speedup over Monte Carlo useful. The three questions go to the heart of whether the method helps in practice: realistic scans, nonlinear PDEs, and the restriction to the last layer. They led us to run several new experiments, which we believe make the paper considerably stronger. Before answering them, we clarify one reading, because it bears on the soundness assessment. Unless stated otherwise, these use the paper's setting (32 modes plus K = 10 discrepancy modes); the all-layers ablation and the Navier–Stokes case use larger ranks, as stated there.

Boundary coverage and QoI coverage. The 83% cited in the second weakness is boundary coverage: the share of boundary rays inside the 5–95% band (Table 1). For the QoI, the linear interval contains the true solution u* at all 16 probes for all three shapes against exact geometry (§4.4, Table 3).

Weaknesses. Each is addressed in the answers below:

  • W1 (representativeness of the perturbations): we test realistic scans with non-uniform sampling, spatially varying noise and missing regions, each over 10 independent scans (Q1).
  • W2 (66% recovered displacement): this reflects our conservative cap of 32 modes; at r = 64, recovered displacement is 0.89 for the flower and 0.85 for the rectangle (Q3). Nonlinear PDEs: Q2.

Q1. How representative are these perturbations of the geometric uncertainties encountered in practical applications, particularly with noisy, incomplete, or nonuniformly sampled point clouds?

We thank the reviewer for raising this; it is the first question a practitioner would ask. The perturbations in the paper describe the measurement noise, but the uncertainty that matters in practice also includes how the reconstruction responds to the sampling. We therefore tested the full pipeline end to end on realistic scans. For each scenario we drew 10 independent scans, trained one network on each with the paper's protocol, built its posterior from that scan alone (r = 32, K = 10), and scored it against the true shape. Any reconstruction error caused by the sampling is therefore part of what the posterior must cover, and calibration is measured over independent scans.

The scenarios put the difficulty on one side of the shape, on the side of the PDE probes:

ScanBoundary coverage (circle / flower / rectangle)QoI coverage (circle / flower / rectangle)
Uniform (paper setting)0.87 / 0.83 / 0.790.89 / 0.81 / 0.86
Non-uniform: 8:1 sparser on one side0.88 / 0.84 / 0.800.82 / 0.83 / 0.77
Noise level 4:1 higher on one side0.87 / 0.82 / 0.780.82 / 0.80 / 0.84
30° arc with no points0.84 / 0.78 / 0.830.55 / 0.15 / 1.00
All three combined0.84 / 0.76 / 0.780.38 / – / 0.87

Non-uniform sampling and spatially varying noise are handled. Coverage stays at the level of uniform scans, and where the noise is higher the band widens by 1.5–1.6×.

Missing regions over curved boundary are not handled. Training interpolates the gap with a chord, which biases the reconstruction where no measurement constrains it, and the posterior cannot detect this bias. In the gap, coverage falls to 0.35 for the circle and 0.02–0.04 for the flower, and around it the error is 1.8–7.3× the band. For the flower, the chord cuts off the lobe tip and the outermost probe falls outside the reconstructed domain (1 of 10 scans with only the gap, all 10 with all three combined). Over a flat side, the chord is exact and coverage holds (0.86–0.94 in the gap for the rectangle).

We report these results as they are. They define the method's scope: scans that observe the whole boundary, or whose missing regions are flat. Treating occlusion would need a prior that reflects the absence of data in the gap, which we list as future work.

AuthorsOpenReview ↗

nonlinear PDEs, Bratu and Navier–Stokes

Q2. How would the proposed first-order uncertainty propagation behave for nonlinear PDEs, such as the Navier–Stokes equations? Can the authors provide evidence that the computational advantages persist without compromising accuracy?

This is a fair concern, and we agree that the paper needed evidence beyond the Poisson equation. The propagation is first order in the geometry, not in the PDE. Even for Poisson, the map from geometry to QoI is nonlinear (§3.4). What matters is how nonlinear that map is over the spread of the posterior. A nonlinear PDE changes how sensitive the QoI is, but not the structure of the method: linear propagation needs one nonlinear solve at the posterior mean and one adjoint solve per QoI with the Jacobian of the converged solve.

