PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Chen Zeng, Qiao Wang

arXiv:2609.14977v1cs.LGcs.AI

TL;DR

Finite element stress surrogates must recover sharp local variations that graph smoothing can suppress, while unitary propagation can require costly approximations. LiftGCN uses Joukowski spectral lifting and a second-order recurrence for efficient, energy-preserving propagation. Experiments report competitive overall accuracy and improved reconstruction of stress concentrations and local gradients at reduced computational cost.

  • Problem

    Stress surrogates must recover sharp local gradients, but graph smoothing attenuates useful high-frequency information and unitary alternatives typically incur O(Ked) propagation cost.

    Source: §1 · pp. 1–2

  • Method

    LiftGCN maps real graph eigenvalues to unit-circle modes through Joukowski lifting and implements the dynamics with a real-valued second-order recurrence and local residual transformations.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    Medium LiftGCN outperforms MeshGraphNets and UniGCN across reported metrics while faster; at comparable latency, large LiftGCN leads on peak-region PSNR at two smallest thresholds and all high-stress gradient metrics.

    Source: §4.2 · p. 6

  • Takeaways & Limitations

    The findings support spectral lifting as an efficient basis for graph-based stress prediction and retaining local physical detail as graph models aggregate wider context.

    Source: §5 · p. 9

Abstract

from arXiv · show

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 $O(Ked)$ 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 $O(ed)$ 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. Our code is available at https://github.com/ChenZeng001/LiftGCN.

1 Introduction

Finite element stress prediction must recover sharp local concentrations and gradients, yet graph aggregation can smooth away the high-frequency information they contain. LiftGCN uses Joukowski spectral lifting to preserve unit-modulus dynamics with one sparse aggregation per layer.

1 Introduction

  • Stress fields can contain sharp concentrations near holes, notches, interfaces, and loading regions, so useful surrogates must recover steep spatial variations as well as the overall field.These local responses matter for assessing structural failure.Source: §1 · p. 1
  • Repeated GNN neighborhood aggregation smooths node features and attenuates graph-frequency components that can encode mechanically meaningful concentrations and gradients.Smoothing can remove local details a stress surrogate needs to retain.Source: §1 · pp. 1–2
  • Unitary graph propagation preserves spectral magnitudes, but approximating its matrix exponential typically costs O(Ked) through repeated sparse operations.A single-aggregation realization would make unit-modulus dynamics more accessible to large finite element meshes.Source: §1 · p. 2
  • LiftGCN realizes unit-modulus spectral dynamics with a real-valued second-order recurrence, positive-definite energy preservation, and one sparse neighborhood aggregation per layer.The resulting graph-propagation cost is O(ed), without evaluating a matrix function.Source: §1 · p. 2

2 Background

The background links the limited regularity of finite element stress fields to the need to retain higher graph frequencies. It contrasts the frequency attenuation of diffusive propagation with unitary propagation, which changes phase without losing spectral magnitude.

2.1 Finite Element Stress Fields and Limited Regularity

  • Stress depends on displacement derivatives, so its natural regularity is generally L2 and global smoothness is not guaranteed.Holes and smooth fillets can produce large finite gradients, while sharp notches, re-entrant corners, and material interfaces can induce stronger local variations.Source: §2.1 · p. 3
  • Higher graph frequencies contribute more strongly to the graph variation measure, and preserving them helps reconstruct localized stress patterns.The section relates graph-frequency content to stress-field variation on a mesh graph.Source: §2.1 · p. 3

2.2 Diffusive and Unitary Graph Propagation

  • Repeated normalized-adjacency aggregation attenuates modes with eigenvalue magnitudes below one, favoring smoother representations and potentially weakening local stress contrasts as receptive fields grow.Components with smaller eigenvalue magnitudes disappear more quickly than low Laplacian frequencies associated with adjacency eigenvalues near one.Source: §2.2 · p. 3
  • Unitary propagation changes each spectral component's phase without losing magnitude, so repeated propagation preserves the feature signal's Euclidean norm.Approximating the matrix exponential requires multiple graph operations, motivating a simpler realization of unit-modulus dynamics.Source: §2.2 · p. 3

3 Joukowski Spectral Lifting and LiftGCN

LiftGCN realizes Joukowski spectral lifting as a real-valued second-order graph recurrence, combining phase-only spectral evolution with local nonlinear residuals. Its linear backbone preserves energy, and each layer needs only one sparse neighborhood aggregation.

