STAMP: Manufacturing-Informed Generative Design and Shape Optimization of Sheet Panels
Authors: Jihoon Kim, Jun Hyeok Song, Wonho Ha, Namwoo Kang
Organizations: Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea · Narnia Labs, Daejeon 34051, Republic of Korea
Generative design of formed sheet panels must preserve geometric manufacturing conditions while exploring layouts and improving performance. STAMP embeds these conditions directly in the shape representation. Each panel is described by its material layout and a reference surface with one height at each occupied position in a plane perpendicular to the drawing direction. Nominal thickness is prescribed along the normal. The representation allows curved surfaces, ribs, and openings. It excludes reference-surface undercuts and keeps nominal thickness fixed without manufacturing penalty terms. Because these conditions belong to the representation, every generated and optimized panel satisfies them regardless of training-set size. Boundary-conditioned diffusion generates alternative layouts, and predictor gradients guide a surface optimization that retains outlines, openings, and designated fixed regions. We demonstrate STAMP on automotive hood inner panels using CarHoods10k. With a comparable number of stored values, depth maps reconstruct surfaces more accurately than point-cloud and voxel baselines. From only five training designs, one of three training runs yields novel and varied layouts. A generator trained on 500 designs, 6.5% of the training set, already reaches the layout variety of the full-data model. Across 11 held-out design families, predictor-based selection among existing dataset designs yields an average gap of 0.18 mm to the best available FEA deformation, versus an expected 1.59 mm for random selection. It does better than random selection in all 11 families. At a 10% depth-change bound, a network not used in the search predicts an average deformation reduction of 24.8% for six generated and existing panels, with a 3.2% average mass increase computed from geometry. Engineers can thus explore and optimize panels within this manufacturing-informed design space.
Figures & tables
Figure 1: STAMP design framework, illustrated with a CarHoods10k case study. Panels show (a) the prescribed boundary, (b) generated depth maps, (c) the reconstructed surface, (d) ensemble deformation prediction, and (e) FFD optimization. The overlaid grid shows the projected FFD controls C , and Dˉ denotes the mean predicted deformation. Depths are normalized. Examples come from three full-data training runs ( N = 7,660 training designs). Optimization uses a 10% depth-change bound and preserves the outline, openings, prescribed boundary heights, and hinge and latch mounts. Depth changes in the optimized 3D view are magnified 2 × for clarity.
Figure 2: Single-depth representation of a formed sheet panel, illustrated with a CarHoods10k example. (a) Reference surface. (b) Schematic orthographic rays give the first-hit distance L , which is converted to the normalized height d and stored per pixel. Openings and the exterior take q=−1 . Nominal thickness t is prescribed along the normal. Colours in (b) span d = 0.64–0.72. (c) The resulting depth map of the surface in (a).
Partition
Designs
Families
Training
7,660
83
Validation
912
10
Test
1,072
11
Total
9,644
104
Table 1: Dataset partitions. Families remain disjoint across partitions.
Figure 3: Representations of the panel in Fig. 2. (a) Reference surface sampled on a 512 × 512 grid. (b–d) Representations that each store about 65,536 values. (b) 256 × 256 depth map. (c) Surface triangulated from 21,845 points, with the stored points marked in the enlarged view. (d) 60 × 78 × 14 voxel occupancy grid. The lower row enlarges the boxed region five times.
Figure 4: Conditional generation under (a) Input A and (b) Input B, shown as depth maps and reconstructed surfaces from the first of three full-data training runs ( N = 7,660, R1). Inputs show the outer domain in grey and the prescribed boundary heights as a coloured band. For Input A, four of the eight saved samples from two dataset entries that share this input are shown. For Input B, the first four samples are shown. Symmetry is imposed by reflecting one half of each sample. Input boundary width is enlarged for visibility.
Figure 5: Effect of training-data size N . Panels show (a) generated depth maps, (b) fine-scale depth fluctuations (roughness, ↓ ), and (c) differences between generated material masks (Jaccard distance, ↑ ). R1–R3 denote independent training runs, with eight samples each under Input A, the same input as in Fig. 4(a). Metrics are averaged over Inputs A and B with eight samples each. Lines and bars show the mean and range across runs. Input boundary width is enlarged for visibility.