Evidence on a nonlinear PDE. We added the Bratu problem −Δu = λeᵘ, u = 0 on the boundary, a standard nonlinear elliptic benchmark with a fold bifurcation: solutions exist only below a critical λ_c that depends on the domain. We ran it at λ = 0.8 λ_c for each shape. There the solution is 62% above its linearised value, the relative QoI uncertainty is about twice that of Poisson, and the problem has the property the reviewer describes: as λ approaches λ_c, small changes of the domain cause large changes of the solution. Against posterior Monte Carlo (500 draws) and the spread over 500 exact-geometry curves:

ShapeLinear / MC stdLinear / exact-geometry stdu* in linear 5–95% intervalMax MC mean shift / stdEstimated speedup
Circle1.010.9516/160.14≈740×
Flower0.960.8914/160.17≈690×
Rectangle1.010.9416/160.10≈300×

No Monte Carlo draw or exact curve fell past its fold. The speedups are estimated from parallel wall-clock time rather than timed on a single core. The computational advantage grows with the cost of the forward solve: linear propagation pays for one nonlinear solve plus adjoints that reuse its Jacobian, while Monte Carlo pays for one nonlinear solve per draw.

Where it fails, and how to detect it. To locate the limit of linearisation, we inflated the posterior covariance by s² for s = 1 … 16 and compared with Monte Carlo. The variance stays within 7% of Monte Carlo up to s = 4. The mean degrades first, by a second-order shift of about half a standard deviation at s = 4. Breakdown coincides with posterior draws whose zero sets are no longer valid boundaries (11–21% of draws at s = 8 for the flower and rectangle), which can be counted without any PDE solve. This gives a practical rule: use linear propagation while realised draws stay valid and the predicted QoI uncertainty is below roughly 10–15% of the QoI, and switch to Monte Carlo otherwise.

Navier–Stokes. We added a steady Navier–Stokes case: a lid-driven cavity at Re = 100 with the circle as an immersed no-slip obstacle (unit cavity, lid speed 1, ν = 0.01, the circle of radius 0.25 at the centre). At this Reynolds number the flow is steady and viscous. The posterior is the one of our answer to Reviewer Ua63, Q1, where the rank rule gives r = 70. The QoIs are both velocity components at 16 probes, placed 0.05 and 0.15 from the obstacle. We compared linear propagation with 100 posterior Monte Carlo draws on two meshes:

MeshLinear / MC std (median [range] over 32 QoIs)Max MC mean shift / stdInvalid draws / failed solvesLinear std / speed (median)
Level 6 (3.4k nodes)0.96 [0.86, 1.04]0.290 / 02.8%
Level 7 (13.4k nodes)0.96 [0.85, 1.08]0.350 / 02.9%

With 100 draws, each Monte Carlo standard deviation is known to about ±7%, so these ratios are consistent with 1. Over the spread of the posterior, the linearisation is as accurate for Navier–Stokes as for Poisson and Bratu.

We have not implemented the adjoint of our Navier–Stokes solver. For this experiment we therefore computed the sensitivities by central finite differences along the 70 posterior directions (140 flow solves; the step was checked against smaller and larger steps, which agreed to within 0.04%). This establishes that the linearisation is accurate, but not that it is cheaper: here it took more solves than the 100 Monte Carlo draws. With an adjoint, the cost would be one steady solve plus one adjoint solve per QoI, as for Bratu. The adjoint of the steady Navier–Stokes equations is standard in shape optimisation, but we have not implemented it, so we make no speed claim for Navier–Stokes. Near a bifurcation, such as the onset of vortex shedding, or for unsteady and turbulent flows, a small change of geometry can move the solution to a different branch, and no local method captures this; the switching rule above would then direct the user to Monte Carlo. We state this scope in the revision.

AuthorsOpenReview ↗

letting all layers vary, and 3D

Q3. Could the authors investigate, through an ablation study, whether allowing the entire INR (or several nonlinear layers) to vary would improve the representation of geometric uncertainty?