3.1 Joukowski Spectral Lifting

  • Setting x = ρ_cλ_j with |ρ_c| < 1 lifts each graph eigenvalue to conjugate unit-circle modes whose phases vary by graph frequency and channel.The Joukowski map sends unit-circle modes to the scaled real graph spectrum and preserves their magnitude.Source: §3.1 · pp. 3–4, Figure 2 · p. 4
  • Equation 7 implements the lifted dynamics in real arithmetic using the previous hidden state, with learned channel-wise coefficients shared across depth.The coefficient constraint keeps the roots in the unit-modulus regime, while channels can learn distinct spectral phases.Source: §3.1 · p. 4

3.2 LiftGCN Architecture

  • LiftGCN applies node-wise nonlinear residual updates and a linear readout of the final two states to predict nodal stress.The recurrence controls inter-node transport while residuals learn local corrections without adding graph aggregations.Source: §3.2 · p. 4, Figure 3 · p. 5

3.3 Spectral and Computational Properties

  • Each graph mode and channel has unit-magnitude roots, and the linear backbone preserves energy across its pair of states.This provides a non-dissipative basis for propagation, while learned residuals add nonlinear corrections.Source: §3.3 · pp. 4–5
  • Each layer costs O(ed) for graph propagation, versus O(Ked) for a K-step matrix-function approximation, while retaining one-hop communication and local second-order memory.Additional layers extend the receptive field through the same sparse connectivity.Source: §3.3 · p. 5

4 Experiment

Across the Mines_Paris_Biaxial_Specimen benchmark, LiftGCN is reported to improve localized stress reconstruction at low propagation cost while retaining competitive overall accuracy. The experiments assess whole-field, peak-region, and high-stress-gradient behavior through comparisons, depth tests, and ablations.

4.1 Dataset and Evaluation Metrics

  • The study uses 100 biaxial-specimen simulations to predict nodal von Mises stress, evaluating whole-field accuracy, top-stress regions, and gradients around them.Source: §4.1 · pp. 5–6

4.2 Main Comparisons, Efficiency, and Depth Sensitivity

  • Medium LiftGCN outperforms MeshGraphNets and UniGCN across reported prediction metrics while remaining faster; large LiftGCN leads on selected peak-region and all high-stress-gradient metrics at comparable latency.Source: §4.2 · p. 6
  • LiftGCN reproduces compact high-stress bands near curved boundaries and heterogeneous specimen patterns, consistent with peak and gradient improvements.Source: §4.2 · p. 6, Figure 4 · p. 6
  • Across model sizes, LiftGCN reconstructs stress concentrations well at low latency, with additional capacity steadily improving peak and gradient accuracy.Source: §4.2 · p. 6, Figure 5 · p. 8
  • At comparable parameter counts, small LiftGCN has the lowest latency and improves whole-field and local accuracy over GCN, while matching or slightly improving UniGCN’s peak-region PSNR with faster inference.Source: §4.2 · p. 6, Table 1 · p. 7
  • As propagation depth increases, LiftGCN improves whole-field and local accuracy while GCN largely saturates and unitary and attention baselines perform best at shallower or intermediate depths.Source: §4.2 · pp. 7–8, Figure 6 · p. 8

4.3 Ablation Studies

  • Ablations show that removing Joukowski propagation or local residual transformations weakens reconstruction, while fixed spectral scaling retains most performance and learned coefficients add a small gain.Source: Table 2 · p. 9, §4.3 · p. 9

5 Conclusion

The conclusion presents LiftGCN as an efficient approach to reconstructing stress concentrations and local gradients, and identifies broader physical prediction tasks as future work.

5 Conclusion

  • Across 10 repetitions, LiftGCN combines efficient inference with accurate stress-concentration and local-gradient reconstruction; ablations show complementary benefits from lifted propagation and local nonlinear transformations.The conclusion also reports that the model uses additional capacity and depth effectively.Source: §5 · p. 9
  • The paper identifies broader geometries, material models, and physical prediction tasks as directions for future work.It characterizes these extensions as a natural direction, not as tested settings.Source: §5 · p. 9

AI use statement

  • The authors report that AI tools assisted with theory, experiments, code, literature work, and manuscript preparation, while the authors independently reviewed and verified the work.The authors state that reported experimental results came from experiments they conducted and evaluated.Source: AI use statement · p. 10

Reproducibility statement

  • Section 3 and Appendices A–C document the recurrence and assumptions, data and graph construction, additional results, and experimental protocols; the code is available publicly.The public Mines Paris data are identified by the citation Kerfriden (2022).Source: Reproducibility statement · p. 10

A Datasets and Graph Representation

This appendix describes the processed finite element datasets and clarifies how dataset statistics and graph-edge counts are reported.

A Datasets and Graph Representation

  • Counts cover all complete processed samples and are measured from coordinate and element-connectivity files.Graph edges follow the constructions described in Appendix A.3.Source: A Datasets and Graph Representation · p. 12
  • Each undirected graph edge is counted once before self-loops are added or message-passing directions are stored separately.Source: A Datasets and Graph Representation · p. 12
  • The dataset statistics table reports variable mesh sizes as minimum–maximum ranges, with arithmetic means on separate rows, and includes the degree feature in feature dimension.The table summarizes dataset statistics.Source: Table 3 · p. 12

A.1 Mines Paris Biaxial Specimen

The Mines Paris dataset uses random material realizations on a common biaxial-specimen mesh, with nodal coordinates, connectivity, and modulus defining the physical prediction task.

