Sheet metal components often consist of a global geometry, such as the corner radii and forming depth of a corner component, and local features added to it. During forming, both affect material flow. A surrogate learning all their combinations may need large training data. However, the same local feature type is often used on many global geometries. This study therefore partly decouples local features from the global geometry. The local-feature effect is first learned on one global geometry, needing fewer simulations than learning all combinations, then transferred to other global geometries via transfer learning, which can need fewer target simulations than training from scratch. A baseline-conditioned residual surrogate, R-BiLoGo, is proposed for ribs on the side walls of hot-stamped corner components. Its inputs are the baseline response and the geometric perturbation introduced by the ribs. Its outputs are the changes in displacement and thinning from the baseline. R-BiLoGo gave lower errors than a hierarchical graph-based model and a Transolver-based model on the source geometry, and on two of three selected target geometries after transfer. Across the 22 target geometries, transfer learning with 20 target simulations gave errors similar to or lower than training from scratch with 80 target simulations for about half of the targets. With the same 80 target simulations, transfer learning reduced errors for 95% of targets. The effect of the global geometry on transfer was studied with four parameter-free geometric descriptors and controlled variations across 4 corner subclasses. The benefit of transfer decreased with increasing surface distance and high-curvature change between source and target, and for targets shallower than the source. These trends were also seen in other corner types with a different magnitude.
Figures & tables
Figure 1: Definition of the residual fields for one rib configuration on a fixed global geometry. Top: final thinning of the rib simulation and the baseline simulation, and the residual thinning. Bottom: the same for the displacement, with the z direction shown as an example. The tool geometries of the two simulations are shown in grey. The dashed box marks the residual fields, which are concentrated near the ribs, while the complete fields are dominated by the forming of the global geometry.
Figure 2: Transfer workflow. The source model is trained with the baseline response of the source geometry and the geometric perturbations of many rib configurations. It is then transferred to a target geometry with a different corner shape, using its baseline response and a few rib configurations. Each target geometry is transferred separately. Although the baseline responses of the two geometries are different, the changes near the ribs show a similar pattern.
Figure 3: (a) Example of the simulation setup for one global geometry, with the blank and the tools of the quarter model. (b) Rib parameters on the formed part. The FE setup and the rib parameterisation are the same as in [ 20 ] .
Subclass
rd
rp
H
Subclass-specific parameters
Standard
25–40
20–27.3
60–120
rpv : 70–250
Chamfered
25–40
20–27.3
60–120
rs : 70–100; dc : 70–140
Elliptical
25–40
20–27.3
60–120
b : 65–100; a/b : 1.20–1.75
Ogee
25–40
20–27.3
60–120
rc : 80–120; rt : 50–80
Table 1: Parameter ranges of the four corner-geometry subclasses. All lengths are in mm. For the Elliptical subclass, b and the ratio a/b were sampled, and a was calculated from them.
Figure 4: The four corner-geometry subclasses with ribs: (a) Standard, (b) Chamfered, (c) Elliptical and (d) Ogee. The top row shows the forming surfaces, and the bottom row shows the plan view of the corner with the subclass-specific parameters.
Parameter
Range
Description
u
25–75%
Position of the rib centre along the top edge of the side wall
A/2
35–50 mm
Half-length of the rib along the side wall
B/2
22.5–30 mm
Outward extent of the rib from the side wall
Table 2: Rib parameters of the source dataset [ 20 ] .
Figure 5: Source geometry and the three selected target geometries: (a) source, (b) S-Med, (c) S-Max and (d) C-Med. From top to bottom: global geometry with ribs, FE thinning, magnitude of the residual displacement and residual thinning of an example simulation. The same rib configuration is used on all four geometries.
Group
Feature
Description
Node, baseline
xi0
Reference node position
xibase
Final node position of the baseline simulation
bix , biy
Boundary flags for nodes on the symmetry planes x=0 and y=0
Element, baseline
cj0
Reference position of the element centre
θjbase
Final thinning of the baseline simulation
Node, perturbation
diD , niD
Distance to the die with ribs, and unit normal of the die at the closest point
Table 3: Input and output features of R-BiLoGo.
Figure 6: Overview and architecture of R-BiLoGo. Top: inputs and outputs of the surrogate, shown for the thinning. Bottom: encoder, processor and decoder. The perturbation stream is shown in orange and the baseline stream in blue. Dashed lines show the features passed directly to later components.
Model
eN (mm)
eE ( ×10−4 )
eEtop (%)
Parameters
R-Transolver
0.0930
11.75
1.46
0.31M
R-CAtt-BiUGNN
0.0822
8.58
2.35
939.9M
R-BiLoGo
0.0743
7.82
1.16
1.48M
Table 4: Accuracy and number of parameters of R-BiLoGo and the baseline models on the source test set. The best value in each column is shown in bold and the second best is underlined.
Figure 7: Distributions of the errors over the 50 source test simulations for R-Transolver, R-CAtt-BiUGNN and R-BiLoGo. (a) Mean positional Euclidean error. (b) Threshold-based thinning error. (c) Top-0.25% critical relative thinning/thickening error. The triangles show the mean values.
