acceptodds
Under review as a conference paper at ICLR 2027

LiftGCN: Efficient Energy-Preserving Graph Learning via Joukowski Spectral Lifting for Finite Element Stress Prediction

Abstract

Finite element stress fields often exhibit strong local non-smoothness, where stress concentrations near holes, notches, and loading regions induce sharp spatial gradients and high-frequency graph components. Although graph neural networks naturally operate on irregular finite element meshes, conventional message passing is inherently smoothing and progressively attenuates such high-frequency information. Unitary propagation alleviates this problem by preserving spectral magnitudes, but typically relies on matrix functions and high-order approximations with propagation complexity. We propose LiftGCN, an efficient spectrally stable graph network based on Joukowski spectral lifting. LiftGCN maps the real spectrum of a normalized graph operator onto the unit circle through the Joukowski relation and realizes the resulting spectral transformation as a simple second-order recurrence, avoiding matrix exponentials, eigendecomposition, and high-order polynomial truncation. We show that the linear Joukowski backbone has unit-modulus characteristic roots and admits an energy-preserving structure under a positive-definite metric, preventing exponential attenuation of graph-frequency components with depth. Each layer requires only one sparse neighborhood aggregation, yielding propagation complexity, while lightweight local nonlinear residuals provide expressive feature transformations. Experiments on finite element stress prediction demonstrate that LiftGCN achieves competitive overall accuracy while improving reconstruction of stress concentrations and local high-gradient structures with substantially reduced computational cost.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
GNNs, Spectral graph learning, Joukowski spectral lifting, Energy-preserving propagation, Finite element analysis, Stress prediction

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

This article proposes LiftGCN, a GNN surrogate model for predicting stress fields from finite element simulations in linear elasticity. The main contribution is a graph layer that projects the features onto the unit circle using the Joukowski transform, in order to better preserve local stress concentrations at a low computational cost. The authors compare it with several GNN models on three datasets and report a competitive overall accuracy and inference time.

Critical strengths
  • The use of the Joukowski transform is an interesting way to limit the smoothing of local stress concentrations by standard GCN layers. It has a similar goal to more advanced spectral approaches, but is simpler and less costly.
  • The paper is clearly written. The mechanical problem is well introduced, and the datasets, the graph construction and the training protocol are documented in detail in the appendices.
Critical weaknesses
  • The contribution is somehow limited to a new graph layer for this mechanical problem. While this is interesting, the improvements should then be clearly significant and supported by an extensive evaluation, which is not completely the case here. The datasets used for the evaluation have very little variability. The model is only tested in linear elasticity, and within each dataset all the samples have the same load and boundary conditions. Only the material properties or the geometry vary between samples, within narrow ranges. As a result, the stress fields are very similar from one sample to another, and a simple GCN already gives good results. The method should be evaluated on more varied cases, with different loads and boundary conditions, or with nonlinear materials such as hyperelasticity or elastoplasticity, to show that it actually better preserves stress concentrations.
  • The comparisons with the baselines are not always done in the same conditions. The small LiftGCN is presented as the fastest model on Mines Paris, but it has fewer layers than GCN. On the two other datasets, with the same number of layers, it is significantly slower than GCN. The latency comparison should be presented more fairly. The medium and large versions of LiftGCN also have many more parameters than the baselines, so their better scores may simply come from their larger size.
  • There is no separate validation set. The checkpoint of each run is selected on the test set, which is also used to report all the results. In addition, only mean values over the 10 runs are reported, without standard deviations, while many of the differences between models are small. This makes it difficult to assess whether the reported improvements are significant and generalize to unseen data.
Questions for authors
  • Why is LiftGCN slower than GCN on the two Abaqus datasets, while it is faster on Mines Paris? How do the two models compare with the same number of layers and the same width?
  • How do the baselines perform when they have as many parameters as the medium and large versions of LiftGCN?
  • The checkpoints should be selected on a separate validation set, and standard deviations over the 10 runs should be reported.
  • How does the method perform on datasets with different loads and boundary conditions, or with nonlinear materials?
2 comments
AuthorsOpenReview ↗

Response to Reviewer FXaG

We sincerely thank you for your detailed and constructive comments. In particular, your concerns about dataset diversity, fair computational comparisons, and model capacity have helped us improve the experimental design and better characterize the scope of LiftGCN. We have uploaded a revised manuscript. Our responses are provided below.