A.1 Mines Paris Biaxial Specimen

  • The processed dataset contains 100 random-material realizations on an identical mesh of 10,502 nodes and 43,399 four-node tetrahedra.The homogeneous reference realization indexed 0 in the source archive is excluded.Source: §A.1 · p. 12
  • The physical task maps nodal coordinates, tetrahedral connectivity, and the sample’s nodal Young’s modulus to a nodal stress target.The passage defines P as coordinates, T as connectivity, and E(s) as the sample’s nodal Young’s modulus.Source: §A.1 · p. 12
  • The network receives standardized coordinates, modulus, and log-degree features and predicts scalar von Mises stress; boundary conditions and load magnitudes are not separate inputs.Displacements and stress components are not input features.Source: §A.1 · p. 12
  • Because coordinates and connectivity are identical across realizations, topology can be reused, while the material field supplies the sample-dependent input.Source: §A.1 · p. 12

A.2 Abaqus Datasets: Geometry, Loading, and Simulation

The custom Abaqus datasets vary geometry under prescribed physical settings, while their graph construction and preprocessing specify how meshes and features enter the models.

A.2 Abaqus Datasets: Geometry, Loading, and Simulation

  • Two custom datasets use Abaqus/Standard static simulations with homogeneous isotropic linear elasticity and C3D10M tetrahedra, remeshed for each parameter combination.Both use Young’s modulus 200 GPa and Poisson’s ratio 0.30.Source: §A.2 · p. 13
  • The connecting-lug dataset varies three geometric parameters across 225 samples and applies concentrated nodal forces whose total resultant is 20,000 N.The shank face is fixed, and force is applied to nodes on the lower half of the pin-hole surface.Source: §A.2 · p. 13
  • The elbow-bracket dataset varies six geometric parameters across 288 combinations, with force distributed over the free end and stresses localized around bends, holes, and constraints.Figure 8’s contour panels each use their own stress colour scale, so comparisons should be read with each panel’s legend.Source: §A.2 · p. 13, Figure 8 · p. 15
  • At shared nodes, element-nodal stress tensors are averaged before computing von Mises stress; training uses that scalar invariant rather than averaged element-wise von Mises values.The exported table contains six tensor components and the scalar invariant.Source: §A.2 · p. 13
  • The experiments assess stress prediction across geometric configurations under fixed material constants, total load, load direction, and boundary-condition rules, without supplying them as extra input channels.The fixed total force is divided among a mesh-dependent number of loaded nodes; comparisons appear in Appendix B.Source: §A.2 · p. 16

A.3 From Finite Element Meshes to Graphs

  • Each mesh node becomes a graph vertex; Mines Paris tetrahedra connect all distinct node pairs, while custom quadratic tetrahedra split each edge at its midside node.For custom meshes, shared edges are deduplicated and stored in both directions for message passing; the ten-node elements are not converted to cliques.Source: §A.3 · p. 16
  • Training-set statistics standardize coordinates, modulus, and targets, while LiftGCN uses a normalized adjacency with self-loops and local-gradient evaluation uses original coordinates.The degree normalization is computed per graph without target values; the base graph has no separately measured edge attributes.Source: §A.3 · p. 16

B Model Comparisons on the Custom Datasets

LiftGCN performs strongly on both custom datasets, with the leading model and latency trade-offs varying by dataset and metric.

B Model Comparisons on the Custom Datasets

C.1 Shared Training and Evaluation Protocol

The shared protocol uses consistent simulation-level splits, checkpoint selection, and inference timing; model configurations are dataset-specific, with shared settings for the custom datasets.

C.1 Shared Training and Evaluation Protocol

  • Each of 10 repetitions uses a seeded simulation-level 80/20 split, shared across models on the same dataset and seed, with normalization recomputed from training data.The seed is 42 + r, and the first floor(0.8N) sample identifiers form the training partition.Source: §C.1 · p. 19, Table 7 · p. 19
  • For each repetition, the checkpoint with the lowest test-set MSE across 50 epochs is retained for reported metrics.Test-set MSE averages node-mean MSE over test meshes in standardized stress units.Source: §C.1 · p. 19
  • Latency is synchronized CUDA-event forward time per complete mesh on an NVIDIA RTX PRO 6000, excluding data loading, graph construction, transfers, normalization, and metrics.LiftGCN timings cache measurements by graph shape and profile a representative of each shape.Source: §C.1 · p. 19

C.2 Model-Specific Configurations

Loading 2609.14977v1…