BiLoGo: Reducing the smoothing of localised responses in Transformer-based neural operators for sheet forming

Yingxue Zhao, Haoran Li, Haosu Zhou, Tobias Pfaff, Nan LiarXiv 2026 · arXiv

TL;DR

BiLoGo is a new "Local-Global" neural simulator that solves the "smoothing" problem in Transformer-based operators for engineering design. By combining local graph message passing with a dual-stream global Transformer, it accurately predicts critical material thinning near sharp features while using up to 69% less memory than current state-of-the-art hierarchical models.

What Problem Does This Paper Address?

In the world of automotive and aerospace engineering, "sheet forming"—the process of stamping flat metal into complex shapes—is traditionally simulated using Finite Element (FE) analysis. While accurate, FE is notoriously slow, often taking hours for a single design iteration.

Researchers have turned to Machine Learning (ML) simulators to speed things up. However, a major hurdle remains: smoothing. Transformer-based neural operators are excellent at capturing how a force on one side of a metal sheet affects the other side (long-range propagation), but they act like a low-pass filter. They tend to "blur" or smooth out localized responses, such as the sudden thinning of metal around a sharp corner or a reinforcing rib. In engineering, these "high-frequency" details are the difference between a successful part and a catastrophic tear.

Methods

The authors introduce BiLoGo (Bipartite Local-Global), a hybrid architecture designed to mimic the logic of actual FE solvers. It consists of three primary modules:

  1. Bidirectional Node-Element GraphSAGE: Unlike standard GNNs that only look at mesh nodes, BiLoGo treats nodes and elements as separate entities in a bipartite graph. This allows the model to predict displacement on nodes and thinning on elements natively, without losing information through interpolation.
  2. Dual-Stream Transformer: This is the "Global" part. Instead of lumping all data together, BiLoGo processes the blank deformation and tool contact in two separate token streams. This prevents the highly localized contact information from being "averaged out" by the state of the entire blank.
  3. Fusion and Element Update: To recover the local details lost during global compression, the model uses "gated residual fusion." It essentially decides, entity-by-entity, how much of the global update to use versus how much of the original local detail to keep. Finally, it performs a kinematics-conditioned update, using the relative positions of an element's nodes to refine its thinning prediction.

The paper explores two variants: BiLoGo-T (using feature-based "physical slices") and BiLoGo-U (using spatially distributed "supernodes").

Datasets

The model was evaluated on three increasingly difficult scenarios:

  • Dome Dataset: Cold forming of a smooth, double-curvature shape.
  • Corner Dataset: Hot stamping of a deep-drawn corner (stronger localized responses).
  • Corner+Rib Dataset: A newly introduced benchmark featuring a fixed corner geometry with reinforcing ribs of varying sizes and positions. This dataset specifically tests a model's ability to handle local geometric features that don't change the global shape but drastically alter local material flow.

Key Insights and Ideas

  • The Locality vs. Global Trade-off: The authors found that while feature-based tokenization (BiLoGo-T) works well for smooth shapes, it struggles when responses are tied to specific local geometry. In those cases, spatial tokenization (BiLoGo-U) is superior because it forces the model to respect the physical "neighborhood" of a feature.
  • Contact is Special: Treating tool contact as a separate stream was a breakthrough. Because contact is often the "driver" of deformation but is spatially sparse, separating it prevents its signal from being drowned out by the rest of the mesh data.
  • Memory Efficiency: Hierarchical GNNs often use edge-specific weights that scale poorly with mesh size. Because BiLoGo uses shared weights in its message-passing and Transformer layers, its parameter count remains constant regardless of mesh density, leading to massive memory savings.

Key Results

  • Accuracy: BiLoGo achieved the lowest positional and thinning errors across all three datasets. On the Corner case, BiLoGo-T reduced mean positional error by 15.5% compared to the previous best model (CAtt-BiUGNN).
  • Anti-Smoothing: Near reinforcing ribs, the "smoothing index" (where 1.0 is perfect) jumped from 0.41–0.56 in standard Transformer models to 0.94 with BiLoGo-U.
  • Efficiency: BiLoGo required 66–69% less training memory than hierarchical GNNs. For the Corner+Rib dataset, a full 10-step rollout took just 7.2 seconds on a single GPU, compared to 1600 seconds for a traditional FE simulation on a 12-core CPU.

Limitations and Future Work

While BiLoGo is a significant step forward, the authors note several areas for expansion:

  • Bending Effects: The current element update is based on membrane deformation (stretching). For processes where bending is a primary driver of thinning, the kinematic descriptors will need to be updated to include curvature changes.
  • Fixed Process Conditions: The current datasets use fixed friction, speed, and temperature. Future versions will need to incorporate these as variable inputs to handle more diverse manufacturing environments.
  • Mesh Topology: While the model's parameters don't depend on the mesh, the current study used the same mesh topology for all simulations in a set. Testing on varying topologies is a "natural extension" of the work.

Summary updated

Abstract

Trial-and-error FE simulation of sheet forming could be computationally expensive, neural-network-based learned simulators are used to accelerate design. These simulators should accurately capture localised responses, such as thinning around sharp radii or near ribs, which determine formability. Transformer-based neural operators propagate information efficiently across the mesh but tend to smooth such localised responses. This study develops BiLoGo, a Bipartite Local-Global learned simulator with an FE-guided design that connects local node-element message passing with a Transformer-based neural operator for long-range information exchange. BiLoGo has three modules: a bidirectional node-element GraphSAGE model that learns nodal-elemental coupling on a bipartite graph and predicts each quantity on its native FE entity without interpolation; a dual-stream Transformer-based neural operator that processes blank deformation and tool contact with separate token sets, preventing localised contact information from being averaged with the whole-blank state; and a fusion and element update module that combines global and local features via gated residual fusion and updates element features conditioned on element-node kinematics. Evaluated on a dome-shaped cold forming case and two corner-shaped hot stamping cases (one with ribs), BiLoGo achieves the lowest positional and threshold-based thinning errors on all three cases compared with single-level and hierarchical graph-based models and Transformer-based neural operators. Near the ribs, the smoothing index of predicted thinning increases from 0.41-0.56 for Transformer-based neural operators to 0.94 for BiLoGo-U. Compared with the hierarchical graph network CAtt-BiUGNN, BiLoGo-T reduces mean positional error by 15.5% on the Corner case and BiLoGo-U by 8.3% on the Corner+Rib case, with 66-69% less training memory.
BibTeX
@misc{zhao2026bilogo,
  author = {Yingxue Zhao and Haoran Li and Haosu Zhou and Tobias Pfaff and Nan Li},
  title = {{BiLoGo: Reducing the smoothing of localised responses in Transformer-based neural operators for sheet forming}},
  year = {2026},
  eprint = {2611.09621},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09621},
}
Figure