Figure 6: Structural prediction on 1,072 test designs from 11 design families. (a, b) Deformation and mass predicted by the mean of the three CNNs versus FEA, with the diagonal marking equality. (c) Pairwise ranking accuracy. (d) FEA deformation of the best available and CNN-selected candidates, and the expectation for uniform random selection, within each family of 92–100 candidates.
Study
Input
Test designs
Deformation MAPE (%)
Mass MAPE (%)
Wollstadt et al. [ 14 ]
Point-cloud latent code
Random 10% of all designs
5.6–8.1
1.9–3.8
Indupally et al. [ 40 ]
Attributed feature graph
301 designs distinct from training, six surfaces
10.6
3.0
Chavare [ 42 ]
Point cloud
20% of each family
5.7
2.0
STAMP
Depth map
11 families absent from training
5.9
1.4
Table 2: Deformation and mass errors of CarHoods10k predictors. Values for earlier studies are those reported in the cited papers on their own test sets. MAPE is the mean absolute percentage error.
Figure 7: Optimization with a 10% depth-change bound. (a, b) Initial and optimized geometry at actual scale, (c) depth change, and (d) section profiles along the line marked in (a) and (b), for the generated panel in Fig. 1. Section heights are in normalized OBJ coordinates, with the vertical axis exaggerated for visibility. (e) Predicted deformation reductions for G1–G3 and O1–O3, with filled markers for the two optimization CNNs and open markers for the evaluation CNN. Letters A and B in parentheses identify the inputs.
Depth-change bound (%)
Deformation reduction, optimization CNNs (%)
Deformation reduction, evaluation CNN (%)
Area-based mass change (%)
5
25.5 (19.4–32.0)
15.0 (5.8–23.9)
+1.3 (+1.0–+1.9)
10
37.0 (28.4–45.7)
24.8 (6.3–42.0)
+3.2 (+2.7–+4.0)
15
43.2 (32.1–51.9)
31.8 (7.5–49.8)
+4.6 (+4.2–+5.2)
20
46.0 (33.0–55.1)
35.2 (8.6–52.3)
+6.0 (+5.1–+6.5)
Table 3: Optimization results, reported as mean (range) across six starting panels. Deformation reductions are predicted. Optimization uses two CNNs, and a third CNN evaluates deformation independently of the search. Mass change is computed from the surface area at a uniform nominal thickness.
Depth-change bound (%)
Panel
Opt. deformation reduction (%)
Eval. deformation reduction (%)
Mass change (%)
5
G1
27.2
15.1
+1.01
5
G2
25.7
18.4
+1.16
5
G3
19.7
5.8
+1.57
5
O1
32.0
23.9
+1.11
5
O2
28.9
19.1
+1.31
5
O3
19.4
7.9
+1.92
Table S1: Optimization results for all starting panels and depth bounds. Deformation reductions are predicted by the optimization CNN mean (Opt.) and the evaluation CNN (Eval.). Mass change is computed from the surface area at a uniform nominal thickness.
Model and GPU
Training designs
Runs
Updates per run
Time (h)
Generator, RTX PRO 6000
5–7,660
12
240,000
31.08–32.76
Generator, A100 80 GB
5–7,660
6
240,000
40.82–41.48
Predictor, RTX PRO 6000
7,660
3
24,000
0.635–0.637
Table S2: Training costs per run for the main experiments. Generator times include validation. Predictor times are elapsed training-run times. Updates denote the scheduled budget. Sampling and reporting are excluded.
Representation
Stored values
Mean distance ( ×10−3 )
95th-percentile distance ( ×10−3 )
Opening IoU
Depth map, 256 × 256
65,536
0.142
0.753
0.94
Point cloud, 21,845 points
65,535
0.210
1.37
0.82
Voxel grid, 60 × 78 × 14
65,520
9.51
16.8
0.70
Table S3: Shape representations that store about 65,536 values, evaluated on the 104 designs of Section S1. Values are medians over designs, and distances are in common normalized coordinates. For the point cloud, the edge-length limit K is chosen for each design to give the smallest mean distance or the largest opening IoU.
