Cross-attention-based bipartite graph neural network for coupled nodal and elemental field prediction in large-deformation sheet material forming

Technical Summary: Cross-attention-based Bipartite Graph Neural Network for Coupled Nodal and Elemental Field Prediction

Problem Statement
Finite element analysis (FEA) is the standard for predicting deformation in sheet material forming, capturing both nodal kinematics (displacements) and elemental deformation measures (strains, stresses, thinning). However, high-resolution FEA is computationally expensive, limiting its use in iterative design and optimization. While machine learning (ML) surrogates, particularly Graph Neural Networks (GNNs), offer acceleration, most existing mesh-based GNNs adopt a node-centred formulation. In these models, the primary predicted quantities are associated with mesh nodes. Consequently, element-level quantities (such as thinning, which is critical for assessing forming quality and failure) must be recovered indirectly via interpolation or post-processing. This approach obscures the native bidirectional coupling structure of the finite element update, where nodal kinematics drive element deformation, and element stresses assemble back into nodal forces.

Methodology
The authors propose CAtt-BiGNN (Cross-attention-based Bipartite Graph Neural Network), a surrogate framework designed to respect the native discretisation of finite element simulations.

1. Bipartite Graph Representation:
   Instead of a single node-centred graph, the model constructs a bipartite graph with two distinct vertex types:

   • Mesh Nodes: Carrying kinematic information (displacement increments) and contact features (tool gap, normal vectors).
   • Elements: Carrying deformation information (specifically thinning values).
   • Directed Edges: Two sets of directed edges connect these vertices: node-to-element (transferring kinematics to update strain/thinning) and element-to-node (transferring stress contributions back to nodal equilibrium). This structure mimics the information flow of explicit dynamic FEA without explicitly solving the governing equations.

2. Edge-Aware Cross-Attention Processor:
   The core of the architecture is a message-passing processor that utilizes a cross-attention mechanism conditioned on geometric edge features.

   • Unlike standard cross-attention where keys are derived from source node/element features, CAtt-BiGNN derives the Key vectors from the directed edge features (representing local geometric relations like relative position and distance).
   • The Query is derived from the receiver state (element or node).
   • This design allows the model to learn adaptive coupling weights that depend on the specific local geometry between a node and an element, enabling more physically grounded message passing than generic attention mechanisms.

3. Hierarchical Extension (CAtt-BiUGNN):
   To address scalability on large industrial meshes, the authors introduce CAtt-BiUGNN. This variant incorporates graph downsampling and upsampling operations (coarsening the bipartite graph) with skip connections. This allows for long-range information propagation while preserving the bipartite node-element formulation at all scales.

4. Adaptive Noise Strategy:
   To improve the robustness of autoregressive rollout (where predictions are fed back as inputs), an adaptive Gaussian noise strategy is employed. Unlike uniform noise, this strategy scales the perturbation based on the inverse of the displacement increment magnitude. This targets slowly moving or near-contact regions (which are sensitive to geometric fluctuations) while avoiding excessive perturbation in areas of large deformation.

Key Contributions

• Novel Representation: Introduction of a finite-element-inspired bipartite graph that retains nodal and elemental quantities on their native discretisation domains, eliminating the need for interpolation to predict thinning.
• Edge-Aware Attention: Development of a cross-attention processor where attention keys are constructed from geometric edge embeddings, enabling the model to adaptively weight node-element interactions based on local geometric relations.
• Hierarchical Scalability: Proposal of CAtt-BiUGNN, which extends the bipartite formulation to large meshes via downsampling/upsampling, maintaining accuracy across different graph sizes.
• Robust Rollout: Demonstration that adaptive noise injection improves the stability of autoregressive predictions, particularly for displacement fields, without degrading thinning predictions.

Results
The model was evaluated on two benchmark sheet forming cases: a dome-shaped case (cold forming, smaller graph) and a corner-shaped case (hot forming, larger graph).

• Dome Case (CAtt-BiGNN): Compared to node-centred baselines (MeshGraphNet, RGNN) and bipartite ablation variants (Vanilla-BiGNN, GAT-BiGNN, CA-BiGNN), the proposed CAtt-BiGNN achieved the best balance between displacement and thinning prediction. It reduced the mean positional Euclidean error by 21.7% and the top-5% relative thinning error by 53.8% compared to the MeshGraphNet baseline. Ablation studies showed that using edge features for attention keys (CAtt-BiGNN) outperformed using source node features (CA-BiGNN).
• Corner Case (CAtt-BiUGNN): In the hierarchical setting, CAtt-BiUGNN outperformed the hierarchical node-centred baseline (UMGN) and other bipartite variants. It reduced the mean positional Euclidean error by 41.8% and the top-1% relative thinning error by 80.0% relative to UMGN.
• Noise Strategy: The adaptive Gaussian noise strategy significantly improved displacement prediction accuracy (reducing errors by ~74% compared to no noise) and provided a more balanced improvement in thinning metrics compared to uniform noise, which degraded thinning localisation.

Significance and Claims
The paper claims that the proposed framework provides an effective surrogate for large-deformation sheet material forming by addressing the limitations of node-centred representations. The authors assert three main conclusions:

1. Representing nodal and elemental quantities on separate graph entities provides a more natural discretisation for coupled displacement-thinning prediction than node-centred formulations.
2. Edge-aware cross-attention is more effective than generic attention mechanisms for modelling directional node-element interactions because the weights are conditioned on local geometric relations.
3. Hierarchical propagation improves the scalability of the bipartite formulation on larger forming meshes.

The work positions FE-inspired graph representations as a promising direction for surrogate modelling, particularly when both nodal kinematics and elemental manufacturability indicators (like thinning) are required simultaneously. The evaluation is limited to tool-geometry variations within the sampled design space, with future work suggested to include material descriptors and process parameters.