We thank the reviewer for suggesting this ablation, which we ran. Letting every layer vary improves recovered displacement by only 0–4 points at the same number of modes, while raising the number of last-layer modes from 32 to 64 improves it by 6–23 points. The limit in the paper came from the number of modes, not from restricting the deformations to the last layer.

Construction. We took the boundary-displacement Jacobian with respect to all network parameters (17,025 for the circle and flower, 33,537 for the rectangle) instead of the last layer alone, and kept everything else as in the paper: the same arc-length-weighted SVD, prior, likelihood and posterior. A coefficient vector is realised as the network with all its parameters perturbed, so the edit acts through every nonlinear layer, and the PDE adjoint is propagated through every layer (verified against finite differences to about 10⁻⁶ relative).

Results (K = 0; last layer → all parameters):

ShaperRecovered displacementBoundary coverage (nominal 0.90)Variance ratiou* in QoI interval
Circle320.90 → 0.940.87 → 0.870.72 → 0.7816/16 → 16/16
Flower320.66 → 0.680.81 → 0.840.72 → 0.7714/16 → 16/16
Rectangle320.66 → 0.670.84 → 0.850.58 → 0.6216/16 → 16/16
Circle640.96 → 0.980.91 → 0.910.99 → 0.9916/16 → 16/16
Flower640.89 → 0.910.87 → 0.870.93 → 0.9616/16 → 16/16
Rectangle640.85 → 0.870.91 → 0.920.83 → 0.8816/16 → 16/16

In both variants linear propagation matches Monte Carlo (Linear / MC std 0.96–1.04 at r = 64, as defined in our answer to Q1 of Reviewer Ua63), and no realised posterior draw was invalid.

Why the last layer is enough here. With the hidden layers frozen, the field is exactly affine in the mode coefficients, so every posterior draw is an exactly realised network and the posterior is closed form. Letting all layers vary gives up this property for a gain of a few points, at a cost that grows with the number of parameters. The quantity that controls representation is the number of retained modes, and the paper's cap of 32 was conservative: with our revised rank rule, recovered displacement at r = 64 is 0.96 / 0.89 / 0.85.

Where this may differ. On detailed 3D surfaces the last-layer span is too small (see the next section), and there deeper layers or a wider final layer may be needed. We have not yet tested letting deeper layers vary in 3D, and we list it as the natural next step.

Three-dimensional geometries

The reviewer rightly asks whether the framework carries over to complex 3D geometries. We will add the 3D results to the updated manuscript. They cover a sphere, with 200 independent scans, and three scanned meshes (Armadillo, Dragon and Happy Buddha), with the Poisson problem solved on 3D octree meshes. In summary:

  • Linear propagation matches posterior Monte Carlo in 3D as in 2D (standard-deviation ratio 0.94–1.14 on the sphere and the Armadillo), with the adjoint sensitivities checked against finite differences.
  • On the sphere the posterior is informative: the last-layer span represents 95% of the boundary error, and over the 200 scans the 95% QoI intervals contain the true solution in 89% of cases.
  • On the detailed scanned meshes the span represents only 20–26% of the boundary error, so the boundary band under-covers. The QoI prediction is affected much less: on the Armadillo, the predicted uncertainty of the solution has the same size as the spread of independently retrained networks.

So in 3D the propagation step carries over directly, while the representation step depends on how detailed the surface is; this is the span question of our answer to Q3 at a larger scale.

Changes in the revision

  • A clearer distinction between boundary coverage (Table 1) and QoI coverage (Table 3).
  • New appendix sections: realistic scans (non-uniform sampling, spatially varying noise, occlusion); the Bratu problem; the Navier–Stokes lid-driven cavity, stating that its sensitivities were computed by finite differences because the Navier–Stokes adjoint is not yet implemented; the all-layers ablation; 3D results.
  • The rank rule that replaces the fixed cap of 32 modes, with main results reported at r = 64.
  • Discussion: the method's scope (occluded regions, bifurcations, unsteady and turbulent flows) and the open question of deeper layers in 3D.
Reviewer_Ua63OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extracts boundary deformation modes from a trained implicit neural representation, builds a geometric posterior from noisy scans, and then estimates output uncertainty through a differentiable PDE solver. For a single output quantity, linear propagation is about 300–350× faster than repeated Monte Carlo solves. Its sweet spot is small deformations near a reference shape; the frozen representation might miss things like sharp corners, and the experiments haven't yet covered 3D or nonlinear PDEs.

