acceptodds
Under review as a conference paper at ICLR 2027

Reliable Surrogate Modeling of PDEs via Physics-Informed Conformal Prediction on Graph Neural Networks

Abstract

Deploying Deep Learning (DL) surrogate models in computational physics requires adherence to physical laws, generalization across diverse topologies of real-world applications, and the ability to quantify uncertainty. The high computational cost of traditional numerical solvers motivated the shift toward Artificial Intelligence (AI). Although Physics-Informed Neural Networks (PINNs) have enabled notable acceleration, the reliance on continuous automatic differentiation with respect to absolute coordinates removes spatial invariance and their ability to generalize to novel geometries. Graph Neural Network (GNN) and Geometric Deep Learning (GDL) surrogate models have addressed this by processing relative topological and graph-based features. Yet, the performance of standard message-passing architectures degrades under major distribution shifts in real-world deployment. Also, neither PINNs nor GNNs inherently quantify uncertainty, making them unreliable for safety-critical applications. To tackle these limitations, we propose the Physics-Informed Conformal Stacked Residual Network (PI-CSRN), based on a multi-resolution architecture with topology-preserving clustering and anisotropic, attention-driven message passing. PDEs are integrated into the training objectives of the surrogate models via discrete differential operators to preserve spatial invariance. Uncertainty estimation is provided via Conformal Prediction, with a residual learning mechanism to improve performance on novel manifolds. A custom benchmark of thousands of complex and diverse auto-generated 2D and 3D manifolds ensures the proper evaluation of surrogate models, using electrostatics and elasticity as the foundational physics. The proposed architecture achieves high accuracy, robustness to distribution shifts, and uncertainty quantification, thus offering a reliable blueprint for deployable AI physics surrogates.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Deep Learning, Geometric Deep Learning, Graph Neural Networks, Uncertainty Quantification, Conformal Prediction, Physics-Informed Machine Learning, Covariate Shift

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Dw85OpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces PI-CSRN, a graph-based surrogate model for solving partial differential equations (PDEs) on variable geometries. The framework integrates a Graph U-Net with Graclus-based graph coarsening, attention pooling, and TransformerConv layers. It incorporates a physics-informed loss based on discrete mesh operators and uses conformalized quantile regression (CQR) to construct prediction intervals. In addition, a stacked residual network is designed to refine the ensemble predictions and adjust the associated prediction intervals under distribution shifts. The proposed PI-CSRN is evaluated on procedurally generated benchmarks covering two-dimensional electrostatics and three-dimensional elasticity.

Critical strengths
  1. The problem is well motivated, addressing the important challenge of uncertainty quantification in mesh-based surrogate models when test geometries differ from those encountered during training.

  2. The use of relative features and discrete mesh operators to preserve invariance to rigid-body transformations is a coherent and well-justified design choice.

Critical weaknesses
  1. The claimed coverage result comes from labelled near-target data, not from the method’s design.

  2. The obvious baseline comparisons are missing including recalibrating CQR on the same small dataset, fine-tuning the base model on it, or weighted and non-exchangeable conformal methods. The current comparisons rely primarily on the authors’ own model variants, limiting the evidence for the proposed method’s advantages.

  3. The distribution shift and benchmark are insufficiently characterized, with no quantitative description of dataset sizes, mesh sizes, or differences between training and test geometries. Results are based on single runs without random-seed analysis, uncertainty estimates, or confidence intervals. Key implementation details, benchmark data, and quantitative results for 3D elasticity are also missing.

  4. The framework combines largely established components, while claims about preserving topology and geodesic structure and preventing representation collapse lack empirical validation. Feature ablations are absent, making it difficult to determine whether the proposed architectural choices provide meaningful benefits.

  5. The use of invariant scalar features raises questions about whether vector displacement fields can be represented correctly without an appropriate coordinate frame or equivariant architecture. The traction-free boundary condition appears incorrectly formulated, and the elasticity physics loss and material parameters are insufficiently specified.

  6. The evaluation focuses on linear, homogeneous, static PDEs that conventional FEM solvers can solve efficiently. Without runtime comparisons or more challenging test cases, the practical advantages of the proposed surrogate remain unestablished.

  7. Several claims, including strict guarantees, optimal performance, and cross-physics generalization, are not adequately supported by the reported experiments. The sigmoid output layer raises questions about predicting unrestricted displacement values, and the discrepancy between training and validation errors requires clarification.

Questions for authors
  1. Does the physics loss apply to the central prediction only, or also to the quantile outputs?