Q3. Experimental protocol and statistical reliability.

We agree that selecting checkpoints based on test-set performance may compromise the reliability of the evaluation. We have therefore adopted a strict 70%/20%/10% training/validation/test split for each of the 10 independent runs (seeds 42–51). Checkpoints are selected using validation MSE, while the test set is used only for final evaluation. All relevant experiments have been rerun, and the revised manuscript reports mean ± standard deviation over 10 runs.

Q1. Fair latency comparisons with GCN.

Following your suggestion, we have revised the model configurations to ensure a more controlled comparison. Both GCN and small LiftGCN use the same number of layers and hidden dimensions, with closely matched parameter counts. The configurations and inference latencies are summarized below.

DatasetHidden Dim.LayersGCN Params.LiftGCN Params.GCN Latency (ms)LiftGCN Latency (ms)
Mines Paris2706444k445k1.7191.758
Connecting lug50061512k1514k0.6910.900
Elbow bracket50061512k1514k0.5702.361
Staggered-notch plate50061513k1515k14.05515.841

The results show that LiftGCN achieves inference latency comparable to GCN across the evaluated datasets. More importantly, our primary efficiency comparison is with UniGCN, which also employs unitary propagation to retain spectral information. On Mines Paris, small LiftGCN requires only 1.758 ms per mesh, compared with 18.866 ms for parameter-matched small UniGCN, achieving approximately 10.7× lower latency.

Q2. Parameter-matched comparisons.

We have added medium-sized UniGCN and Hybrid LG configurations on Mines Paris, with parameter counts closely matched to medium LiftGCN (3.299M, 3.323M, and 3.268M, respectively).

ModelSizeParametersLatency (ms)PSNR@1% (dB) ↑HSG-NMSE@1%(×10−2\times 10^{-2}) ↓
Hybrid LGSmall438k13.36520.472 ± 0.4368.188 ± 0.545
UniGCNSmall437k18.86620.123 ± 0.0957.108 ± 0.137
LiftGCN (ours)Small445k1.75820.818 ± 0.1257.252 ± 0.137
Hybrid LGMedium3.268M24.17419.467 ± 0.44710.462 ± 1.107
UniGCNMedium3.323M56.36419.883 ± 0.2027.456 ± 0.274
LiftGCN (ours)Medium3.299M5.04121.976 ± 0.1335.683 ± 0.152

At the small scale, the three models achieve comparable results, with LiftGCN obtaining the highest PSNR@1%. More importantly, when increasing the model size from small to medium, the performance of UniGCN and Hybrid LG does not improve accordingly. Both models exhibit a slight decline in peak-region reconstruction and gradient accuracy. In contrast, LiftGCN improves consistently, with PSNR@1% increasing from 20.818 to 21.976 dB and HSG-NMSE@1% decreasing from 7.252 to 5.683.

This trend is also reflected in Figures 5 and 6. As model size and propagation depth increase, UniGCN and Hybrid LG tend to saturate or fluctuate, with some metrics even deteriorating. LiftGCN, on the other hand, exhibits consistent improvements across the tested configurations, benefiting from additional model capacity without sacrificing local reconstruction accuracy.

Q4. Evaluation under more diverse loading conditions and nonlinear material behavior.

To extend the evaluation beyond linear elasticity and fixed loading conditions, we have introduced a new Abaqus dataset: the Staggered-notch plate (Appendix B.2). This dataset uses elastoplastic steel with an initial yield stress of 355 MPa and subsequent plastic hardening. It contains 100 simulations with variations in notch depth (5–9 mm), axial force (28–58 kN), and load eccentricity (−6 to 6 mm). The resulting stress fields exhibit visibly different concentration patterns across samples, as illustrated in Figure 10(b). The key metric comparisons are summarized in the table below.

