acceptodds
Under review as a conference paper at ICLR 2027

Out-of-Distribution Generalization of Neural Operators via Active Geometry Synthesis

Abstract

Training neural operators for PDEs over complex geometries requires high-quality simulation datasets that are often designed by careful parameterization of a few specific shapes, yielding geometries that cover a range of physical dynamics close to those of the base shapes. Still, as seemingly small changes in the geometry can induce widely different physical dynamics, neural operators trained with such data generalize poorly to unseen geometric features that may arise in engineering and design optimization. In this work, we propose Active Geometry Synthesis (AGS), a novel geometric active learning method that improves the out-of-distribution (OOD) generalization of neural operators by synthesizing the geometries the operator is most uncertain about. Specifically, AGS equips a few base geometries with low-dimensional free-form deformations, and maximizes the predictive variance of the solution field, obtained by the Laplace approximation, with respect to the deformation control points using gradient ascent. As we demonstrate empirically, the Laplace predictive variance provides a surrogate for the operator errors over unseen shapes. As such, AGS samples diverse new geometries in regions of the geometry manifold where the physical dynamics are most challenging for the operator, expanding the coverage of the training dataset over the geometry manifold, and significantly improving the OOD generalization compared to i.i.d. uniform sampling. Meanwhile, by avoiding the manual collection and augmentation of geometries, AGS does not restrict the neural operator to a predetermined set of shapes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning for Physics, Neural Operators, Geometric Domains, Active Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qAE9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Active Geometry Synthesis (AGS) is an active learning framework designed to improve the out-of-distribution generalization of neural operators by synthesizing complex PDE geometries that maximize model uncertainty. To synthesize valid new shapes, AGS parameterizes geometries using differentiable free-form deformations and converts a trained deterministic neural operator into a function-valued Gaussian process (GP) via Laplace approximation.

The primary technical contributions include formulating the GP's predictive variance as a computable surrogate for worst-case error across the geometry manifold, alongside kernel conditioning and trust-region regularization to enforce diversity during batch synthesis. Methodologically, this provides the first synthesis-based active learning framework explicitly designed for complex, continuous geometric distributions in operator learning rather than scalar parameters or initial conditions.

Validated on 2D Euler transonic airfoil flow and bistable reaction-diffusion systems, AGS achieves up to a 40% reduction in volumetric relative L2L^2 error and a 50% to 70% reduction in drag coefficient error on unseen geometric families compared to uniform i.i.d. sampling.

Critical strengths
  1. Principled Active Synthesis Framework for High-Dimensional Geometric Domains

Existing active learning techniques for neural operators are largely restricted to scalar PDE parameters or initial conditions. AGS bridges this gap by unifying Laplace-approximated neural operators (function-valued Gaussian processes) with differentiable Free-Form Deformations (FFD). Differentiating the GP predictive variance directly with respect to FFD control points enables gradient-based optimization to synthesize challenging shapes.

  1. Significant Out-of-Distribution (OOD) Generalization Gains on Complex Dynamics

On challenging benchmarks where subtle geometry changes drastically alter physical behaviors (e.g., shock wave formation in 2D transonic Euler flow and ignition thresholds in bistable reaction-diffusion systems), AGS achieves up to a 40% reduction in volumetric relative L2L^2 error and a 50% to 70% reduction in drag coefficient (CDC_D) error on unseen shape families.

Critical weaknesses
  1. Breakdown of the Uncertainty Surrogate Far OOD and Heavy Reliance on Heuristics

The core theoretical foundation relies on the Laplace predictive variance v(a)v(a) acting as a reliable upper-bound surrogate for the operator's worst-case error. However, the authors acknowledge that local linearization via the Laplace approximation becomes severely overconfident far from the training distribution, causing the correlation between predictive standard deviation and actual prediction error to saturate and break down on distant geometries. To prevent gradient ascent from driving geometries into degenerate or overconfident regimes, the method relies on regularizing heuristics, such as a tuned trust-region penalty μ\mu, domain-specific shape filters, and manually replacing invalid/rejected synthesized samples with uniform draws (e.g., 74 samples rejected in the reaction-diffusion experiment).

  1. Systematic Degradation of In-Distribution Performance

Across both the Airfoil and Reaction-Diffusion benchmarks, training with AGS consistently leads to worse in-distribution performance compared to standard i.i.d. uniform sampling and competing synthesis baselines like PaPQS. For example, on the NACA 0012 in-distribution dataset with Ntr=2048N_{tr} = 2048, AGS yields a volumetric relative L2L^2 error of 0.047 compared to 0.037 for uniform sampling and 0.035 for PaPQS.

Questions for authors

N.A.