Figure 8: Complete fields predicted by R-BiLoGo for two source test simulations. (a)–(f) First test simulation. (g)–(l) Second test simulation. (a), (g) FE displacement magnitude. (b), (h) Predicted displacement magnitude. (c), (i) Position error. (d), (j) FE thinning. (e), (k) Predicted thinning. (f), (l) Absolute thinning error. The predicted fields were obtained by adding the predicted residuals to the baseline response.
Figure 9: Distributions of the errors of R-BiLoGo, R-Transolver and R-CAtt-BiUGNN over the 50 test simulations of S-Med, S-Max and C-Med with K=20 . (a)–(c) Training from scratch. (d)–(f) Transfer learning. (a), (d) Mean positional Euclidean error. (b), (e) Threshold-based thinning error. (c), (f) Top-0.25% critical relative thinning/thickening error.
Target
Model
eN (mm)
eE ( ×10−3 )
eEtop (%)
ZS
TL
S
ZS
TL
S
ZS
TL
S
S-Med
R-Transolver
2.1960
0.1868
0.1632
50.88
2.27
1.99
60.96
2.82
2.46
R-CAtt-BiUGNN
0.7039
0.1532
0.3159
5.94
1.47
3.01
8.91
2.33
8.52
R-BiLoGo
1.7410
0.1005
0.1595
16.41
1.20
1.81
29.21
1.93
2.76
S-Max
R-Transolver
2.8866
0.1107
0.0972
70.94
1.97
1.75
193.37
4.26
3.84
R-CAtt-BiUGNN
1.2070
0.1456
0.1608
10.70
2.04
2.42
40.23
8.28
9.44
Table 5: Mean test errors on the three selected target geometries with K=20 . ZS denotes zero-shot prediction, TL denotes transfer learning and S denotes training from scratch. For each target geometry, the best value in each column is shown in bold and the second best is underlined.
Model
Training time per epoch (s)
Peak GPU memory (GiB)
R-Transolver
0.46
6.7
R-CAtt-BiUGNN
5.93
15.3
R-BiLoGo
1.25
11.6
Table 6: Computational cost of transfer learning on S-Max with K=20 , measured on an NVIDIA A40 GPU. One epoch is one pass over the 20 target training simulations. The batch size was 4 for R-Transolver and R-BiLoGo and 1 for R-CAtt-BiUGNN. The best value in each column is shown in bold and the second best is underlined.
Model
eN (mm)
eE ( ×10−4 )
eEtop (%)
R-BiLoGo
0.0743
7.82
1.16
Complete-field targets
0.1548
7.92
1.13
w/o baseline response
0.0783
8.18
1.18
w/o baseline fusion
0.0775
10.3
1.41
Table 7: Ablations of R-BiLoGo on the source test set. The best value in each column is shown in bold and the second best is underlined.
Model
eN (mm)
eE ( ×10−3 )
eEtop (%)
ZS
TL
S
ZS
TL
S
ZS
TL
S
R-BiLoGo
1.7410
0.1005
0.1595
16.41
1.20
1.81
29.21
1.93
2.76
w/o baseline response
1.7852
0.1020
0.1778
25.21
1.30
2.00
81.17
2.06
3.57
Table 8: Errors of R-BiLoGo and the model without the baseline response on S-Med with K=20 . ZS denotes zero-shot prediction, TL denotes transfer learning and S denotes training from scratch. The better value in each column is shown in bold.
Figure 10: Slice assignment of R-BiLoGo for an example test simulation of the source dataset. (a)–(f) Baseline stream. (g)–(l) Perturbation stream. (a), (g) Dominant slice of each node. (b), (h) Dominant slice of each element. (c)–(f) Assignment weights of baseline slices 12, 13, 11 and 5. (i)–(l) Assignment weights of perturbation slices 0, 10, 6 and 1. Each weight panel uses its own colour scale.
Figure 11: Test errors of R-BiLoGo against the number of target simulations K for transfer learning (solid lines) and training from scratch (dashed lines). The shaded bands show the standard deviation of the errors over the 50 test simulations. The dotted lines show the errors of the oracle, which was trained from scratch on the full training pool of 294 simulations. Rows: S-Med, C-Med and S-Max. Columns: mean positional Euclidean error, threshold-based thinning error and top-0.25% critical relative thinning/thickening error.
Figure 12: K80 transfer gain over the controlled variations of the Standard subclass. (a) δmax . (b) Δc0.99w . (c) ΔH . (d) δc , for S- δc -low, S- Δc -neg-high and S- δc -high. The solid lines show the mean of the node and element gains, and the dotted lines show the node and element gains. The shaded regions show the ranges in which the gain changed.
Target
Δc0.99w ( ×10−3 mm -1 )
δc ( ×10−4 mm -1 )
K80 gain (%)
S- Δc -zero
−0.20
0.87
26.2
S- Δc -pos-mid
2.00
2.68
25.9
S- Δc -pos-high
4.68
6.49
−15.3
Table 9: Δc0.99w , δc and K80 gain of the positive variation of Δc0.99w .
