acceptodds
Under review as a conference paper at ICLR 2027

DeAL: Density-Agnostic Learning for Sampling-Stable Geometry-to-Field Surrogates

Abstract

Geometry-to-field surrogates predict physical fields from geometry and operating conditions, reducing the need for repeated numerical simulation. Because practical models receive finite meshes or point clouds rather than the underlying continuous domain, remeshing can redistribute local sampling density and connectivity without changing the physical case yet still alter predictions, a phenomenon we call sampling-distribution sensitivity. We introduce DeAL (Density-Agnostic Learning), a backbone-agnostic training principle that draws two fixed-budget views from a continuous inverse-density family for each case, supervises both predictions at common physical queries, and penalizes their disagreement. Dual supervision anchors both views to the physical target, while prediction consistency suppresses representation-dependent variation without changing the backbone or single-view inference. Across nine surrogate architectures and four physical systems, we find that DeAL reduces error and inconsistency among predictions by roughly half on average over a large number of matched comparisons when the same physical cases are evaluated under held-out spatial redistributions and independently generated remeshes, demonstrating transfer across distinct encoder designs and physical settings. Together, these results show that DeAL substantially improves predictions when the same physical case is presented through different numerical representations, without changing the underlying architecture or deployment path. In this way, our method increases robustness and reliability of neural surrogates as a key component for industrial applications.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
geometry-to-field surrogates, sampling-distribution sensitivity, density-agnostic learning, point clouds, remeshing robustness, neural operators, PDE surrogates

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_odAqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DeAL targets sampling-distribution sensitivity in geometry-to-field surrogates: predictions can change when the same physical case is supplied through a different surface sampling or remesh. A coordinate-only kernel concentration score defines a continuous inverse-density family; paired fixed-budget views with independently drawn strengths are supervised at common queries, and a symmetric stop-gradient Smooth-ℓ1\ell_1 term penalizes their disagreement; the backbone and single-view inference are unchanged. Across diverse backbones and CFD, heat-exchange, and magnetostatic systems, DeAL lowers error and prediction disagreement relative to matched single-view training under held-out spatial redistributions and remeshes.

Critical strengths
  1. Consistent matched gains in accuracy and stability. The single-view Base degrades under held-out shifts in every task–architecture pair, and DeAL lowers both error and direct prediction disagreement in every pair (Sec. 6.1; Table 6; App. E, Table 8), with remesh geometry audited (App. I), paired intervals for SMART (Table 14), and seed replication on a subset (App. H).

  2. Drop-in design with matched component ablations. DeAL changes only view construction and the loss (Eqs. (6)–(10); Algorithm 1), needs no density computation at inference, and matches Base in total supervised-view passes rather than optimizer updates (Sec. 5.1; App. C). With SMART, prediction consistency improves on density-varied supervision alone, and continuous density draws outperform fixed-intensity and endpoint pairs (Sec. 6.3; Fig. 4; App. Fig. 7).

Critical weaknesses
  1. DeAL's advantage over other view generators rests on a narrow comparison. Downsampling and masking controls are compared across systems but only with SMART (Sec. 6.2; Table 9), as are all other ablations in App. F. Each control is run at one fixed strength (App. A), with masks removing only a small fraction of points in the illustrated example (Fig. 20), whereas DeAL's preselected settings receive a sensitivity analysis (Table 11). Although Sec. 6.2 limits its conclusion to the tested controls, Sec. 7 draws a broader implication about view construction, which Sec. 4 presents as DeAL's defining difference from consistency-based training.

  2. Practical value relative to simple resampling preprocessing is not established. Remeshing motivates the work (Sec. 1), and Sec. 2.2 contrasts density-biased inputs with area-uniform sampling; yet Base trains on source-uniform selections, remeshed inputs are drawn uniformly from vertices (App. A; App. D), and no single-view baseline resamples inputs to reduce density bias in both training and inference (e.g., area-uniform resampling of each surface mesh, or, for point clouds, fixed-β\beta inverse-density resampling via Eqs. (6)–(7)). Such a baseline is needed to establish DeAL's incremental practical value over simpler preprocessing in the re-meshing workflows targeted in Sec. 7.

Questions for authors
  1. (Related to W1; suggested experiment.) How sensitive is Table 9 to control strength (App. A), and how do the downsampling and masking controls compare with DeAL for an already evaluated backbone with different aggregation (e.g., hierarchical PointNet++ SSG or radius-graph PointGNN; Sec. 5.1)?

  2. (Related to W2; suggested experiment.) On the existing remesh conditions (Sec. 5.2), how does DeAL compare with a single-view Base trained and evaluated on area-uniform samples of the native and remeshed surfaces (e.g., for SMART)?

If any part of this review reflects a misunderstanding of the paper, I would appreciate the authors pointing it out, and I am happy to revise my assessment accordingly.

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

This paper studies sampling-distribution sensitivity in geometry-to-field surrogates, where different discretizations of the same physical case can produce inconsistent field predictions. It proposes DeAL, a training framework that constructs two fixed-budget point-cloud views using continuously varied inverse-density sampling, supervises both predictions at common physical query locations, and penalizes their disagreement. The framework leaves the backbone and single-view inference procedure unchanged. Experiments across nine architectures and four physical systems show improved accuracy and prediction consistency under held-out spatial redistributions and standard remeshing operations. Controlled ablations examine the contributions of density-specific view construction, continuous density variation, and prediction consistency.