3 comments

We thank the reviewer for the recognition of the strengths of our works and the careful review. Below, we address the raised Weaknesses (W). In addition, we will soon submit a revised manuscript that incorporates the new results and discussions.

W1:

We agree with the diagnosis, and we would like to clarify how AGS is designed around it. The linearized Laplace approximation is a local approximation of the posterior around the MAP weights θ⋆\theta^\star, so its variance is informative only where the operator error is approximately captured by the linearized operator, i.e. near the training data. Overconfidence far from the training data is a known limitation of approximate posteriors for neural operators, and Figure 3b measures it for our setting, where the variance tracks errors with Spearman's ρ=0.81\rho = 0.81 and saturates for the two farthest families. For this reason, Eq. 8 serves only as a motivation (Remark 3), and AGS does not require the variance to be calibrated far from the data. Instead, AGS uses the variance as a local ranking signal. Each round synthesizes geometries near the current dataset, where the variance is informative, then retrains the operator and recomputes the posterior, so the region where the variance is reliable expands with the dataset. Figure 3a shows the outcome, as the distant OOD families end up much closer to the AGS dataset than to the Uniform and PaPQS datasets.

The trust region follows from the same reasoning, in the spirit of trust-region optimization, where a local model is trusted only within a radius. Our new ablation on Airfoil (in-loop models at 512 labels, see our response to Reviewer yF5J for details) shows that μ\mu controls a trade-off between in-distribution and OOD accuracy:

rel-L2L^2 (512 labels)In-dist.E387FX 63-137MH 113
μ=0\mu = 00.0760.1150.1630.181
μ=0.1\mu = 0.10.0740.1300.1940.197
μ=0.3\mu = 0.3 (paper)0.0690.1330.2120.216
μ=1.0\mu = 1.00.0710.1580.2450.269

A strong trust region (μ=1\mu = 1) loses most of the OOD gains. Without it (μ=0\mu = 0), the OOD errors are lowest after three rounds, but the designs drift towards the FFD box bound (43% of the design coordinates in the third round, against 11% for μ=0.3\mu = 0.3), as the ascent targets unreachable designs outside the box. Such saturated designs can harm the diversity over more rounds, even with kernel conditioning. We note that the best μ\mu may change with more acquisition rounds.

Regarding the shape filters and replacements, the validity checks are a property of the geometric design space. They are applied identically to Uniform, PaPQS, and AGS samples, and both the checks and the replacements are automatic, with no manual intervention per sample. For Airfoil, the thickness constraint is built into the projection step of the gradient ascent, and no synthesized geometry was rejected. In RD, the 74 of 448 rejected geometries all failed the same check (an over-diffused source). Our new RD ablation (see our response to Reviewer yF5J) shows that the trust region itself halves these rejections at μ=0.3\mu = 0.3. Of 192 synthesized geometries, 80 failed this check without a trust region (μ=0\mu = 0), 41 with μ=0.3\mu = 0.3, and 1 with μ=1\mu = 1. Building this check into the projection in the same way would remove the need for replacements, which we leave for future work.

W2:

We understand the reviewer's concern regarding the in-distribution performance. As we acknowledged in Section 5 and correctly pointed out by the reviewer, AGS may deteriorate the ID performance compared to Uniform sampling, which we observe in Airfoil and the new elastoplastic benchmark (see our response to reviewer yF5J). In RD, however, AGS achieves the lowest in-distribution error at N=512N = 512 (0.020, against 0.026 for Uniform and 0.022 for PaPQS, Table 6). Furthermore, we would like to highlight that, for both Airfoil and RD, AGS consistently improves the ID performance as well as OOD generalization with increasing label budget (see Fig. 2). On the other hand, adding uniform samples does not always improve the OOD generalization (e.g. Fig.2, MH113 airfoil and Ellipse RD). As such, AGS brings about a trade-off, while its errors consistently decrease on almost all evaluation sets with increasing budget. We note that when the target geometric distribution can be exactly predetermined and restricted, such trade-off may not be desirable. But when the target geometric settings are flexible and broad, as in many engineering and shape optimization problems, AGS offers a more reliable performance that consistently improves with additional computational and label budget.

Thank the author for the clarification. 1. I think the argument of gradually pulling OOD closer to training manifold is convincing; 2. The trade-off between in-distribution and OOD is clarified, and therefore ID performance should not be understood as limitation alone.

I have no further concern. Therefore, I keep my score and recommend acceptance.

Dear Reviewer qAE9,

Thank you for your response and for recommending acceptance. We are glad that our clarifications addressed your concerns. We have uploaded the revised manuscript (changes in blue), which incorporates the points from our discussion.

