120 epochs on six datasets

Excluded: AC28_DT_RR_UPR_STD_1_20260827_18_01__NG_7194c0e36c7b

6 datasets / 1,227 weak label targets. Both before/after columns evaluate the same six retained datasets. Epochs increased from 30 to 120; all other model settings, input arrays, labels, exclusion and seed are unchanged. Identical ordered training folds reuse an exact fit within this run, with separate checkpoints and predictions. Gaussian Process is refitted with its unchanged non-iterative procedure. Weak-label agreement is not independently reviewed accuracy.

Comparison CSV · Exclusion and checkpoint audit · Full results

Full 120-epoch learning curves

Model	Protocol	30 epoch BCE	120 epoch BCE	30 epoch AP	120 epoch AP	30 epoch IoU	120 epoch IoU
cnn2d_unet	leave_one_design_out	0.2280	0.4378	0.0979	0.1213	0.0000	0.0000
cnn2d_unet	train_all	0.0705	0.0211	0.8441	0.9451	0.2326	0.8667
dgcnn	leave_one_design_out	0.1925	0.1965	0.0666	0.0503	0.0000	0.0000
dgcnn	train_all	0.1720	0.1529	0.0624	0.1654	0.0000	0.0000
fno2d	leave_one_design_out	0.3211	0.7015	0.0377	0.0434	0.0000	0.0000
fno2d	train_all	0.0912	0.0258	0.7726	0.9411	0.0294	0.8622
gaussian_process	leave_one_design_out	0.4530	0.4530	0.0745	0.0745	0.0000	0.0000
gaussian_process	train_all	0.1292	0.1292	0.7332	0.7332	0.4473	0.4473
meshgraphnet	leave_one_design_out	0.1874	0.2085	0.0808	0.0690	0.0000	0.0000
meshgraphnet	train_all	0.1725	0.1562	0.0921	0.0916	0.0000	0.0000
physics_risk_gnn	leave_one_design_out	0.1879	0.2060	0.0886	0.0372	0.0000	0.0000
physics_risk_gnn	train_all	0.1719	0.1535	0.0975	0.1393	0.0000	0.0000
pointnet	leave_one_design_out	0.1920	0.1967	0.0575	0.0517	0.0000	0.0000
pointnet	train_all	0.1720	0.1601	0.0886	0.0778	0.0000	0.0000
tabular_mlp	leave_one_design_out	0.1945	0.1911	0.0784	0.0501	0.0000	0.0000
tabular_mlp	train_all	0.1692	0.1576	0.0711	0.0815	0.0000	0.0000
unet3d	leave_one_design_out	0.3507	0.4631	0.0348	0.0367	0.0000	0.0000
unet3d	train_all	0.0427	0.0276	0.9093	0.9333	0.7565	0.8093
uvnet	leave_one_design_out	0.2129	0.2043	0.0562	0.2203	0.0000	0.0889
uvnet	train_all	0.1691	0.1542	0.0950	0.1476	0.0000	0.0000

UV-Net held-out detail

Macro IoU improved from 0 to 0.0889 at threshold 0.5. Only 2 of 49 weak positive nodes were detected, with 5 false positives. Both true detections occur in AC28 scans; performance remains limited. These targets are automatic weak projections.

Per-dataset evidence

Dataset	IoU	TP	FP	FN
AC28_DT_RR_UPR_STD_20260827_16__NG_008adde420f0	0.3333	1	0	2
AC28_DT_RR_UPR_STD_20260828_09_21__NG_f4c68f38e467	0.2000	1	2	2
AC28_DT_RR_UPR_STD_2_20260827_18_01__NG_8df93616f7ea	0.0000	0	0	3
ME1a_DT_FRT_UPR_LHD_20260608_14_06__NG_ffc4fce5d996	0.0000	0	0	16
MX5a_PE_DT_RR_AREST_1_20260626_13_08__NG_fe0e1b9b8b53	0.0000	0	0	12
MX5a_PE_DT_RR_AREST_2_20260626_13_08__NG_41ec5d7a218e	0.0000	0	3	12