Critical strengths

1.Trains just one implicit neural representation, avoiding repeated retraining.

2.Connects boundary deformation, observation posterior, and PDE sensitivity.

3.Uses two reference geometries to separate representation error from propagation error.

4.Linear propagation of a single quantity of interest is ~300–350× faster than posterior Monte Carlo.

5.Reports ablation results on mode count, network architecture, and noise model.

Critical weaknesses

1.Frozen features and truncated modes struggle to represent sharp corners and similar shape changes.

2.Local linearization breaks down for large deformations or topological changes.

3.The posterior is sensitive to the noise model and small noise terms.

4.Experiments are limited to 2D shapes and a linear Poisson equation.

5.Coverage for external geometries doesn't always reach the nominal level.

Questions for authors

1.How do you choose and validate the number of modes for a new shape?

2.How do you estimate correlated noise covariances in real scans?

3.How do you quantify systematic shape deviations outside the captured modes?

4.When should you stop using linear propagation and switch to Monte Carlo?

5.How do accuracy and speed scale to 3D, nonlinear PDEs, or large numbers of quantities of interest?

5 comments
AuthorsOpenReview ↗

rank selection. Raising the rank above the paper's cap of 32 improves recovered displacement, coverage and QoI accuracy on all three shapes.

We thank the reviewer for the review. We answered each point with new experiments on the paper's networks. The main outcome is that our fixed cap of 32 modes was conservative: with the rank rule of Q1, recovered displacement and coverage improve on all three shapes.

Unless stated otherwise, the new experiments use the paper's setting (32 modes plus K = 10 discrepancy modes), so that each one changes a single factor relative to the submitted results. The rank sweep of Q1, the span diagnostics of Q3, the all-layers ablation and the Navier–Stokes case use other ranks, and say so where they are reported.

Weaknesses. Each is addressed in the answers below:

  • W1 (sharp corners): with the rank rule of Q1, the rectangle reaches 0.85 recovered displacement and 0.91 coverage.
  • W2 (large deformations): the variance stays accurate up to four times the paper's posterior spread, and a cheap check flags when to switch to Monte Carlo (Q4).
  • W3 (sensitivity to the noise model): a correlation length wrong by a factor of 2 costs at most 0.06 in coverage; the noise level matters more (Q2).
  • W4 (2D and linear PDE only): we add a nonlinear PDE and a steady Navier–Stokes case, and report preliminary 3D results (Q5).
  • W5 (coverage below nominal): 0.87–0.91 at r = 64, with u* inside the QoI interval at all probes (Q1).

Q1. How do you choose and validate the number of modes for a new shape?

We agree that the paper did not make this clear. In the revision the singular values set an upper limit on the rank: we keep at most the modes with s_k/s_1 ≥ 10⁻³. This is the threshold the paper already used, but we no longer cap the rank at 32. We then check this limit by drawing from the posterior and confirming that the zero set of each draw is still a valid closed boundary. Both steps use only the trained network and the scan, without ground truth, and the check needs only zero-set extraction, not PDE solves.

We validated this by sweeping r = 8 … 128 against the 500 exact-geometry curves. Invalid draws first appear where s_r/s_1 falls below about 10⁻³ (circle r = 80, flower r = 88–96, rectangle r = 96–128), and the parameter-update norm ‖Tc‖ jumps by an order of magnitude at the same point. Below that point, the geometry and the QoI uncertainty both get closer to the reference as the rank grows:

Shape (r = 32 → 64)Recovered displacementBoundary coverage (nominal 0.90)Variance ratiou* in QoI intervalLinear / MC stdLinear / exact-geometry std
Circle0.90 → 0.960.87 → 0.910.72 → 0.9916/16 → 16/160.99 → 0.950.90 → 0.94
Flower0.66 → 0.890.81 → 0.870.72 → 0.9314/16 → 16/160.91 → 0.990.88 → 0.94
Rectangle0.66 → 0.850.84 → 0.910.58 → 0.8316/16 → 16/160.98 → 0.990.92 → 0.95

