# ME1a_DT and AC28 experiment 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.

## Graph node changes and representation fidelity

Values compare each part with its own finest graph. Surface P95 is sampled source-to-simplified distance; image IoU uses six 128² views; voxel Jaccard uses 32-cell surface occupancy; weak-label transfer IoU maps each graph's 10 mm CSV-seeded mask back to the finest graph.

| Requested | 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 |

The ME1a 100-node request produced 144 nodes, a 387.356 mm surface P95 error and 0.151 image silhouette IoU. The 3,000-node request produced 3,083 nodes, a 0.563 mm surface P95 error and 0.988 silhouette IoU. Small 10 mm weak defect masks lose substantial support even at 3,000 nodes; geometry fidelity alone does not preserve localization labels.

## Scan and annotation inspection

- All six PLY files passed full vertex/bounds checks with no nonfinite vertices. The five component files are views of the same production run, not independent training samples. Sample-to-fused P95 distances span 1.214–4.610 mm.
- The inferred merged-height raster-to-PLY mapping has 0.048 mm height P90 error on sampled scan points. All six CSV anchors are within 30 mm of a CAD node after a diagnostic proper-rigid fit.
- The fitted pose has capped 80 mm RMSE 43.08 mm and is not independently validated. The six CSV seeds produce 13 positive nodes at an assumed 10 mm geodesic radius. All `training_valid` masks remain false.

## Model execution and weak-label fits

- 18 forward/backward architecture checks passed on ME1a inputs; 13 synthetic checks/demos are separate from real labels.
- The constrained risk/deformation GNN ran 400 weak-fit epochs: loss 1.884 → 1.023; same-run threshold IoU 0.036, Dice 0.070. Its defect-type head cannot learn from one run-level class.
- All ten feasible model families were fitted to the same ME1a weak target. Values below are same-run training diagnostics and cannot rank production quality or be compared as held-out AC28 vs ME1a accuracy.

| Model | Epochs | ME1a loss | ME1a AP | ME1a IoU | ME1a Dice | CAD mapping 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 |

## Conditional diffusion variants

Four representation-specific diffusion denoisers and a paper-inspired TAG Graph U-Net were also fitted on the one ME1a run. Deformation MAE uses a nearest-scan proxy under unvalidated registration; risk AP/IoU are same-run weak-label diagnostics.

| Variant | Nodes/run | Best loss | ME1a 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 |

## Joint AC28 + ME1a GNN experiment

The residual geometric GNN was compared with the prior Mesh GNN and a position-aware MLP control at 3k, 5k, 6k and 12k requested nodes. A second 12k fit used hard, class-balanced weak masks for 180 epochs. The numbers below are weak-label diagnostics, not verified prediction accuracy.

| Run | Individual AP | Individual IoU | Joint AP | Joint IoU | Held-out AP | Held-out IoU |
|---|---:|---:|---:|---:|---:|---:|
| AC28_16h | 0.385 | 0.200 | 0.147 | 0.145 | 0.005 | 0.000 |
| AC28_18h01 | 0.446 | 0.136 | 0.671 | 0.465 | 0.005 | 0.000 |
| ME1a_14h06 | 0.629 | 0.476 | 0.476 | 0.053 | 0.001 | 0.000 |

ME1a individual-fit AP reached 0.629 and IoU 0.476, exceeding the earlier FNO same-run AP 0.444 and IoU 0.143. Its held-out AP stayed near its 0.001 prevalence and IoU was zero. Adding ME1a to AC28 holdout training did not consistently improve either AC28 run. The GNN fits ME1a's weak training mask more closely. Cross-part prediction remains unproven.

For localized risk output, fix the current graph at **12,000 requested nodes** (AC28 11,956; ME1a 12,324 actual). The 5k ME1a graph has 5,132 actual nodes and transfers only 0.175 of the finest weak mask by IoU; its 6k graph transfers 0.333 and 3k transfers 0.129. Lower-node graphs can be context inputs, but should not define the spatial output grid. This is a representability choice, not a statistically validated predictive optimum.

- [Four-size architecture/pooled-fit comparison](gnn_joint/index.html) and [full CSV](gnn_joint/report.csv)
- [Optimized 12k residual GNN results](gnn_joint_hard/index.html), [prediction maps](gnn_joint_hard/prediction_maps.png), and [full CSV](gnn_joint_hard/report.csv)


## Open results

- [Interactive report](index.html)
- [All risk models compared on AC28 and ME1a](model_ranking/index.html) and [full metrics CSV](model_ranking/report.csv)
- [ME1a preprocessing gallery](../../experiments/ME1a_DT/data_processed/index.html)
- [ME1a graph viewer](../../experiments/ME1a_DT/results/graphs/viewer.html)
- [ME1a all-model training gallery](../../experiments/ME1a_DT/results/models/cad_predictions/index.html)
- [ME1a weak GNN gallery](../../experiments/ME1a_DT/results/models/weak_gnn/index.html)
- [ME1a conditional diffusion gallery](../../experiments/ME1a_DT/gen_models/outputs/index.html) and [report JSON](../../experiments/ME1a_DT/gen_models/outputs/report.json)
- [Graph representation CSV](representation_sensitivity.csv) and [per-resolution training CSV](graph_fit/report.csv)
- [AC28 preprocessing gallery](../../data_processed/index.html) and [AC28 all-model training gallery](../models/cad_predictions/index.html)

The fused PLY is an outcome for target construction, not an input to pre-process prediction. Missing independently verified registration, reviewed spatial labels, OK examples and held-out production runs prevent a valid prediction-accuracy claim.