acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Error Diagnosis for Data Acquisition and Training Strategy Selection in Graph Neural Network

Abstract

High-fidelity finite element (FE) analysis is widely used for simulating complex physical phenomena, but its computational cost limits repeated evaluations over large parametric design spaces. Graph neural network (GNN)-based surrogate models offer a promising alternative by directly representing irregular FE meshes and enabling rapid structural-response prediction. However, their accuracy can deteriorate in localized high-stress regimes. When large prediction errors occur, it is often unclear whether the relevant geometric information is present in the model input but sparsely represented in the training data, or whether the stress-driving geometry is insufficiently exposed by the input representation. Because additional FE data require substantial simulation effort, distinguishing these failure mechanisms is important for deciding whether to acquire more data, modify the representation, or use both. We investigate this problem using three-dimensional FE data from 2,315 parametric perforated-plate specimens for nodal von Mises stress prediction in a fixed region of interest. We evaluate two interventions—targeted FE data enrichment and explicit geometric conditioning—in a controlled 2 by 2 design consisting of a baseline model, enrichment alone, conditioning alone, and their combination. Geometric conditioning augments the baseline input with node-wise relational geometry and structure-level descriptors derived from CAD and mesh information. Across two seeds, adding 315 targeted training specimens changes node-level RMSE from 0.0889 to 0.0862 MPa, whereas geometric conditioning applied to the same 1,605-specimen base training set reduces RMSE to 0.0425 MPa. Using both interventions gives 0.0363 MPa. The value of conditioning is strongly geometry-dependent. In the narrow-ligament regime, conditioning reduces median peak relative error by 1.158–1.858 percentage points across the two seeds, and the conditioned model trained on 1,605 specimens outperforms the unconditioned model trained on 1,920 specimens in both runs. Thus, the investigated 315-specimen enrichment is insufficient to substitute for conditioning in this regime. Sample-level analysis further illustrates distinct failure mechanisms. In one specimen, a non-intersecting hole drives the stress concentration but is not explicitly described at the prediction nodes; all 242 severely underpredicted nodes persist after enrichment but disappear under geometric conditioning, consistent with representation insufficiency. In another specimen from the Intersecting regime, the stress-driving geometry is locally reflected in the prediction-region input; targeted enrichment reduces severe underpredictions from 40 nodes to 15, and the combined condition removes them. Feature ablation further shows that most of the observed improvement is associated with node-wise relational geometry rather than structure-level scalars alone. These results show that additional simulation data and geometric conditioning are not uniformly interchangeable across the design space. Their effects depend on the geometric response regime and are not simply additive. The results therefore support assessing whether stress-driving geometry is sufficiently observable before committing additional FE simulation budget.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, finite element analysis, surrogate modeling, stress prediction, targeted data acquisition

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

This paper compares two approaches to improving GNN-based finite element stress prediction: adding targeted training samples and adding explicit geometric features. The experiments show that geometric features provide much larger improvements than adding more simulation data, especially in narrow-ligament regions.

Critical strengths

The paper addresses a practical problem and uses a clear experimental design. The results are easy to interpret, and the analysis of different geometric regimes is helpful.

Critical weaknesses

1.1. The novelty is quite limited. Using geometric information, such as distances to boundaries or holes, to improve GNN-based stress prediction has already been explored in previous studies. The paper mainly compares two existing improvement strategies rather than proposing a new algorithm or a general method for selecting them.

  1. I also have concerns about the experimental setup. The baseline GNN does not have access to some important geometric information, while the geometry-aware model explicitly receives it. Therefore, the large improvement is not particularly surprising. In addition, the validation set was used to guide data acquisition and is also used for the reported evaluation. Without an independent test set, it is difficult to assess how well the conclusions generalize.
Questions for authors
  1. How does the proposed approach compare with existing geometry-aware GNNs that already receive the complete geometry information?

  2. Can the authors provide results on an independent test set that was not involved in selecting the data acquisition regions?

  3. Would the same conclusion hold under different data acquisition budgets or with a stronger sampling strategy?

Reviewer_qEDbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This article addresses stress prediction on perforated plates with a GNN surrogate trained on finite element (FE) simulations. The authors compare two ways of reducing the prediction error on specific geometries: adding targeted FE simulations to the training set, or adding geometric features derived from the CAD model and the mesh to the model input. Both approaches are evaluated separately and in combination, over two seeds. The authors report a larger error reduction with the geometric features, especially when a hole lies close to the loaded face. An ablation study evaluates the contribution of the different geometric features.