The last two columns concern the PDE output u at the 16 probe points. Both compare the standard deviation of u predicted by linear propagation (one forward solve and one adjoint solve per probe) with a reference standard deviation, at each probe, and report the median over the 16 probes. A value of 1 means linear propagation predicts the reference spread exactly; below 1 it predicts too little. The references are:

  • MC: the spread of u over 300 posterior draws, each solved with the PDE. This tests only the linearisation, since both use the same posterior.
  • Exact geometry: the spread of u over the 500 exact-geometry curves, each solved with the PDE. This is the true geometric uncertainty, so it tests the posterior and the linearisation together.

At r = 64 linear propagation is within 5% of Monte Carlo and within 6% of the exact-geometry spread on all three shapes.

For the three shapes the rule gives an upper limit of r ≈ 70–88. Below it, recovered displacement peaks near r = 64–72 and declines slightly beyond, so we report r = 64 as the main setting in the revision. This choice within the limit was made against the reference curves; the limit itself uses no ground truth. The Navier–Stokes case (our answer to Reviewer c5mM, Q2) uses the limit directly, r = 70 for the circle. The cap of 32 came from the tanh networks of our ablations, which break at rank 15–17.

AuthorsOpenReview ↗

noise covariance from the sensor, and the effect of a misspecified covariance

Q2. How do you estimate correlated noise covariances in real scans?

We thank the reviewer for this question. It is central to applying the method to real scans, and it gives us the opportunity to explain this part of our work in more detail.

The covariance comes from the sensor, not from the scan being reconstructed. Range errors of laser scanners and depth cameras are spatially correlated: they vary with range, incidence angle and reflectance, and include calibration and registration errors (Nguyen et al., 2012), which is why correlated stochastic models are standard for terrestrial laser scanning (Wujanz et al., 2017; Kauker & Schwieger, 2017; Kermarrec et al., 2018). The variance and correlation length can be measured from repeated scans of a calibration artefact, or taken from the instrument's error model. Our method takes the covariance as an input (§3.1).

The scan itself cannot supply it. Maximising the marginal likelihood over the correlation length reached the edge of the search grid (8× the true value) for all 33 scans we tested, because the network is trained on that same scan and absorbs most of its noise (χ²/dof ≈ 0.003).

To measure the cost of a wrong covariance, we generated scans with the true model and built the posterior with a misspecified one, over 10 independent scans per shape. As noted at the start of this response, these experiments use the paper's setting (32 modes plus K = 10), not the larger rank of Q1. They measure how coverage changes when the covariance is wrong, relative to the correct covariance on the same chart, so the absolute values in the "correct" row are those of the paper's setting:

Likelihood usedBoundary coverage (circle / flower / rectangle)QoI coverage
correlation length ℓ/20.89 / 0.80 / 0.760.76–0.79
correct0.87 / 0.83 / 0.790.81–0.89
correlation length 2ℓ0.86 / 0.77 / 0.730.82–0.98
noise level σ/20.56 / 0.52 / 0.470.54–0.66
noise level 2σ1.00 / 1.00 / 0.991.00

A correlation length wrong by a factor of 2 costs at most 0.06 in coverage. The noise level matters more: underestimating it by 2× roughly halves coverage, while overestimating it is conservative. A noise level that varies 4:1 across the shape, with its profile known, leaves coverage unchanged (0.78–0.87, against 0.79–0.87 for uniform noise), and the band widens 1.5–1.6× where the noise is larger. The sensitivity to a vanishing nugget (Table 3) is the near-singularity of squared-exponential covariances on dense scans; a small nugget or a Matérn kernel avoids it.

AuthorsOpenReview ↗

deviations outside the captured modes, and when to switch to Monte Carlo

Q3. How do you quantify systematic shape deviations outside the captured modes?

This is an important question, because these deviations are exactly what a last-layer posterior cannot represent. We quantify them by projecting onto the span of the last layer. Any boundary deformation h splits into a part Ph inside the span, which the posterior can represent, and a remainder (I − P)h, which no change of the last layer can produce. The paper measured this against reference curves through the recovered displacement (Appendix B). For the rebuttal we added two diagnostics that do not need a reference geometry.