ModelLatency (ms)R²PSNR@1%HSG-NMSE@1%(×10−2\times 10^{-2}) ↓
LiftGCN-small15.8410.850±0.00720.08±0.4162.41±1.37
UniGCN96.1920.800±0.00317.90±0.4058.97±2.69
Hybrid LG137.5620.777±0.08517.81±1.5285.66±12.33

These results further demonstrate that LiftGCN achieves strong prediction accuracy while maintaining relatively low inference latency.

Thank you for your detailed answers and the additional experiments. They clarify the points I raised, and the results on the new elastoplastic dataset are interesting. I will raise my rating accordingly.

Reviewer_qc4vOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LiftGCN for finite element stress prediction. The method uses Joukowski spectral lifting to construct a second-order graph propagation mechanism that preserves spectral information while requiring only one sparse neighborhood aggregation per layer. Node-wise nonlinear residual blocks are incorporated to improve feature expressiveness. Experiments on three finite element datasets demonstrate competitive prediction accuracy, particularly in stress concentration and local gradient reconstruction, with favorable inference efficiency.

Critical strengths

The paper addresses an important problem in graph-based physical simulation, where conventional message passing may suppress high frequency information needed for accurate stress prediction. The proposed propagation mechanism is computationally efficient, and its energy preservation property for the linear backbone is mathematically well motivated. The experiments cover multiple datasets and include comparisons with various graph architectures, depth analysis, and ablation studies. The improvements in local stress reconstruction are encouraging.

Critical weaknesses

My main concern is the methodological novelty of this work. The proposed Joukowski recurrence in Eq. (7) closely resembles the classical Chebyshev polynomial recurrence, which has already been used in spectral graph convolution methods. Although the nonlinear residual blocks introduce additional flexibility, the paper does not clearly explain its methodological contribution beyond existing Chebyshev based approaches. I would like to see a direct comparison with ChebNet to better establish the novelty of the proposed method. Another concern is that the energy preservation analysis only applies to the linear propagation mechanism, rather than the complete network with nonlinear residual blocks. The paper does not provide sufficient evidence that the full model actually preserves high frequency information during training. Moreover, the ablation results in Table 2 show almost identical performance between fixed and learnable spectral scaling, leaving the contribution of the learnable coefficients unclear.

Questions for authors

(1)The proposed Joukowski recurrence appears closely related to the classical Chebyshev recurrence. Could the authors clarify the key methodological differences and include a comparison with ChebNet under comparable experimental settings? (2)Appendix C.1 indicates that the best checkpoint is selected based on test set performance. Could the authors explain this choice and report results using a separate validation set for model selection? It would also be helpful to include standard deviations across runs. (3)The energy preservation analysis only applies to the linear backbone. Could the authors provide empirical evidence that the full model retains high frequency information after nonlinear transformations? Additionally, can the authors provide parameter matched comparisons to better isolate the contribution of the proposed propagation mechanism?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qc4v

We sincerely thank you for your insightful comments. They have helped us clarify the mathematical foundations and architectural contributions of LiftGCN. We have revised the manuscript accordingly.

Q2. Checkpoint selection and experimental variability.

We agree that selecting checkpoints based on test-set performance may lead to an optimistic evaluation. We have corrected this issue by introducing disjoint training, validation, and test sets with a ratio of 70%/20%/10% for each of the 10 independent runs. Checkpoints are now selected exclusively using validation MSE, and the test set is reserved for final evaluation. All relevant experiments have been rerun, and the revised tables report mean ±\pm standard deviation over 10 runs.

Q1. Relationship between LiftGCN and ChebNet.

We appreciate this observation. Indeed, the linear recurrence of LiftGCN is mathematically connected to the classical Chebyshev recurrence. With x=ρcλjx=\rho_c\lambda_j, its mode-wise response satisfies hj,c(l)=Tl(ρcλj)hj,c(0)h_{j,c}^{(l)}=T_l(\rho_c\lambda_j)h_{j,c}^{(0)}.

We have added Section 3.4 to explicitly establish this connection and clarify the architectural differences.