  2. The output layer in Fig. 1 ends in a sigmoid. How does that work for displacements outside [0, 1]?

  3. Why is the single stacked model’s train MAE above its validation MAE? Is it a typo?

2 comments

We would like to thank the reviewer for recognizing the motivation of uncertainty quantification, especially under geometric distribution shifts, and the design choice of combining relative geometric invariants with discrete mesh operators to preserve rigid-body invariance.

Regarding the questions addressed:

1) Does the physics loss apply to the central prediction only, or also to the quantile outputs?

Both Lpde\mathcal{L}_{pde} and Lbc\mathcal{L}_{bc} of the multi-objective loss function in Equation 4 apply directly to the central prediction y^\hat{y}. However, all three output heads (q^low\hat{q}_{low}, q^high\hat{q}_{high}, and y^\hat{y}) share a common encode-process-decode backbone. So, the two CQR heads whose predictions are correlated with the central prediction head also benefit, in line with Multi-Task Learning principles. This is demonstrated in Section 5 (Table 1), where adding Lpde\mathcal{L}_{pde} and Lbc\mathcal{L}_{bc} to the Conformal Anisotropic U-Net GNN improves all test-set metrics simultaneously: Test MAE drops from 0.0250 to 0.0220, coverage rises from 0.740 to 0.800, and interval width tightens from 0.066 to 0.064 compared to the Conformal Anisotropic U-Net GNN without these loss terms.

2) The output layer in Fig. 1 ends in a sigmoid. How does that work for displacements outside [0, 1]?

The sigmoid activation in Fig. 1 was applied only to the 2D Electrostatics experiments, where Dirichlet ports are ϕ∈[0,1]\phi \in [0, 1], thus, the interior potential is strictly bounded. For 3D Linear Elasticity, an unrestricted linear output head was used to predict signed displacement vectors. We have clarified in the revised manuscript that the use of sigmoid activation is optional and task-dependent.

3) Why is the single stacked model’s train MAE above its validation MAE? Is it a typo?

That was indeed a typographic error in Table 1 (Conformal Stacked Residual GNN), which was corrected in the revised manuscript from 0.0190 to 0.0120.

Regarding the weaknesses identified:

Section 5 (Table 1) systematically ablates each architectural and loss component, alongside standard Encode-Process-Decode baselines (Simple NN, SAGE, Residual SAGE, and Residual Transformer GNN), following the paradigm of MeshGraphNets (Pfaff et al., 2021), alongside Isotropic U-Net GNN and a Deep Ensemble UQ baseline (Conformal Anisotropic Ensemble GNN). Crucially, coverage improvement does not actually occur from near-target data. Even without fine-tuning, our Conformal Stacked Residual Ensemble GNN improves OOD coverage from 0.800 to 0.840 and lowers Test MAE from 0.0220 to 0.0208 over base CQR purely through stacked residual learning. Fine-tuning the residual corrector on a small, disjoint, previously unseen set (†\dagger) with reduced complexity relative to the test set led to further improvements in coverage (0.901) and achieved the best Test MAE (0.0159).

Across more than 10,000 simulations spanning 2D and 3D topologies, each experiment used 5,000 auto-generated topologies and 500 independent test topologies. Training meshes contain 2,000 to 10,000 nodes, whereas test meshes range from 2,000 to 100,000 nodes and include perturbed procedural parameters to achieve higher complexity with non-convex topologies with multiple holes and sharp concavities. The statistics mentioned are also included in the revised manuscript.

I thank the authors for their response, which answered my three clarification questions and added useful dataset statistics. However, it leaves most of my main concerns unaddressed. First, valid coverage still depends on labelled near-target data. Without fine-tuning, OOD coverage reaches only 0.800–0.840, below the nominal 0.9 level. The rebuttal also does not say whether the fine-tuning data are separate from the conformal calibration data. If they overlap, the coverage guarantee does not hold. Second, the baselines I requested are still missing: recalibrating CQR on the same near-target data, fine-tuning the base model on that data, and weighted or non-exchangeable conformal methods. Without these, it is unclear whether the stacked residual network adds anything beyond access to labelled near-target data. Third, the distribution shift is still not characterised quantitatively. All results come from single runs without variance or confidence intervals. The 3D elasticity results and the size of the fine-tuning set are still not reported. Finally, the rebuttal does not respond to my questions on the correctness and scope of the method: representing vector fields with invariant features, the boundary-condition and physics-loss formulation, feature ablations, a runtime comparison with FEM, and the claims of strict guarantees and cross-physics generalization. The reported OOD coverage contradicts the claim of strict guarantees.

