ajYdZddlmZddlZddlZddlmZeejjdZ e dz Z e dz Z d dZ dd Zddd Zdd Zed k(reyy)zz)htmlescape)r labels r linkrs, $++d:K9LBt{{[`OaNbbf.g'gr c|d|dS)Nz,.f)valuedigitss r numberrsr&l8K/Lr c(+ttdz d}ttdz d}ttdz }ttdz }ttdz }ttdz }ttdz }ttdz d }ttd z d }ttd z } ttd z } ttd z } ttdz } ttdz } d}g}|D]+t +fd|D}t +fd|D}t +fd|D}t +fd|D}|j d+dd|ddd|dddt |ddt |ddt |ddt |ddt |ddt |ddt |ddt |ddg}|djD]\}}|dj|}|j dtj|d|d dt |d!dt |d"dt |d#dt |d$dt |d%d|rt |d"nd&d|rt |d#nd&dg}|d'D]~}|d(d)}|j dtj|d*d|d+ddt |d,dt |d-dt |d.d/dt |d.d0d d1}d2}g}|D]}|||d3|}|||d4}|||d5|} |j d|dt |d6dt |d7dt |d6dt |d7dt | d6d8dt | d7d| d D!cgc]}!|!d9 }"}!td:|dD}#d;| d<d=| d>d?| d@dAddB| dCddD| dEddF|#dGt|ddHt|d'dI}$dJdK|$dKdLdKdMdKdNdOg }%|D]+t +fdP|D}t +fdQ|D}t +fdR|D}t +fdS|D}|%j dTdUj+d|dd|ddt |dt |dt |dt |dt |dt |dt |dt |df zdVz|%dKdWdKdXdKdYt|"dZd[t|"dZd\d]| d^d_dZd`da| dbdcdddedf| dgdhdKdidKdj|#dkt|dldmdn|dodZdp|d!dZdq|drdZds|dtdZdu dvdKdwdxgz }%|djD]\}}|dj|}&|%j dTdUj|t!|d t |d!t |d"t |d#t |d$t |d%|&rt |&d"nd&|&rt |&d#nd&f zdVz|%gdyz }%|d'D]t}|d(d)}|%j dTdUj|d*|d+dt |d,t |d-t |d.d/t |d.d0fzdVzv|%gdzz }%|D]r}|||d3||||d4|||d5|g}'|%j dTdUj|g|'D!(cgc]}!d{D]}(t |!|(c}(}!zdVzt|%gd|z }%|%gd}z }%td~z j#dj|%dd})dKjgd|)dtj|$d| dEdd| ddcddd| dCdd| d@dAdd|#dt|ddt|d'ddKj|dt%dddt%dddt%ddd| d^d_dZd| dbdcdddedt%dddt%dddt%ddd|drdZddKj|dt%dddt%dddt%dddt%dddt%ddddKj|dt%dddt%dddt%ddddKj|dt%dddt%dddt%dddt%dddt%dddt%dd«dt%ddīdt%ddƫdt%ddȫdt%ddʫdt%d~d˫d̑}*tdz j#|*dt'tdz ycc}!wcc}(}!w)Nzrepresentation_sensitivity.jsonrecordszgraph_fit/report.jsonz*results/models/cad_predictions/report.jsonzresults/models/report.jsonz#results/models/weak_gnn/report.jsonzgen_models/outputs/report.jsonzgnn_joint/report.jsonrowszgnn_joint_hard/report.jsonz%results/inspection/target_review.jsonz'results/inspection/component_audit.jsonz%analysis/results/source_manifest.jsonz!data_processed/brep/metadata.jsonz data_processed/graphs/index.json).ipii iii,dc3FK|]}|ddk(s |dk(s|ywpartAC28requested_nodesNr.0rbudgets r zmain..&*T#Q6F!2q9J7KV7Sq# !!!c3FK|]}|ddk(s |dk(s|ywr ME1a_DTr"Nrr#s r r'zmain..'+W#Q6I!5!.(+[9a& 6(9a@Q>RTZ>Z9r)c3FK|]}|ddk(s |dk(s|ywr+rr#s r r'zmain..)+^9a& 9(<CTAUW]A]9r)z,z actual_nodesz source_to_graph_p95_mmimage_silhouette_iou_128surface_voxel_jaccard_32'weak_label_transfer_iou_on_finest_graphz modelsepochs loss_finalaverage_precision threshold_iouthreshold_dicedisplay_coverageu—variantsmetricsvariant nodes_per_run best_lossMAE_mmriskAPz IoU_at_0.5c0tfd|DS)Nc3fK|](}|ddk(s |ddk(s|dk(r |dk(r|*yw)r&rkind residual_geoeval_runsetupNr)r$r%runrMs r r'z*main..gnn_score..AsEBt!q{E'9ai>W* s*qz5/@ts 111)next)rrNrMs ``r gnn_scorezmain..gnn_score@s BtBB Br )AC28_16h AC28_18h01 ME1a_14h06single_jointholdout_ mapped_ap mapped_iousample_to_fused_p95_mmc3,K|] }|ddk(yw)statusforward_backward_smokeNr)r$r%s r r'zmain..NsR/Q1X; 88/sz ME1a used