  • W1, uncertainty surrogate and heuristics: The ablation of the trust region, including the rejected geometries for each trust-region weight, is in the new Appendix F. The role of the trust-region weight is also added to the limitations in Section 5, and the validity checks are discussed in the new Appendix B.6.
  • W2, in-distribution performance: The trade-off between in-distribution and OOD accuracy is clarified in the closing paragraph of Section 4.2 and in Section 5.

If you have any further questions or concerns, we would be happy to address them.

Best regards,

The authors

Reviewer_oek5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Training neural operators over complex geometries requires datasets, which are usually based on a few specific shapes. Small changes in geometry can lead to different physical dynamics. The authors propose AGS to handle this problem. AGS uses Laplace approximation to estimate uncertainty in the solution field and uses the variance of the field as an acquisition function for new training samples.

Critical strengths
  1. AGS improves performance on OOD geometries.
  2. AGS can synthesize good geometries to capture the physical dynamics on new geometries.
  3. AGS empirically show that predictive uncertainty is related to prediction error near the training distribution.
Critical weaknesses
  1. AGS reduces ID performance in the Airfoil experiment.
  2. AGS requires additional optimization for acquisition.
  3. The authors point out that uncertainty is related to prediction error. But errors may come from a coarse mesh, not enough training, or a model architecture that has not been optimized.
Questions for authors
  1. How is the proposed method related to A-optimal or D-optimal experimental design? It would be interesting to understand when these methods choose similar or different samples.

  2. The current method uses FFD to generate new geometries. Can it work well with more complicated geometries or shapes that are difficult to represent with FFD?

  3. Can uniform sampling achieve similar performance with the same total computational cost by increasing the number of training samples or training for longer?

3 comments

We thank the reviewer for the careful review and the insightful questions. Below, we address the raised Question (Q) and Weaknesses (W). In addition, we will soon submit a revised manuscript that incorporates the new results and discussions.

Q1:

Thank you for this question. Under the linearized Laplace approximation, the neural operator is a linear-Gaussian model in its active weights, so the classical design criteria apply conceptually. A- and D-optimality summarize the posterior uncertainty of the weights by its average variance and its volume, while V-optimality, the prediction-space counterpart of A-optimality, averages the predictive variance over a reference distribution of inputs and thus favours representative inputs. AGS is most directly related to G-optimality, which minimizes the worst-case predictive variance (Eq. 8) and favours inputs where the variance is largest, typically towards the edges of the design space, consistent with the trade-off between in-distribution and OOD accuracy we observe. Repeatedly labelling the geometry with the largest variance mirrors the Wynn-Fedorov sequential algorithm for D-optimal design, and for a fixed linear model and continuous designs, the Kiefer-Wolfowitz theorem shows that the G- and D-optimal designs coincide. Since AGS retrains the operator every round, we see this connection as an interpretation rather than a guarantee. Our batch conditioning (Eq. 11) is similar to the pool-based MaxDet [2], which Holzmüller et al. relate to D-optimal design. The two coincide for a scalar output and noise-free observations, whereas for the solution field, Eq. 11 uses the trace rather than the determinant of the conditioned covariance block.

We added MaxDet and LCMD [1,2] as pool-based baselines, both implemented with the bmdal_reg package [2] using its linear last-layer kernel, selecting from a pool of 15K random FFD designs with the same Transolver and training protocol (mean over three seeds, rel-L2L^2):

LabelsMethodIn-dist.RAE 2822E387FX 63-137MH 113
512Uniform0.0650.1060.1560.2690.262
LCMD0.0670.1040.1470.2180.206
MaxDet0.0670.1060.1240.2060.208
AGS0.0780.1150.1300.1950.204
2048Uniform0.0370.0790.1300.2440.246
LCMD0.0360.0770.1230.2280.225
MaxDet0.0390.0790.1420.2630.272
AGS0.0470.0880.1020.1470.148

At small budgets, MaxDet and AGS perform best on the OOD families (except for RAE 2822), consistent with their shared greedy rule. As the budget grows, MaxDet loses its OOD advantage, while AGS keeps improving. We attribute this difference mainly to selection versus synthesis, as MaxDet is restricted to random FFD candidates, whereas AGS optimizes the geometry directly.

[1] Musekamp, Daniel, et al. "Active learning for neural PDE solvers." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Holzmüller, David, et al. "A framework and benchmark for deep batch active learning for regression." Journal of Machine Learning Research 24.164 (2023): 1-81.