Reviewer_Guv9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues most recent learning-based surrogate physical models lack the ability to quantify uncertainty, which is needed to deal with strong distribution shifts in real-world problems. They therefore introduce a new method that is a composition of well-established components, e.g., Graph-Unet, PINN loss, quantile regression, and ensemble. The resulting architecture looks very complicated, where each individual component itself is either computationally heavy (hierarchical graph and ensemble) or requires careful tuning with the hyperparameters (PINN loss and the quantile loss).

Critical strengths

None

Critical weaknesses
  • As noted in the summary, the resulting architecture is too complicated and seems brittle; such complexity would create no impact in the field.
  • The experiment is validated on a simple elastic PDE, though with a large number of nodes. The problem is static, where there is no time evolution involved, which I believe is not considered too challenging and interesting for engineers. Note that for this kind of problem, I believe most local graph networks with sufficient depth layers or sufficient treatment to resolve oversmoothing (such as using proper normalization) would work equally well.
  • This would lead to another concern: the authors claimed that their uncertainty prediction helps reduce the "geometric covariate shift" problem. For local GNNs, I would expect them to be able to deal with arbitrary geometry by design; see MeshGraphNet [1]. Therefore, I am not convinced that "generalizing to new geometry" is a benefit brought by the capability of quantifying the uncertainty.

[1] T. Pfaff, et al. Learning Mesh-Based Simulation with Graph Networks. (ICLR 2021)

Questions for authors

None

1 comment

First, we would like to thank the reviewer for the constructive comments.

Regarding the critical weaknesses identified:

1) As noted in the summary, the resulting architecture is too complicated and seems brittle; such complexity would create no impact in the field.

In the ablation study in Section 5 (Table 1), numerous model variants of the proposed architecture are presented in order of increasing capabilities, performance, and complexity. Different variants can be selected depending on the deployment requirements. For example, the Anisotropic U-Net GNN, which does not require the base model ensemble or post-processing, already reduces Test MAE from 0.0762 (SAGE GNN) to 0.0180 and can be deployed directly in applications where uncertainty quantification is not required, whereas the full PI-CSRN is designed for providing uncertainty quantification with increased robustness to distribution shift.

2) The experiment is validated on a simple elastic PDE, though with a large number of nodes. The problem is static, where there is no time evolution involved, which I believe is not considered too challenging and interesting for engineers. Note that for this kind of problem, I believe most local graph networks with sufficient depth layers or sufficient treatment to resolve oversmoothing (such as using proper normalization) would work equally well.

The proposed architecture was validated on two different elliptic PDEs, 2D electrostatics and 3D linear elasticity, across thousands of arbitrary, non-convex topologies with multiple holes and sharp concavities. While extending to time-evolving systems is an important direction noted in Section 6, research on elliptic PDEs presents unique challenges. In the ablation study in Section 5 (Table 1), all baseline and widely used flat GNN architectures exhibit significantly inferior performance compared to the proposed architecture. This highlights the need for wider receptive fields and the importance of the multi-resolution architecture for capturing global dynamics, compared to time-evolving systems with local information propagation within a finite radius. Also, just increasing the number of layers in a flat GNN architecture will not be sufficient for real-world applications where mesh density and thus the leading receptive field may vary.

3) This would lead to another concern: the authors claimed that their uncertainty prediction helps reduce the "geometric covariate shift" problem. For local GNNs, I would expect them to be able to deal with arbitrary geometry by design; see MeshGraphNet [1]. Therefore, I am not convinced that "generalizing to new geometry" is a benefit brought by the capability of quantifying the uncertainty.

The proposed PI-CSRN architecture was introduced to increase robustness to covariate shift, which negatively affected both MAE and coverage. The base ensemble's uncertainty intervals and central predictions are concatenated into the input feature space, thereby enriching the features for the final residual corrector. So, a model's generalization power and robustness to covariate shift are not a benefit of its ability to provide uncertainty quantification, but as mentioned at the end of Section 3.3, "the ensemble's predictive uncertainty is an informative feature for the stacked architecture, allowing the residual corrector to dynamically tune its localized refinements".

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

The paper implements a multi-res Graph U-Net for PDEs on irregular meshes.

  • Physics is injected through PINN-style training losses using discrete differential operators.
  • Uncertainty is quantified with standard split Conformalized Quantile Regression at the node level, and an ensemble with a stacked residual corrector is added to recover coverage under geometric distribution shift.

The method is evaluated on a newly generated benchmark of 2D electrostatics and 3D elasticity problems.