The first shows which spatial frequencies the span can represent. We projected the Laplace–Beltrami eigenfunctions of the boundary onto the span; on a closed curve these are Fourier modes in arc length. With r = 32 the span retains at least 0.93 of every mode up to frequency k = 12, but only 0.16–0.41 at k = 20. With r = 64 it retains at least 0.97 up to k = 24. The span therefore acts as a low-pass filter on the boundary, with a cut-off near k ≈ r/2.

The second combines the span with the scanner covariance C of Q2 to predict the recovered displacement without any reference curve, as 1 − (1 − tr(PC)/tr(C))^{1/2}. The prediction agreed with the measured value to within 0.00–0.10 at every rank we tested:

Shape (predicted / measured)r = 16r = 32r = 48r = 64
Circle0.63 / 0.640.90 / 0.900.98 / 0.981.00 / 0.99
Flower0.34 / 0.400.62 / 0.660.82 / 0.860.93 / 0.93
Rectangle0.35 / 0.450.61 / 0.670.76 / 0.780.87 / 0.88

These diagnostics account for the flower and rectangle results in the paper. The scan noise on these shapes has energy up to k ≈ 20–30, above the cut-off of r = 32. Because both diagnostics use only the network and the scanner model, a user can check in advance whether the span resolves the deviations the scanner produces.

If the span is too small, it can be enlarged. The simplest way is to raise the rank, as in Q1. Another is to train auxiliary output heads together with the main head, with targets given by the scan's SDF displaced by low-order Laplace–Beltrami modes. The shared features then have to represent these deformations, so they enter the span of the main head. This follows multi-task learning with shared representations (Caruana, 1997), shared-trunk multi-head networks (Lee et al., 2015; Osband et al., 2016) and shape-family decoders (Park et al., 2019; Dupont et al., 2022), and a construction of this kind for INRs is given in [Anonymous, 2026]. The targets come from the scan, so the method remains truth-free, and the frequency diagnostic above would show which modes have entered the span.

Some deviations are not described by the noise model, for example reconstruction bias where the scan has no data. The diagnostics do not predict them and the posterior does not capture them, as our occlusion experiment in the response to Reviewer c5mM shows. We state this as a limitation in the revision.

Q4. When should you stop using linear propagation and switch to Monte Carlo?

We make this cut off 2hen realised posterior draws stop being valid geometries, or when the predicted QoI uncertainty exceeds roughly 10–15% of the QoI. We measured this by keeping the paper posterior's mean and inflating its covariance, Σ → s²Σ (equivalent to a scan s times noisier), against 500 realised Monte Carlo draws per level:

sLinear / MC stdMax mean shift / stdInvalid realised geometriesLinear QoI std / QoI (wall probe)
1 (paper)0.93–1.020.02–0.2103–4%
20.93–0.980.22–0.2806–9%
40.95–1.000.39–0.56012–17%
80.96–1.100.77–0.950–21%24–35%
161.06–1.221.05–1.4420–96%48–70%

(Ranges over the three shapes. Linear / MC std is defined as in Q1: the standard deviation of u from linear propagation divided by that from the 500 Monte Carlo draws, median over the 16 probes.)

The variance matches Monte Carlo within 7% up to s = 4. The mean degrades first, through a second-order shift of about half a standard deviation at s = 4; if the mean is needed more accurately at that spread, Monte Carlo or a second-order correction should be added. The variance breaks down once invalid geometries appear. Counting invalid draws needs no PDE solves (1000 draws take seconds), and it also detects the breakdown in the rank sweep of Q1, where the Monte Carlo spread jumps to 7–70 times the linear one. A small pilot Monte Carlo is not a reliable check: with 30 draws, its estimate of the std ratio ranged from 0.72 to 1.23 where the true ratio was close to 1.

Comments on the paper

1 comment
Area_Chair_ett3OpenReview ↗

Please engage in the discussion

Authors: you have not responded to any raised question. Please reply to these comments and engage in the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.