The key distinction lies in how polynomial propagation is parameterized and incorporated into the network. ChebNet constructs a polynomial spectral filter by independently learning weights for different polynomial orders. In contrast, LiftGCN uses a channel-wise spectral coefficient ρc\rho_c, shared across propagation layers, and carries two consecutive hidden states through a second-order cross-layer recurrence. Each propagation step is followed by a node-wise nonlinear residual update, and the final prediction uses both consecutive states.

To provide a direct empirical comparison, we have added ChebNet and ChebNetII to Table 1. All three models have comparable parameter counts. Their results on Mines Paris are summarized below.

ModelParametersPSNR@1% (dB) ↑\uparrowHSG-NMSE@1% (×10−2\times 10^{-2}) ↓\downarrowLatency (ms) ↓\downarrow
ChebNet441k19.896±0.16719.896\pm0.1679.092±0.2489.092\pm0.2483.227
ChebNetII444k18.103±0.04218.103\pm0.04212.165±0.07212.165\pm0.0720.731
LiftGCN (ours)445k20.818±0.12520.818\pm0.1257.252±0.1377.252\pm0.1371.758

LiftGCN achieves the best performance on both peak-region reconstruction and local gradient accuracy. Compared with ChebNet, it improves PSNR@1% by 0.922 dB and reduces HSG-NMSE@1% by approximately 20.2%, while requiring less inference time. ChebNetII achieves a lower latency, but its peak-region and gradient reconstruction performance is weaker in this experiment.

Q3.1. Energy preservation and spectral retention in the full architecture.

To further examine the behavior of the complete architecture, we have added a spectral-mode retention experiment in Appendix A. Using the same graph eigenvector probes, we compare repeated GCN propagation, pure Joukowski propagation, and full LiftGCN. For the full architecture, the residual blocks are randomly initialized and kept untrained to isolate their effect on mode propagation.

The results show that GCN rapidly attenuates most graph modes, whereas pure Joukowski propagation exhibits oscillatory retention consistent with the theoretical analysis. The full LiftGCN architecture retains visible responses across spectral bands, although the nonlinear residuals modify their amplitudes. In particular, the low-frequency gain is associated with the normalization and affine bias in the residual blocks.

Q3.2. Contribution of learnable spectral coefficients.

Following your suggestion, we have expanded the ablation study to compare learnable channel-wise coefficients against three fixed settings: ρ=0.25\rho=0.25, 0.50.5, and 0.750.75. The results on Mines Paris are summarized below.

ModelSNR (dB) ↑\uparrowR2R^2 (×10−2\times 10^{-2}) ↑\uparrowPSNR@1% (dB) ↑\uparrowHSG-NMSE@1% (×10−2\times 10^{-2}) ↓\downarrow
Fixed-ρ\rho (0.25)15.834±0.04015.834\pm0.04085.846±0.12985.846\pm0.12919.740±0.11919.740\pm0.1199.626±0.1309.626\pm0.130
Fixed-ρ\rho (0.5)15.982±0.03615.982\pm0.03686.320±0.11586.320\pm0.11519.977±0.05719.977\pm0.0578.378±0.1138.378\pm0.113
Fixed-ρ\rho (0.75)16.109±0.03016.109\pm0.03086.716±0.09286.716\pm0.09220.467±0.07720.467\pm0.0778.952±0.1638.952\pm0.163
Learnable ρ\rho (LiftGCN)16.393±0.02916.393\pm0.02987.557±0.08287.557\pm0.08220.818±0.12520.818\pm0.1257.252±0.1377.252\pm0.137

The results reveal that no single fixed coefficient optimizes all evaluation metrics. For example, ρ=0.75\rho=0.75 achieves the best PSNR@1% among the fixed settings, whereas ρ=0.5\rho=0.5 achieves the lowest HSG-NMSE@1%. In contrast, the learnable coefficients achieve the best mean performance across all four metrics.

This comparison supports the use of channel-wise learnable coefficients: different feature channels can adapt their spectral responses to the prediction task, rather than relying on a single manually selected scaling factor.