Q2:

In principle, FFD deforms the space in which a geometry is embedded rather than the geometry itself, so it applies to many discretized representations, including point clouds and surface or volume meshes, in both 2D and 3D, and its cost does not depend on the complexity of the embedded shape (see [1] for an example in aerospace shape optimization). However, as we acknowledged in Section 5, the flexibility of FFD comes without a prior over valid or plausible shapes. For a complex base geometry, perturbing the control points can easily produce self-intersecting or physically implausible shapes, and keeping γ(s)(Z(s))\gamma^{(s)}\big(\mathcal{Z}^{(s)}\big) meaningful therefore requires problem-specific constraints. The constraints we used in our experiments are described in Appendix C. FFD also preserves the topology of the base shape, so geometries that differ from every base shape in topology, such as a shape with an added hole, are not reachable.

A natural direction for future work is to replace FFD with a parameterization that carries a learned prior, such as the latent space of a generative shape model or a Signed Distance Function (SDF) decoder (DeepSDF [2]). Plausibility and topological changes would then be handled by the decoder, and AGS would operate on its latent code in place of zz.

[1] Kenway, Gaetan, Graeme Kennedy, and Joaquim RRA Martins. "A CAD-free approach to high-fidelity aerostructural optimization." 13th AIAA/ISSMO multidisciplinary analysis optimization conference. 2010.

[2] Park, Jeong Joon, et al. "Deepsdf: Learning continuous signed distance functions for shape representation." 2019 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2019.

W3 and Q3:

To address the reviewer's concerns, we trained additional Uniform Airfoil models to test each alternative explanation. In the submission, all Airfoil models were trained on up to 2048 samples for 500 epochs with L=14L = 14K mesh nodes per sample (Reaction-Diffusion uses the full solver grid). The new models below use fewer or more nodes (LL = 6K, 20K), longer training (1000 epochs), and 2×2\times and 4×4\times more labels. All models are Transolver with one random seed, so the values differ slightly from the three-seed means of Table 1. We report both the relative errors and the Spearman correlation between the predictive STD and the errors (similar to Figure 3b) below. For brevity, we only report two OOD families in the table. The same pattern holds for FX 63-137, while on RAE 2822, the family closest to the training distribution, AGS remains slightly behind, as in Table 1.

rel-L2L^2# LabelsIn-dist.E387MH 113Spearman ρ\rho
Uniform20480.0380.1600.2720.81
Uniform, LL = 6K / 20K20480.034 / 0.0390.169 / 0.1540.290 / 0.2740.84 / 0.84
Uniform, 1000 epochs20480.0340.1370.2810.80
Uniform, 2×2\times / 4×4\times labels4096 / 81920.028 / 0.0220.136 / 0.1070.231 / 0.2560.84 / 0.80
AGS (ours)20480.0470.0980.139–

We agree that the Laplace variance captures mainly the data-related (epistemic) error. The table shows that the node count, training length, and more labels all improve or preserve in-distribution accuracy but leave the OOD gap largely open, so this gap is mainly a coverage gap, which is what AGS targets.

The last column is the Spearman correlation between the per-sample predictive STD and the per-sample error, pooled over the in-distribution set and all four OOD families as in Figure 3b. It stays between 0.80 and 0.84 for every model, in line with the 0.81 that Figure 3b reports for the Uniform model with 2048 labels. As in Figure 3b, the correlation within each of the two farthest families (FX 63-137 and MH 113) is weaker (ρ≤0.42\rho \le 0.42 for every model). The correlation also holds for a different architecture: the Uniform GNOT model of Table 7 (1024 labels), whose OOD errors are similarly high, gives ρ=0.81\rho = 0.81. The relationship between the variance and the error is therefore not an artifact of the mesh, the training length, the number of labels, or the architecture.

Q3. Generating the 8192 Uniform samples took about 47 hours on 30 CPU cores (different machine than in Table 4), comparable to the full 15-round AGS run (about 50 hours), excluding the final model training in both cases. At this matched computational budget, Uniform with 4×4\times labels nearly matches AGS on E387 but plateaus on MH 113 (and FX 63-137) well above AGS.

W1:

We understand the reviewer's concern regarding the in-distribution performance. As we acknowledged in Section 5 and correctly pointed out by the reviewer, AGS may deteriorate the ID performance compared to Uniform sampling, which we observe in Airfoil and the new elastoplastic benchmark (see our response to reviewer yF5J). In RD, however, AGS achieves the lowest in-distribution error at N=512N = 512 (0.020, against 0.026 for Uniform and 0.022 for PaPQS, Table 6). Furthermore, we would like to highlight that, for both Airfoil and RD, AGS consistently improves the ID performance as well as OOD generalization with increasing label budget (see Fig. 2). On the other hand, adding uniform samples does not always improve the OOD generalization (e.g. Fig.2, MH113 airfoil and Ellipse RD). As such, AGS brings about a trade-off, while its errors consistently decrease on almost all evaluation sets with increasing budget. We note that when the target geometric distribution can be exactly predetermined and restricted, such a trade-off may not be desirable. But when the target geometric settings are flexible and broad, as in many engineering and shape optimization problems, AGS offers a more reliable performance that consistently improves with additional computational and label budget.

