Retrained with 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. Settings, input arrays, labels and seed are unchanged. Weak-label agreement is not independently reviewed accuracy.

Comparison CSV · Exclusion and checkpoint audit · Full results

Model	Protocol	Before BCE	After BCE	Before AP	After AP	Before IoU	After IoU
cnn2d_unet	leave_one_design_out	0.1983	0.2280	0.1219	0.0979	0.0000	0.0000
cnn2d_unet	train_all	0.0599	0.0705	0.8680	0.8441	0.4876	0.2326
dgcnn	leave_one_design_out	0.1948	0.1925	0.0673	0.0666	0.0000	0.0000
dgcnn	train_all	0.1715	0.1720	0.0631	0.0624	0.0000	0.0000
fno2d	leave_one_design_out	0.3462	0.3211	0.0420	0.0377	0.0000	0.0000
fno2d	train_all	0.0746	0.0912	0.8543	0.7726	0.0727	0.0294
gaussian_process	leave_one_design_out	0.4651	0.4530	0.1001	0.0745	0.0000	0.0000
gaussian_process	train_all	0.1723	0.1292	0.5517	0.7332	0.3694	0.4473
meshgraphnet	leave_one_design_out	0.1894	0.1874	0.0810	0.0808	0.0000	0.0000
meshgraphnet	train_all	0.1714	0.1725	0.0940	0.0921	0.0000	0.0000
physics_risk_gnn	leave_one_design_out	0.1894	0.1879	0.0919	0.0886	0.0000	0.0000
physics_risk_gnn	train_all	0.1708	0.1719	0.0992	0.0975	0.0000	0.0000
pointnet	leave_one_design_out	0.1944	0.1920	0.0561	0.0575	0.0000	0.0000
pointnet	train_all	0.1705	0.1720	0.1009	0.0886	0.0000	0.0000
tabular_mlp	leave_one_design_out	0.1968	0.1945	0.0779	0.0784	0.0000	0.0000
tabular_mlp	train_all	0.1670	0.1692	0.0734	0.0711	0.0000	0.0000
unet3d	leave_one_design_out	0.3564	0.3507	0.0358	0.0348	0.0000	0.0000
unet3d	train_all	0.0370	0.0427	0.9222	0.9093	0.7824	0.7565
uvnet	leave_one_design_out	0.2125	0.2129	0.0566	0.0562	0.0000	0.0000
uvnet	train_all	0.1604	0.1691	0.0906	0.0950	0.0000	0.0000