Reviewer_hXu2OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LiftGCN, a graph network for predicting nodal von Mises stress on finite element meshes. Each graph eigenvalue is mapped onto a pair of unit-circle modes via the Joukowski relation and realised as the second-order recurrence Hl+1=2A~HlR−Hl−1H^{l+1}=2\tilde{A}H^lR-H^{l-1}. The linear backbone therefore preserves a quadratic energy (Eq. 9–11) at the cost of one sparse aggregation per layer, O(ed)O(ed) instead of the O(Ked)O(Ked) of UniGCN, and node-wise nonlinear residuals are added on top. Experiments compare against 11 graph baselines on the Mines Paris biaxial specimen and two Abaqus-generated datasets, reporting whole-field, peak-region and gradient metrics, latency, depth sensitivity and ablations.

Reviewer note: this submission is outside my area of expertise (computer graphics), and I have informed the AC. My assessment is limited to presentation and general experimental protocol; I cannot reliably judge the novelty relative to the spectral/oscillatory GNN literature.

Critical strengths

S1. The recurrence is simple and well motivated: one aggregation per layer gives unitary-like spectral behaviour. At a comparable size, small LiftGCN matches UniGCN's PSNR@1% (20.327 vs 20.096) at roughly 10× lower latency (1.528 vs 15.000 ms, Table 1).

Critical weaknesses

W1. Checkpoints are selected by the lowest test-set MSE (p.6; App. C, p.18), with no validation split. Tables report 10-run means without variance, on 20 test meshes for Mines Paris, so the reported margins cannot be assessed reliably.

Questions for authors

Q1. How do UniGCN and Hybrid LG perform when scaled to the parameter count of medium LiftGCN?

2 comments
AuthorsOpenReview ↗

Response to Reviewer hXu2

We sincerely thank you for your constructive suggestions on the experimental protocol and model capacity comparisons. Your comments have helped us strengthen the experimental evaluation. We have uploaded a revised version of the manuscript. Our responses are provided below.

W1. Checkpoint selection and experimental variability.

We agree that selecting checkpoints based on test-set performance is not an appropriate evaluation protocol, as it may introduce bias into the final results. We have therefore revised the entire experimental pipeline.

For each of the 10 independent runs (seeds 42–51), complete simulations are divided into disjoint training, validation, and test sets with a ratio of 70%/20%/10%. Checkpoints are selected according to the lowest validation MSE, and the test set is used exclusively for final evaluation. We have rerun the relevant experiments under this protocol and now report both the mean and standard deviation over the 10 runs in all result tables.

Q1. Parameter-matched comparisons with UniGCN and Hybrid LG.

Following your suggestion, we have added medium-sized UniGCN and Hybrid LG configurations with parameter counts closely matched to medium LiftGCN. All three medium configurations use eight layers, with 3.299M, 3.323M, and 3.268M parameters, respectively. The results are summarized below.

ModelParametersPSNR@1% (dB) ↑HSG-NMSE@1% (×10⁻²) ↓Latency (ms) ↓
UniGCN (medium)3.323M19.883 ± 0.2027.456 ± 0.27456.364
Hybrid LG (medium)3.268M19.467 ± 0.44710.462 ± 1.10724.174
LiftGCN (medium)3.299M21.976 ± 0.1335.683 ± 0.1525.041

With nearly identical parameter counts, LiftGCN achieves better stress-concentration reconstruction while requiring substantially less inference time than both baselines.

We also notice that increasing the parameter count of UniGCN and Hybrid LG leads to a slight decrease in their performance rather than an improvement. This trend can also be observed in Figures 5 and 6. Under our experimental settings, simply increasing the model capacity does not consistently translate into better prediction performance for these two architectures, possibly indicating difficulties in effectively exploiting the additional parameters. In contrast, LiftGCN consistently improves its performance as the model scales from small to medium and large configurations. This difference suggests that the proposed propagation and state-transition mechanism may provide a more effective way to utilize additional model capacity. Further investigating the relationship between LiftGCN's scalability, stability, and expressive power is an interesting direction for future work.

Thank you for your response and clarifications, which partially address my concerns. I will maintain my original rating of Weak Accept, as my limited expertise in this area does not allow me to confidently reassess the remaining issues.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.