W2:

We reported the matched-budget results in our response to W3 and Q3. As reported, at comparable computational cost, uniform samples plateau at some of the OOD families, while AGS consistently improves with increasing budget. We acknowledge that the computational cost of AGS is not negligible (AGS times are reported in Appendix D.2). However, in engineering and industrial applications where numerical simulations are expensive and can take days to complete, AGS provides a more reliable generalization at a reasonable computational cost.

Dear Reviewer oek5,

Thank you again for your careful review and insightful questions. We have uploaded the revised manuscript (changes in blue), and we would like to point you to the parts that address your questions and concerns.

  • Q1, pool-based baselines: The comparison with the pool-based MaxDet and LCMD methods is reported in Table 5 and Appendix E.5 (Figure 5).
  • Q2, geometry parameterization: The choice of FFD, its limitations, and alternative parameterizations are discussed in Section 5 and the new Appendix B.6.
  • W3 and Q3, sources of error and matched compute: The additional Uniform models (different mesh resolutions, longer training, and 2× and 4× more labels), the matched compute comparison, and the Spearman correlations are in the new Appendix E.6 (Table 9, Figure 6).
  • W1, in-distribution performance: The trade-off between in-distribution and OOD accuracy is clarified in the closing paragraph of Section 4.2 and in Section 5.
  • W2, acquisition cost: The matched compute comparison in Appendix E.6 also covers the additional cost of the acquisition.

We hope that the new results and discussions address your concerns, and we would greatly appreciate it if you could consider them in your final assessment. If you have any further questions or concerns, we would be happy to address them.

Best regards,

The authors

Reviewer_yF5JOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Active Geometry Synthesis (AGS), an active learning framework for improving the generalization of neural operators to unseen geometric distributions. AGS uses Laplace-based predictive uncertainty to optimize free-form deformation parameters and synthesize informative geometries. Kernel conditioning promotes batch diversity, while a trust-region penalty limits excessive drift from existing samples. The generated geometries are labeled by numerical PDE solvers and used to retrain the operator. Experiments on airfoil flow and reaction-diffusion problems demonstrate improvements on several out-of-distribution (OOD) test sets.

Critical strengths

Well-motivated problem. The paper addresses an important limitation of neural operators: small geometric changes can induce substantial changes in physical behavior, making generalization to unseen shapes challenging.

Meaningful empirical improvements. AGS achieves substantial gains on difficult OOD airfoil families. For example, it reduces the field error on MH 113 from 0.243 to 0.148 compared with PaPQS under the same label budget. Reaction-diffusion results and additional GNOT experiments provide further evidence of its usefulness.

Controlled experimental comparisons. The methods share the same label budgets and reachable geometry space, and model selection uses in-distribution validation data. These choices help attribute the observed improvements to the data acquisition strategy.

Critical weaknesses

Weaknesses

  1. Limited validation across industrial design applications. The experiments on two-dimensional airfoils and reaction-diffusion provide limited evidence for generalization to broader industrial design problems. The authors should evaluate AGS on additional industrially relevant datasets and conduct ablation studies of kernel conditioning for batch diversity and trust-region regularization for local geometric constraints (§3.2–3.3). These experiments are needed to establish whether the benefits of both components remain consistent across different datasets and geometric settings.

  2. Unclear manual effort required for solver compatibility. Does configuring the deformation mapping γ\gamma and its admissible variables zz require substantial manual design to ensure that the generated geometries can be discretized and solved reliably? A valid geometric deformation does not necessarily guarantee a suitable mesh or successful numerical solution. Although Appendix C describes task-specific validity checks, the effort required to adapt these constraints to new geometries and solvers remains unclear. The authors should clarify this adaptation process and provide evidence of its practical feasibility in industrial applications.

The paper addresses a problem of clear importance to industrial design. I would be willing to raise my score if the authors clarify these concerns and provide the corresponding experimental evidence.

Questions for authors

See weakness.

3 comments

We thank the reviewer for the constructive feedback. Below, we address the raised Weaknesses (W). In addition, we will soon submit a revised manuscript that incorporates the new results and discussions.

W1:

To address the reviewer's concerns and suggestions, we added a new benchmark, ablations, and pool-based AL baselines.