Critical strengths
  • The idea of comparing the benefit of additional features versus additional data is interesting, as FE simulations are costly. The comparison is done with the same architecture and training setup, which isolates the effect of each option.
Critical weaknesses
  • The "region of interest" is referred to several times but never clearly defined, and it is difficult to understand how this region is selected. This is a problem, because the model only sees the nodes of this region. When a hole lies outside of it, the baseline model has no way of knowing it is there, so it is not surprising that adding geometric features helps much more than adding data. A comparison with a model that sees the whole plate would be needed before drawing such conclusions.
  • There is no independent test set. The same validation set is used to choose where to add data, to select the checkpoints and to report all the results. In addition, only two seeds are used, and some of the differences between models change sign from one seed to the other, which makes the conclusions fragile.
  • The writing and organization of the article make it hard to follow, and Tables 1 to 5 are never referenced in the text. The FE model used to generate the data is only briefly described, and the 18 geometric features are only described in words, without precise mathematical definitions. As a result, it is not possible to understand precisely how the data and the model inputs are built, nor to reproduce the experiments.
  • Predictions and FE results are only compared visually on two examples (Figure 3), and no error maps are shown, which makes it difficult to put the reported errors into perspective.
Questions for authors
  • What exactly is the region of interest, and how is it chosen? How does the baseline perform when the model sees the whole plate?
  • How were the 18 geometric features chosen?
  • Why is a calibration FE solve needed, given that the problem is linear? How many FE solves were run in total for the 315 targeted specimens? If it is more than one per specimen, the data budget should be counted in solves rather than in specimens.
  • The article should report results on an independent test set, with more seeds.
  • How does the method behave with a different mesh or element type? Such experiments would help to show that the conclusions generalize beyond this specific setup.
Reviewer_pms9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether localized prediction errors in GNN-based finite element surrogates are better addressed through targeted data enrichment or geometric conditioning. The authors compare these interventions on stress prediction for perforated plates and find that conditioning provides substantially larger gains, particularly for narrow-ligament geometries. The paper further examines how these effects vary across geometric regimes.

Critical strengths
  1. The controlled experimental setup isolates the effects of adding training data and geometric conditioning while keeping the underlying architecture and training procedure fixed. This allows for a relatively clear comparison between the interventions.

  2. The results demonstrate the importance of geometric context for accurate stress prediction. The feature ablation further shows that node-wise relational geometry accounts for most of the improvement, whereas structure-level descriptors alone provide considerably smaller gains.

  3. The regime-wise analysis goes beyond aggregate prediction accuracy and examines localized stress concentrations. The sample-level examples also help illustrate how different geometric configurations respond to the interventions.

Critical weaknesses
  1. The baseline construction limits the novelty of the contribution. The model deliberately excludes mesh information outside the prediction region, including geometric information relevant to stress concentration. The conditioned model then receives this missing information explicitly. Consequently, the substantial performance improvements are somewhat expected rather than demonstrating a fundamentally new capability.

  2. The experimental comparison lacks stronger geometry-aware baselines. In particular, there is no comparison against models that already receive the complete geometric information through simpler representations, such as global hole parameters or more complete mesh connectivity. Without these comparisons, it is difficult to isolate the actual contribution of the proposed conditioning from the benefit of simply providing missing information.

  3. The proposed diagnostic framework remains largely retrospective. The paper compares the outcomes of different interventions but does not establish a practical procedure for distinguishing sampling insufficiency from representation insufficiency beforehand. This limits the applicability of the central idea of guiding data acquisition and training strategy selection.

  4. The evaluation has important limitations. The same validation set is used to guide targeted enrichment and evaluate the models, which introduces dependence between the acquisition strategy and the reported results. Furthermore, the comparison lacks alternative acquisition strategies under a matched simulation budget.

  5. The experiments focus on a single structural family and GNN architecture, with limited random-seed evaluation. This makes it difficult to determine whether the observed behavior extends beyond the specific geometry and representation choices considered here.

Questions for authors
  1. How does the proposed conditioning compare against simpler baselines that already receive the complete geometric information? This seems essential for establishing the novelty of the approach.

  2. What happens when hole parameters are broadcast to all prediction nodes, or when the relevant hole geometry is retained in the input graph? Would the proposed relational features still provide a meaningful advantage?

  3. Can the authors evaluate the models on a strictly independent test set and compare targeted enrichment against random or geometry-stratified sampling under the same simulation budget?

  4. What actionable criterion allows a practitioner to identify representation insufficiency versus sampling insufficiency without first training and evaluating all the alternative configurations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.