Critical strengths
  • The generated benchmark could be a useful resource for the community.
Critical weaknesses
  • Framing it as "physics-informed conformal prediction" overstates what is shown. The physics enters only as a training regularizer, and the conformal step is the textbook split CQR whose score and calibration are physics-free.
  • There are no external baselines like MeshGraphNets or neural operator models for accuracy, and no baselines like deep ensembles or weighted conformal prediction for uncertainty.
Questions for authors
  • Was the conformal threshold recomputed on a fresh held-out split after the stacked corrector and the fine-tuning?

  • How does the method compare with weighted conformal prediction or with field-level scores (e.g. the maximum or other stats over nodes) that give simultaneous coverage per geometry?

  • Given that the input features include geodesic distances to both ports, how does a simple feature-only baseline perform on the electrostatic benchmark?

1 comment

First, we would like to thank the reviewer for recognizing the usefulness of our benchmark comprising more than ten thousand arbitrary, auto-generated geometries and simulations.

Regarding the questions addressed:

1) Was the conformal threshold recomputed on a fresh held-out split after the stacked corrector and the fine-tuning?

Yes, this is correct. Strictly disjoint held-out splits were used. The stacked residual corrector was first fine-tuned on a small, disjoint dataset, and the conformal threshold was recomputed on an isolated calibration split to prevent any data leakage.

2) How does the method compare with weighted conformal prediction or with field-level scores (e.g. the maximum or other stats over nodes) that give simultaneous coverage per geometry?

In Weighted Conformal Prediction (WCP), calibration points matching the test distribution receive likelihood-ratio weights. In the presence of significant distribution shift, where the likelihood at almost all calibration points is very low, WCP concentrates almost all weight on a tiny fraction of calibration samples, leading to coverage degradation or even degenerate intervals. On the other hand, field-level scores yield a single global threshold per geometry, thereby losing the node-level resolution needed to localize high-error regions. Our method instead uses the augmented feature space to dynamically learn localized residual corrections from both the input features and the base ensemble's disagreement and interval widths.

3) Given that the input features include geodesic distances to both ports, how does a simple feature-only baseline perform on the electrostatic benchmark?

In the ablation study in Section 5 (Table 1), the first entry (Simple NN, Test MAE: 0.0870), which is the feature-only MLP baseline, and the flat message-passing baselines (SAGE GNN, Residual SAGE GNN, and Residual Transformer GNN, Test MAE: 0.0762 to 0.0780) have significantly inferior performance compared to our proposed PI-CSRN, which achieves a Test MAE of 0.0159 (an 81% error reduction). This confirms that port geodesic distances, with a simple feature-only baseline, are insufficient for complex topologies with irregular cavities without multi-scale topological message passing.

Regarding the critical weaknesses mentioned:

The multi-objective loss in Equation 4 integrates discrete differential operators (Lpde\mathcal{L}_{pde}) and boundary constraints (Lbc\mathcal{L}_{bc}) to enforce physical compliance while maintaining rigid-body spatial invariance. Although Lpde\mathcal{L}_{pde} and Lbc\mathcal{L}_{bc} are applied directly to the central prediction head y^\hat{y}, all three output heads (q^low\hat{q}_{low}, q^high\hat{q}_{high}, and y^\hat{y}) share a common encode-process-decode backbone. So, the two CQR heads whose predictions are correlated with the central prediction head also benefit, in line with Multi-Task Learning principles. This is demonstrated in Section 5 (Table 1), where adding Lpde\mathcal{L}_{pde} and Lbc\mathcal{L}_{bc} to the Conformal Anisotropic U-Net GNN improves all test-set metrics simultaneously: Test MAE drops from 0.0250 to 0.0220, coverage rises from 0.740 to 0.800, and interval width tightens from 0.066 to 0.064 compared to the Conformal Anisotropic U-Net GNN without these loss terms.

Finally, the ablation study of Section 5 (Table 1) evaluates architectures of incrementally increasing capabilities to isolate each component's contribution. The first section of Table 1 includes standard Encode-Process-Decode architectures (Residual SAGE GNN and Residual Transformer GNN) that directly follow the paradigm of established mesh surrogates such as MeshGraphNets (Pfaff et al., 2021). For uncertainty quantification, Table 1 explicitly includes both standard split CQR (Conformal Anisotropic U-Net GNN, MAE: 0.0220, Coverage: 0.800, Width: 0.064) and a Deep Ensemble baseline (Conformal Anisotropic Ensemble GNN, MAE: 0.0240, Coverage: 0.870, Width: 0.075), both of which are outperformed by PI-CSRN (MAE: 0.0159, Coverage: 0.901, Width: 0.065).