Training designs
Roughness ( ×10−3 )
Mask Jaccard distance
5
1.02 (0.99–1.06)
0.035 (0.017–0.049)
10
0.80 (0.68–0.90)
0.029 (0.021–0.038)
50
0.65 (0.65–0.66)
0.087 (0.068–0.105)
100
0.65 (0.62–0.67)
0.108 (0.098–0.114)
500
0.66 (0.60–0.71)
0.141 (0.111–0.168)
7,660
0.55 (0.55–0.57)
0.127 (0.106–0.148)
Table S4: Generation metrics for one design from each of the ten validation families, with four samples each and no reflection. Values are the mean and, in parentheses, the range across three training runs.
Figure S1: Reconstruction distance across grid resolutions. (a) Mean and (b) 95th-percentile distances for 104 panels. Lines connect values for the same panel, dots show individual values, and boxes show quartiles and medians, with whiskers extending to 1.5 times the interquartile range. Distances use normalized coordinates and 512 × 512 visible-surface reference samples.
Figure S2: All five training designs for each five-design training run (R1–R3). Columns follow the saved training subsets. Depth maps are shown as stored, without added symmetry or smoothing. All depth maps share the same colour scale.
Figure S3: Input A samples from the five-design models at four fixed training budgets. R1–R3 denote the three training runs, each shown with its first four noise samples per budget. Depth uses a common normalized 0–1 scale. Smoothing and reflection match Fig. 5. The 240,000-update samples come from the original models of Fig. 5.
Figure S4: Five-design training-budget comparison of (a) surface roughness and (b) mask diversity. Each training run weights the two inputs equally, using eight samples per input and all 28 pairs for diversity. Shading and reference whiskers show the observed range across the three runs. Open markers denote the original 240,000-update models. All curves use the same sampling, smoothing, and reflection as Fig. 5.
Deep generative design can produce complex geometries, but unconstrained representations may create undercuts that prevent release from a two-part mold. We present CAST, which reformulates three-dimensional shape generation under a selected mold-opening constraint as a two-dimensional image generation problem using paired top and bottom depth images. Each valid depth pair defines one material interval along a selected mold-opening direction, so reconstructions exclude undercuts in that direction by construction. Diffusion models generate candidate shapes in this representation, and a finite element analysis (FEA) guided optimizer refines the same two boundaries. We evaluate diffusion-based generation with limited training data comprising 211 SimJEB bracket designs. On SimJEB brackets, FEA-guided refinement with a volume penalty reduced compliance in 193 of 200 generated candidates. The median reduction across all candidates was 36.8 percent in the optimization FEA model. Reconstruction of existing brackets revealed a mass--stiffness trade-off: filling occluded regions increased material volume and reduced displacement under the evaluated loads. Paired depth images thus connect generative exploration and structural refinement in a shared design space that preserves the selected pull-direction restriction.
Jihoon Kim, Yongmin Kwon, Namwoo Kang
Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea · Narnia Labs, Daejeon 34051, Republic of Korea · Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, Daejeon 34051, Republic of Korea
An inner hood panel must meet a deflection target, stay below a stress limit, and hit a mass target. Machine-learned surrogates have made the forward direction, geometry to performance, fast and routine. The inverse direction, producing geometry from a stated requirement, remains largely unaddressed for industrial parts whose design space is organized into discrete topology families rather than a continuous parameterization. This work presents a two-stage pipeline for that inverse problem. A reachability stage determines which topology families can satisfy a given requirement vector. A conditional variational autoencoder then generates point-cloud geometry within a selected family, and a neural-operator surrogate estimates the performance of each candidate. The pipeline is built entirely from public data and freely available compute, and is deployed as an interactive tool. The pipeline works, with qualifications that are reported as primary findings rather than caveats. The surrogate is accurate in aggregate, but its error is comparable to the performance differences it is asked to discriminate, which bounds what can be claimed for any individual generated design. That ratio of surrogate error to within-class signal is argued to be the quantity that determines whether a pipeline of this kind can work at all.
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.