New benchmark: Elastoplastic plate

Plates with cutouts are a standard problem in structural design, such as access holes in load-bearing panels. We model a unit-square plate with a single interior cutout under plane-strain J2 plasticity with isotropic hardening, loaded by a uniform tension close to its limit load. Near this load, small changes in the cutout decide whether yielding stays local or spreads across the plate to its free edges. The operator maps the cutout boundary to the stress field and the equivalent plastic strain. Training cutouts are a superellipse deformed by FFD. OOD sets use unseen circle, square and five-pointed star cutouts, deformed by uniform random perturbations. Labels are computed with the finite-element library FElupe. We trained AGS (μ=0.3\mu = 0.3) starting from 128 uniform samples for Nal=3N_{al}=3 rounds with an acquisition batch size of Nb=128N_{b}=128, resulting in 512 total samples. Similar to RD, 70 synthesized cutouts that failed the validity checks (mainly self-intersections) were replaced with uniform samples, and all models are retrained with three seeds.

Stress rel-L2L^2MethodIn-dist.CircleStarSquare
256 labelsUniform0.0690.1380.2540.256
AGS0.0840.1280.2550.218
512 labelsUniform0.0460.1200.2260.243
AGS0.0540.0860.2070.170

Similar to Airfoil, AGS trades some in-distribution accuracy for OOD accuracy, and its OOD gains grow with the label budget. At 512 labels, AGS reduces the stress error by 28% on circles, 8% on stars, and 30% on squares, and is more accurate than Uniform on over 80% of the circle and square samples at every seed. For star cutouts, the gain is smaller, which may reflect the limited reach of FFD deformations of the rounded superellipse towards the five sharp spikes of the star.

Ablation: Kernel Conditioning and Trust Region

We ablate kernel conditioning and the trust region (TR) by running AGS without both (variance maximization only, Eq. 9), with conditioning but without the TR (μ=0\mu = 0), and with conditioning and μ∈{0.1,0.3,1}\mu \in \{0.1, 0.3, 1\}. Due to limited compute during the rebuttal period, we run 3 acquisition rounds on Airfoil (512 labels) and RD from the same initial model and seeds, and report the in-loop models that AGS uses for acquisition. These are single-seed models, trained and evaluated at L=6L = 6K nodes for Airfoil, so the numbers are not comparable with Table 1.

Airfoil rel-L2L^2 (512 labels)Cond.μ\muIn-dist.E387FX 63137MH 113
Variance only (Eq. 9)✗00.0830.1480.2460.280
+ conditioning✓00.0760.1150.1630.181
AGS✓0.10.0740.1300.1940.197
AGS (paper)✓0.30.0690.1330.2120.216
AGS✓1.00.0710.1580.2450.269

On Airfoil, kernel conditioning substantially reduces the OOD error and also improves the in-distribution error. The TR controls a trade-off between in-distribution and OOD accuracy. A strong TR (μ=1\mu = 1) restricts exploration and loses most of the OOD gains, while removing the TR gives the lowest OOD errors at this budget but a higher in-distribution error, and pushes many designs to the FFD box bound (43% of the FFD design coordinates on the bound in the third round, against 11% for μ=0.3\mu = 0.3). This is the saturation discussed in Section 3.3, where the ascent targets designs outside the FFD box, which can harm diversity over more rounds. RD behaves differently. There, the TR reduces the invalid geometries that occur without it (halved at μ=0.3\mu = 0.3), and since rejected geometries were not replaced in this ablation, it also yields more valid labels. In RD, μ=0.3\mu = 0.3 outperforms both μ=0\mu = 0 and μ=0.1\mu = 0.1, μ=1\mu = 1 performs comparably, and conditioning alone brings no improvement. Overall, effective conditioning is the main source of the OOD gains on Airfoil, while the TR keeps the designs valid and away from the box bound. We selected μ=0.3\mu = 0.3 based on the in-distribution validation error while avoiding higher values that prevent manifold exploration, and we note that the best μ\mu may change with more acquisition rounds.

W1 (Cont'd)

Pool-based baselines

We have added new pool-based baselines for Airfoil. For brevity, we briefly summarize the results here and invite the reviewer to see our response to Reviewer oek5 if interested. We generated a pool of 15K uniform deformations of NACA 0012, and trained with the pool-based LCMD algorithm [2], one of the best performing methods of AL4PDE [1], and with MaxDet [2], a pool-based counterpart of AGS that similarly ranks candidates by maximizing the variance. In summary, LCMD improves the OOD errors at the smallest budget, but performs similarly to PaPQS and Uniform at larger budgets, slightly improving the in-distribution performance at some budgets, while MaxDet significantly improves the OOD generalization at low budgets, outperforming AGS on RAE 2822 and E387 at the N=512N=512 label budget. As the budget grows, however, MaxDet fails to maintain the OOD improvements and falls behind LCMD and Uniform.