Comments on the paper

1 comment

We would like to thank all reviewers for their time, constructive feedback, and insightful comments. We appreciate that the reviewers recognized the motivation of quantifying uncertainty under geometric distribution shifts, the usefulness of our large-scale benchmark of over 10,000 simulations across arbitrary 2D and 3D manifolds, and the design choice of combining relative geometric invariants with discrete mesh operators to preserve rigid-body spatial invariance.

Below, we summarize our responses to the core technical questions shared across the reviews. Detailed point-by-point responses are also provided directly under each individual review.

1. Baselines, Architectural Modularity, and Elliptic PDEs

  • Accuracy and Baselines: Section 5 (Table 1) systematically ablates each architectural and loss component to isolate its contribution. It includes a feature-only MLP baseline (Simple NN, Test MAE: 0.0870) and flat Encode-Process-Decode message-passing architectures (SAGE GNN, Residual SAGE GNN, and Residual Transformer GNN, achieving Test MAE of 0.0762 to 0.0780) following the paradigm of established mesh surrogates such as MeshGraphNets (Pfaff et al., 2021). For uncertainty quantification, Table 1 compares our method against both standard split CQR (Conformal Anisotropic U-Net GNN, Coverage: 0.800, Width: 0.064) and a Deep Ensemble baseline (Conformal Anisotropic Ensemble GNN, MAE: 0.0240, Coverage: 0.870, Width: 0.075), both of which are outperformed by PI-CSRN (MAE: 0.0159, Coverage: 0.901, Width: 0.065).
  • Global Dynamics in Elliptic PDEs & Modularity: Unlike time-evolving systems with local information propagation within a finite radius per step, elliptic PDEs (2D electrostatics and 3D linear elasticity) require capturing global dynamics across variable-density meshes. Simply increasing the depth of flat GNNs is insufficient, whereas our multi-resolution Anisotropic U-Net GNN reduces Test MAE to 0.0180. Moreover, the Anisotropic U-Net GNN can be deployed directly in applications where uncertainty quantification is not required, while the full PI-CSRN provides robustness under distribution shift.

2. Physics-Informed Multi-Objective Loss, and Stacked Residual Learning

  • Multi-Objective Loss: Both the discrete PDE residual loss (Lpde\mathcal{L}_{pde}) and boundary condition loss (Lbc\mathcal{L}_{bc}) in Equation 4 are applied directly to the central prediction head y^\hat{y} using discrete differential operators to preserve rigid-body spatial invariance. Because all three output heads (q^low\hat{q}_{low}, q^high\hat{q}_{high}, and y^\hat{y}) share a common encode-process-decode backbone, these loss terms do not only influence the central prediction but also benefit the correlated CQR quantile heads in line with Multi-Task Learning principles. In Section 5 (Table 1), adding Lpde\mathcal{L}_{pde} and Lbc\mathcal{L}_{bc} to the Conformal Anisotropic U-Net GNN improves all test-set evaluation metrics simultaneously: Test MAE decreases from 0.0250 to 0.0220, OOD coverage increases from 0.740 to 0.800, and interval width tightens from 0.066 to 0.064 compared to training with the Ldata\mathcal{L}_{data} loss alone.
  • Architectural Gain vs. Fine-Tuning: Coverage increase under distribution shift does not occur from fine-tuning on near-target data. As shown in Section 5 (Table 1), even without fine-tuning (†\dagger), our Conformal Stacked Residual Ensemble GNN improves out-of-distribution (OOD) coverage from 0.800 to 0.840 and reduces Test MAE from 0.0220 to 0.0208 compared to the base CQR model purely through stacked residual learning. Fine-tuning the residual corrector on a small, disjoint set (†\dagger) with reduced complexity compared to the test set further increases coverage to 0.901 and achieves the lowest Test MAE of 0.0159.

3. Summary of Clarifications and Updates in the Revised Manuscript

  • Benchmark Dataset Statistics: Clarified in Section 4 that across the 10,000 simulations spanning 2D and 3D domains, each experiment comprises 5,000 auto-generated topologies (up to 10,000 nodes per graph) and 500 independent OOD test topologies (up to 100,000 nodes per graph) generated with perturbed procedural parameters featuring higher complexity, non-convex geometries, and sharp concavities.
  • Task-Dependent Output Activation: Clarified in Section 2 (Figure 1) that the sigmoid output activation was applied only to the 2D Electrostatics experiments where Dirichlet boundary potentials are strictly bounded within [0,1][0, 1], whereas an unrestricted linear output head was used for 3D Linear Elasticity to predict signed displacement vectors.
  • Table 1 Typographical Fix: Corrected the typographical error in Table 1 for the Train MAE of the single Conformal Stacked Residual GNN from 0.0190 to 0.0120.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.