Critical strengths
  1. Broad evidence of improved robustness under the evaluated representation shifts. DeAL reduces average shifted-input error in all 36 task–architecture comparisons, and the separate prediction-disagreement metric supports the claim of improved representation stability (Section 6.1; Tables 6 and 8). The evaluation includes standard remeshing methods absent from training, with geometric-fidelity audits and additional training-seed experiments. These results provide substantial support for the method’s effectiveness within the evaluated setting.

  2. Informative controlled ablations and straightforward integration. The matched view-generator comparisons retain paired supervision and consistency, while the consistency ablation retains both density-varied supervised views. Together with the continuous-versus-fixed-density comparisons, these experiments help establish that the density-specific construction and consistency term both contribute (Sections 6.2–6.3; Tables 9–12). The framework also preserves the existing backbone and single-view inference procedure.

Critical weaknesses
  1. Incremental methodological contribution. The framework primarily adapts established data-augmentation and consistency-regularization principles to sampling sensitivity in geometry-to-field surrogates. The continuous inverse-density view construction is relevant and empirically useful, but the paired supervision and prediction-consistency objective do not appear to constitute a substantial conceptual departure. I therefore view the methodological contribution as incremental, while recognizing the care and breadth of its empirical validation.

  2. The connection to practical deployment remains insufficiently established. The experiments demonstrate robustness to unseen resampling and standard remeshing, but the practical circumstances in which these gains are most consequential are less clearly established. In the presented setup, the underlying meshes and source point clouds are available, so applying a common sampling or preprocessing policy during training and deployment appears feasible. The paper does not sufficiently explain when such consistency is unavailable or inadequate, or demonstrate DeAL’s advantage under a concrete constraint of that kind relative to feasible preprocessing alternatives. A shared sampling rule need not eliminate every density difference, but the remaining practical problem and the method’s benefit in addressing it need clearer evidence. This limits my assessment of the significance of the otherwise well-supported improvements.

Questions for authors
  1. Practical setting — clarification. What concrete deployment setting makes a common sampling or preprocessing policy unavailable or insufficient at training and inference? The paper motivates the problem through remeshing, but what representation differences remain after the preprocessing that is feasible in that setting?

  2. Generalization advantage — evidence relevant to the recommendation. For that setting, which reported experiment establishes DeAL’s advantage under the identified constraint? If the existing experiments do not directly address it, a focused comparison against a feasible common-preprocessing baseline, where applicable, would materially strengthen the practical significance of the contribution.

1 comment

Thank you for the clarifications and additional experiments. The SMART baseline using FPS at both training and inference directly addresses my request for a common-preprocessing comparison, and DeAL achieves lower mean errors across the reported sampling-shift conditions. The recomputed C-core fields and broader augmentation controls also strengthen the evidence for the reported robustness gains.

I still regard the methodological contribution as incremental. The practical benefit also remains less directly established: the rebuttal explains why geometry preparation may differ between training and deployment, but does not directly measure DeAL’s benefit in an actual design workflow with those differing input preparations. Thus, the experiments demonstrate improved robustness, while the magnitude of the benefit in actual engineering use remains uncertain.

Nevertheless, the additional results meaningfully address my original concerns. Together with the breadth of the evaluation across architectures and physical systems, they provide a sufficiently strong empirical contribution for weak acceptance. I therefore raise my rating from 2 to 3 (weak acceptance), while retaining the above reservations about methodological novelty and practical validation.

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

This paper addresses an important but easily overlooked issue: for the same object, differences in surface point-cloud sampling density can lead a model to produce different predictions. To address this issue, the authors propose DeAL. The method samples points with probabilities inversely related to their local density: points in densely sampled regions are less likely to be selected, while points in sparse regions are more likely to be selected. A parameter β\beta controls the strength of this reweighting; as β\beta increases, sampling increasingly favors sparse regions. DeAL generates two views of the same object and uses both to train the model. The training objective encourages each prediction to match the ground truth while also reducing the discrepancy between the two predictions. Experiments show that the method is effective across different tasks and physical systems.

Critical strengths

1.This paper addresses an interesting problem: for the same object, a model may produce different predictions when given point clouds with different sampling densities. The paper introduces a clear objective: when the point-cloud density changes but the object remains the same, predictions should stay close to the ground truth and remain consistent with one another. 2.DeAL can be readily applied to existing models without changing their architectures or inference procedures. The method is evaluated across a range of models and physical systems, with consistent benefits.

Critical weaknesses

The remeshing experiments use the original physical field as the ground truth. Although the paper provides a detailed geometric-fidelity audit showing that the shape changes are small, it does not report how the physical field would change if it were recomputed after remeshing, particularly near locally sensitive features such as the magnetic-core gap.The authors should analyze how such small local geometric changes may affect the performance of the proposed method.

Questions for authors

I would appreciate further analysis of regions where the physical field may be sensitive to small local geometric changes, such as the magnetic-core gap.

Comments on the paper

1 comment
Reviewer_odAqOpenReview ↗

Response to Authors

Thank you for the careful answers to my questions. The additional experiments show that DeAL outperforms other view generators at every tested strength and on a second backbone. Also, DeAL has lower error than the FPS baseline without extra inference cost.

While an area-uniform comparison on the remesh conditions is still missing, but it is unlikely to be a concern since DeAL also leads on the densely meshed car remeshes.

Overall, I confirm my positive assessment of 3 (weak accept).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.