[1] Musekamp, Daniel, et al. "Active learning for neural PDE solvers." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Holzmüller, David, et al. "A framework and benchmark for deep batch active learning for regression." Journal of Machine Learning Research 24.164 (2023): 1-81.

W2:

In AGS, the deformation map γ(s)\gamma^{(s)} and its admissible set Z(s)\mathcal{Z}^{(s)} define the reachable manifold M\mathcal{M}, so they determine which geometries can be synthesized at all. The acquisition itself only requires γ(s)\gamma^{(s)} to be differentiable in zz and Z(s)\mathcal{Z}^{(s)} to admit a projection. We chose FFD for its low-dimensional representation, its flexibility across geometry formats, and its wide use in shape optimization, for example through pyGeo in the MACH-Aero and OpenFoam's DAFoam frameworks [1,2] and through the native FFD design variables of SU2 solver. However, as we acknowledged in Section 5, FFD provides no prior over plausible shapes, so Z(s)\mathcal{Z}^{(s)} and the validity checks must be configured with problem-specific knowledge. This configuration is done once per problem, and no individual sample required manual intervention.

In our experiments, for Airfoil, we restrict the control points to vertical displacements and pin the leading- and trailing-edge columns. Both surfaces therefore remain single-valued in xx, and a self-intersection can only occur as the upper and lower surfaces crossing, which the gap check detects together with the bound zmax⁡z_{\max} and the thickness to chord retraction during the ascent. Meshing is fully automated with identical settings for every sample, and no airfoil geometry failed to mesh or solve in any of our experiments. RD requires no meshing. Its rejections came from FFD perturbations that stretched the point cloud of the support, so that the source term became over-diffused and no longer reached its full amplitude. For the new elastoplastic benchmark, the checks reject self-intersecting cutouts and cutouts too close to the plate boundaries, and most of its rejections were self-intersections. Since all FFD control points were free to move in both directions (with a maximum perturbation of 0.08), more careful FFD constraints, as for Airfoil, could likely avoid these rejections.

We agree that the required constraints and checks grow with the flexibility asked of γ\gamma. Rescaling or locally reshaping a component with FFD needs only a few constraints, while synthesizing genuinely new designs in complex geometric problems with FFD would require many coupled constraints and substantial manual effort to keep the shapes valid and plausible. A natural direction for future work is to replace FFD with a parameterization that carries a learned prior, such as the latent space of a generative shape model or an SDF decoder (e.g. DeepSDF [3]), where plausibility comes from the training shapes.

[1] Kenway, Gaetan, Graeme Kennedy, and Joaquim RRA Martins. "A CAD-free approach to high-fidelity aerostructural optimization." 13th AIAA/ISSMO multidisciplinary analysis optimization conference. 2010.

[2] He, Ping, et al. "DAFoam: An open-source adjoint framework for multidisciplinary design optimization with OpenFOAM." AIAA journal 58.3 (2020): 1304-1319.

[3] Park, Jeong Joon, et al. "Deepsdf: Learning continuous signed distance functions for shape representation." 2019 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2019.

Dear Reviewer yF5J,

Thank you again for your constructive review. We have uploaded the revised manuscript (changes in blue). We would first like to highlight two new results that were not included in our earlier response.

  • Complete Elastoplastic results with the PaPQS baseline: The revised manuscript now reports the full Elastoplastic benchmark for Uniform, PaPQS, and AGS, with the budget curves in Figure 2c (Section 4.2) and the stress and plastic strain errors at every label budget in Table 8 (Appendix E.4).
  • Ablation on Elastoplastic: The ablation of the kernel conditioning and the trust region now also covers the Elastoplastic benchmark (Table 12, Appendix F), in addition to Airfoil and RD.

The remaining changes that address your concerns are listed below.

  • W1, additional benchmark: The Elastoplastic dataset is described in Appendix C.3.
  • W1, ablations: The ablations on Airfoil and RD are in Appendix F (Tables 10 and 11).
  • W1, additional baselines: The pool-based LCMD and MaxDet baselines for Airfoil are reported in Table 5 and Appendix E.5 (Figure 5).
  • W2, configuration effort: The choice of FFD, the problem-specific constraints, and the validity checks are discussed in Section 5 and the new Appendix B.6, and the checks used for the new benchmark are described in Appendix C.3.

We hope that the new results and discussions address your concerns, and we would greatly appreciate it if you could consider them in your final assessment. If you have any further questions or concerns, we would be happy to address them.

