ME1a_DT + AC28 results

ME1a used 50 XLS process rows, 6 CSV defect marks, 57,428,249 fused PLY vertices, 5,663 B-Rep faces and 315,967 welded STEP-mesh vertices. The same eight graph budgets as AC28 were prepared. All 18 recorded architecture forward/backward checks passed; 10 prediction-model families and 5 diffusion variants completed weak-label diagnostic fitting.

Interpretation: PLY-to-CAD registration and CSV defect extents are not independently validated. Training metrics below measure fit to weak masks on the same physical run. They are not held-out accuracy or evidence of design transfer.

Graph nodes change what the representations retain

Each part is compared with its own finest graph. P95 is sampled source-surface error in mm; six-view image silhouette IoU and surface voxel Jaccard are measured after regenerating each representation from every graph.

Requested nodes	AC28 actual	ME1a actual	AC28 P95 mm	ME1a P95 mm	AC28 image IoU	ME1a image IoU	AC28 voxel Jaccard	ME1a voxel Jaccard	AC28 weak-label transfer IoU	ME1a weak-label transfer IoU
12,000	11,956	12,324	0.075	0.066	1.000	1.000	1.000	1.000	1.000	1.000
6,000	5,925	6,155	0.252	0.198	0.997	0.994	0.983	0.990	0.741	0.333
5,000	4,920	5,132	0.328	0.276	0.996	0.993	0.983	0.987	0.034	0.175
3,000	2,906	3,083	0.613	0.563	0.991	0.988	0.963	0.974	0.278	0.129
1,500	1,393	1,543	1.488	1.403	0.990	0.978	0.931	0.943	0.361	0.079
750	634	764	3.552	2.770	0.964	0.965	0.884	0.926	0.341	0.057
300	329	279	20.535	7.963	0.787	0.918	0.674	0.812	0.326	0.030
100	329	144	20.535	387.356	0.787	0.151	0.674	0.309	0.326	0.017

ME1a at 3,083 nodes: 0.563 mm P95 and 0.988 image IoU. ME1a at 144 nodes: 387.356 mm P95 and 0.151 image IoU. Fine geometry views can remain similar while 10 mm defect masks lose support.

AC28 and ME1a depth views by graph budget

Representation quality versus node count

Weak label and training fit by node count

Representation metrics CSV · All graph-resolution fits CSV · ME1a interactive graph viewer

Source and visual inspection

All six PLY files have finite vertices. The component scans are five views of one run. The merged-height raster matches sampled PLY heights at P90 0.048 mm. The automatic rigid pose has capped 80 mm RMSE 43.08 mm, so residual/deformation values remain diagnostic.

CAD and fused PLY fit with diagnostic targets

Annotation and pose review JSON · All PLY audit JSON · ME1a preprocessing gallery

All model weak-label fits

Ten model families fit the same ME1a diagnostic mask. AP, IoU and Dice are measured on that training run. AC28 columns show its own separate weak-label training run and are included for traceability, not a performance ranking. The constrained GNN's 400-epoch training IoU was 0.036.

Model	Epochs	ME1a loss	ME1a AP	ME1a IoU	ME1a Dice	CAD map coverage	AC28 AP	AC28 IoU
2D CNN/U-Net	80	1.084	0.298	0.143	0.250	0.922	0.622	0.333
FNO2d	60	0.765	0.444	0.143	0.250	0.922	0.620	0.296
Dense 3D U-Net	100	0.592	0.097	0.019	0.037	1.000	0.358	0.048
Sparse 3D U-Net	100	0.672	0.112	0.019	0.037	1.000	0.359	0.047
PointNet	120	1.486	0.076	0.009	0.017	0.997	0.093	0.114
DGCNN	100	1.298	0.038	0.065	0.122	0.630	0.002	0.000
Mesh GNN	100	1.162	0.045	0.021	0.040	1.000	0.758	0.147
Risk GNN	400	1.023	0.253	0.036	0.070	1.000	0.764	0.309
UV-Net	100	1.208	0.011	0.011	0.021	0.966	0.076	0.043
Tabular MLP	120	0.000	0.001	0.000	0.000	1.000	0.002	0.000

All ME1a weak-fit model predictions on one CAD

ME1a model image gallery and checkpoints · All model metrics JSON · Constrained GNN training · Architecture execution report · AC28 model results

Conditional diffusion diagnostics

Four representation-specific denoisers and the TAG Graph U-Net were fitted on the one ME1a run. Deformation MAE uses an unvalidated nearest-scan target. Weak risk AP and IoU are measured on that same run, so these values do not establish generation or prediction quality.

Variant	Nodes/run	Best loss	Deformation MAE mm	Weak risk AP	Weak risk IoU
image_2d	6,000	0.598	5.738	0.143	0.055
voxel_3d	2,592	0.548	6.103	0.235	0.057
point_6000	6,000	0.576	5.645	0.058	0.011
graph_6000	6,155	0.430	5.914	0.177	0.039
graph_unet_6000	6,155	0.738	9.635	0.006	0.001

Diffusion gallery, generated clouds and checkpoints · Five-variant report JSON · AC28 diffusion gallery

Can one GNN learn from both parts?

The residual geometric GNN was compared with the prior Mesh GNN and a coordinate MLP control at 3k, 5k, 6k and 12k graph budgets. The optimized 12k fit uses hard, class-balanced provisional masks. Individual and joint scores are measured on their training runs; held-out scores predict a run excluded from fitting.

Run	Individual AP	Individual IoU	Joint AP	Joint IoU	Held-out AP
AC28_16h	0.385	0.200	0.147	0.145	0.0047
AC28_18h01	0.446	0.136	0.671	0.465	0.0046
ME1a_14h06	0.629	0.476	0.476	0.053	0.0010

Result: ME1a individual fit AP 0.629 / IoU 0.476 exceeds its previous FNO same-run 0.444 / 0.143, but held-out ME1a AP is near chance and IoU is zero. Adding ME1a did not consistently improve held-out AC28 prediction. Node choice: use 12k requested nodes (11,956 AC28; 12,324 ME1a actual) for spatial outputs. ME1a's 5k graph has 5,132 actual nodes and retains only 0.175 of the finest weak mask by IoU; its 6k graph retains 0.333 and 3k retains 0.129. This choice protects localized labels; it is not a validated predictive optimum.

GNN variants and node sizes, joint and held-out results

Individual, pooled and held-out GNN prediction maps

Risk GNN versus every other model · Four-size architecture experiment · Optimized 12k GNN and pooling experiment · All optimized per-run metrics CSV

Prepared inputs

ME1a has six CAD image views at 128²/256², sparse surface voxels at 32/64/128, one 8,192-point CAD set, a separate 8,192-point post-scan diagnostic set, tabular XLS conditions and the trimmed B-Rep face graph. The scan is an outcome used for target construction, not a pre-process predictor.

Images · Voxels · Points · Process table · B-Rep · AC28 inputs · Full written summary

ME1a_DT with AC28: preprocessing and model results