Target
δmax (mm)
Δc0.99w ( ×10−3 mm -1 )
ΔH (mm)
δc ( ×10−4 mm -1 )
S- δc -low
29.7
−0.20
0.6
0.87
S- Δc -neg-high
29.8
−3.72
0.6
1.95
S- δc -high
29.8
−0.20
0.6
3.37
Table 10: Descriptors of the geometries in Fig. 12 (d).
Figure 13: K80 transfer gain of the Chamfered, Elliptical and Ogee geometries. (a) δmax : C- δmax -low and C- δmax -high. (b) Δc0.99w : E- Δc -pos-mid and E- Δc -pos-high, which also differ in ΔH . (c) ΔH : E- ΔH -neg-mid and E- ΔH -neg-high. (d) δc : O- δc -low and O- δc -high, and C- δc -low and C- δc -high. The solid lines show the mean of the node and element gains, and the dotted lines show the node and element gains. Each panel has its own vertical range, and all panels have the same span.
Figure 14: Test errors of all target geometries for transfer learning with K20, training from scratch with K20 and training from scratch with K80. (a) Node error. (b) Element error. The arrows point from training from scratch to transfer learning with K20, and the values give the K20 transfer gain. The dotted lines separate the subclasses.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 15: Controlled variations of (a) δmax , (b) Δc0.99w and (c) ΔH , measured from the regenerated CAD geometries. Circles, squares and triangles denote δmax , Δc0.99w and ΔH . The vertical axis shows the quantile of each descriptor in the 128 Standard geometries. In (a), the open markers at the high case show the result before the null-space refinement.
Figure 16: Nested subsets of rib configurations for (a) K=5 , (b) K=10 , (c) K=20 , (d) K=40 , (e) K=80 and (f) K=160 . The grey points show the pool of 294 rib configurations, and the blue points show the selected configurations.
Figure 17: Test errors of R-Transolver against the number of target simulations K for transfer learning (solid lines) and training from scratch (dashed lines). The shaded bands show the standard deviation of the errors over the 50 test simulations. The dotted lines show the errors of the R-BiLoGo oracle, which was trained from scratch on the full training pool of 294 simulations. Rows: S-Med, C-Med and S-Max. Columns: mean positional Euclidean error, threshold-based thinning error and top-0.25% critical relative thinning/thickening error.
Figure 18: Predictions of R-BiLoGo after transfer learning with K80 for one target geometry of each subclass. From left to right: predicted final geometry, FE final geometry and absolute position error, and predicted thinning, FE thinning and absolute thinning error.
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.
Yingxue Zhao, Haoran Li, Haosu Zhou +2
Dyson School of Design Engineering, Imperial College London, London, UK · Prometheus, UK
A geometric style learned from one family of shapes is typically applied to new shapes from the same family. We consider a more challenging setting: learning the geometric design language of a stylized Latin alphabet and applying it to new forms, from Greek glyphs to entirely different structures such as human poses. We introduce Open-Form Geometric Style Transfer, where the forms to which a learned geometric style can be applied are not restricted to the shape family in which that style was observed. This setting is deceptively challenging: the difficulty is not merely learning a style, but learning the style without learning the shape family in which it was observed. We represent target forms by structural sketches and disentangle structural adherence from style realization using a Geometry LoRA and a style-specific LoRA. Critically, open-form transfer requires the Geometry LoRA to be universal. To achieve this, we devise a geometry-enrichment strategy that constructs a large and purposefully diverse collection of synthetic structural sketches spanning a broad range of unfamiliar geometries. Our experiments and ablations demonstrate the importance of both disentanglement and this universal geometry prior, enabling geometric styles learned from glyphs, pictograms, and other shape families to be transferred to substantially different, unseen forms.
Traditional sheet metal forming relies on time-consuming and expensive Finite Element Analysis (FEA) for design validation, a process that significantly prolongs design cycles. While surrogate models offer faster iteration, current approaches have limitations: scalar-based methods cannot capture comprehensive field-based FEA results, while existing image-based models often ignore the critical role of material properties by focusing solely on geometry. To address this gap, we develop a physics-guided deep learning framework, namely StampFormer, which simultaneously uses component geometry and material stress-strain responses to predict FEA outcomes. The StampFormer framework uses three core components to process data. A Material-Augmented Geometric Network (MAGN) first fuses geometric and material data. This information is then integrated at various levels by a Hierarchical Material Embedding Injection Unit (HMEIU) before being processed by the primary network backbone, an adapted Swin-UNet. We evaluated our model on the stamping of a crossmember panel with two simulation datasets for steel and aluminium panels, and results demonstrate that StampFormer provides high-fidelity predictions of critical physical fields - including thinning, major strain, minor strain, plastic strain, and displacement - in under a second. Compared with ground truth FEA, our model achieved an average relative error of less than 8.5% on the four 2D fields and a mean squared error of less than 1.2 mm2 for the 3D displacement field. In summary, we introduce a practical and efficient framework that integrates multimodal information, namely geometry and material properties, to provide fast and accurate predictions, enabling designers to perform real-time manufacturability assessments.