Best regards, The authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Revision

We thank all the reviewers for their insightful reviews and constructive feedback. We have uploaded a revised manuscript that incorporates the new results and discussions. Below, we summarize the major changes. All new text in the revised manuscript is coloured blue\textcolor{blue}{blue}.

  • New Elastoplastic benchmark (R1 yF5J, W1): We added a new Elastoplastic plate benchmark in Section 4.2, with Uniform, PaPQS, and AGS results. The dataset description and the errors per budget are provided in Appendix C.3 and Appendix E.4 (Table 8).
  • Ablation of kernel conditioning and the trust region (R1 yF5J, W1; R3 qAE9, W1): We added ablations of the kernel conditioning and the trust region on all three benchmarks in the new Appendix F.
  • Pool-based AL baselines, LCMD and MaxDet (R1 yF5J, W1; R2 oek5, Q1): We trained new pool-based AL baselines for Airfoil. The errors are reported in Table 5, and the budget curves and discussion in Appendix E.5 (Figure 5).
  • Uniform sampling with more compute (R2 oek5, W3/Q3 and W2): We trained Uniform Airfoil models with larger, computation-matched uniform datasets. The new results are reported in Appendix E.6 (Table 9).
  • Variance as an error surrogate (R2 oek5, W3): Spearman correlations for all additional Uniform models (Table 9) and std-vs-error plots for GNOT and the 8192-label model (Figure 6) are also provided in Appendix E.6.
  • Choice of geometry parameterization (R1 yF5J, W2; R2 oek5, Q2; R3 qAE9, W1): We elaborated on the choice of FFD, the validity checks, and the design constraints in Section 5 and Appendix B.6.
  • In-distribution and OOD trade-off (R2 oek5, W1; R3 qAE9, W2): We clarified the trade-off in the closing paragraph of Section 4.2 and in Section 5.
AuthorsOpenReview ↗

Summary of Discussions and Revisions

Dear AC, SAC, and Reviewers,

We sincerely thank everyone for their time and constructive feedback. As the rebuttal period is coming to an end, we would like to provide a final summary of our contributions, discussion, and revisions.

Contributions and strengths recognized by the reviewers

  • Well-motivated problem: small geometric changes can induce very different physics, which limits the generalization of neural operators to unseen shapes, "a problem of clear importance to industrial design" (yF5J).
  • Novel framework: the first synthesis-based active learning framework designed for continuous geometric distributions in operator learning (qAE9).
  • Significant OOD gains: up to 40% lower field error and 50% to 70% lower drag error on unseen airfoil families (yF5J, oek5, qAE9), with controlled comparisons at equal label budgets (yF5J).

Substantial revisions (locations are listed in our "Rebuttal Revision" comment)

  • A third benchmark, an elastoplastic plate with a cutout, where AGS reduces the OOD stress error of Uniform by 28%, 8%, and 30% on the three unseen cutout families (yF5J).
  • Ablations of the kernel conditioning and the trust region on all three benchmarks (yF5J, qAE9).
  • Pool-based AL baselines, LCMD and MaxDet (yF5J, oek5), and Uniform models trained with more compute (oek5).
  • An expanded discussion of the geometry parameterization, the in-distribution trade-off, and the limitations (yF5J, oek5, qAE9).

Summary of the discussion

  • Reviewer yF5J (2 → 3): In response to their concerns about the limited validation and the solver compatibility, we added a new benchmark and ablations and clarified the configuration effort. The reviewer wrote that our response "has clarified my concerns" and raised their score, and we incorporated the new results and discussion into the revised manuscript.
  • Reviewer oek5 (2): The reviewer raised concerns about (i) whether the errors stem from a coarse mesh, insufficient training, or the architecture, (ii) whether Uniform sampling could match AGS at the same compute, and (iii) the lower in-distribution accuracy. In response, we trained Uniform models with coarser and finer meshes, 2× longer training, and 2× and 4× more labels. None of these closes the OOD gap, including the 4× model, whose compute cost is comparable to the full AGS run, and the variance-error correlation holds across all of these models and a second architecture. We also added the pool-based baselines in response to their question on optimal design. The in-distribution trade-off is the same one that Reviewer qAE9 found clarified. We have not yet received feedback from Reviewer oek5 on these results, and we hope that they can be taken into account by the reviewer and the AC in the upcoming discussion.
  • Reviewer qAE9 (3, maintained): In response to their concerns about the uncertainty surrogate far from the training data and the in-distribution accuracy, we clarified the design of AGS and the trade-off. The reviewer found the argument "convincing" and the trade-off "clarified."

We hope this summary assists in your evaluation. Thank you once again for your time and consideration.

Best regards,

Submission 49820 authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.