process_rowsz XLS process rows, defect_rowsz CSV defect marks, fused_check vertex_countz fused PLY vertices, face_countz B-Rep faces and source_verticeszT welded STEP-mesh vertices. The same eight graph budgets as AC28 were prepared. All z7 recorded architecture forward/backward checks passed; z prediction-model families and z< diffusion variants completed weak-label diagnostic fitting.z%# ME1a_DT and AC28 experiment resultsz1## Graph node changes and representation fidelityuValues 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.z| 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 |z8|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|c3FK|]}|ddk(s |dk(s|ywrrr#s r r'zmain..]r(r)c3FK|]}|ddk(s |dk(s|ywr+rr#s r r'zmain..^r-r)c3FK|]}|ddk(s |dk(s|ywrrr#s r r'zmain.._r/r)c3FK|]}|ddk(s |dk(s|ywr+rr#s r r'zmain..`r1r)z| z | z |a[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.z!## Scan and annotation inspectionz- 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 z.3fu–z mm.z7- The inferred merged-height raster-to-PLY mapping has raster_mapping_checkheight_abs_p90_mmz 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.z(- The fitted pose has capped 80 mm RMSE registration_candidatesrcapped_rmse80_mmz.2fzB mm and is not independently validated. The six CSV seeds produce weak_10mm_positive_nodesz] positive nodes at an assumed 10 mm geodesic radius. All `training_valid` masks remain false.z&## Model execution and weak-label fitsz- z= forward/backward architecture checks passed on ME1a inputs; syntheticz6 synthetic checks/demos are separate from real labels.zE- The constrained risk/deformation GNN ran 400 weak-fit epochs: loss loss_initialu → z; same-run threshold IoU training_threshold_iouz, Dice training_threshold_dicez=. Its defect-type head cannot learn from one run-level class.z- 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.zk| Model | Epochs | ME1a loss | ME1a AP | ME1a IoU | ME1a Dice | CAD mapping coverage | AC28 AP | AC28 IoU |z-|---|---:|---:|---:|---:|---:|---:|---:|---:|)rdz!## Conditional diffusion variantsrdzFour 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.rdz\| Variant | Nodes/run | Best loss | ME1a deformation MAE mm | Weak risk AP | Weak risk IoU |z|---|---:|---:|---:|---:|---:|)rdz### Joint AC28 + ME1a GNN experimentrdaThe 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.rdz\| Run | Individual AP | Individual IoU | Joint AP | Joint IoU | Held-out AP | Held-out IoU |z#|---|---:|---:|---:|---:|---:|---:|)rWrX)rdaaME1a 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.rdaFor 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.rdzk- [Four-size architecture/pooled-fit comparison](gnn_joint/index.html) and [full CSV](gnn_joint/report.csv)z- [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)rd)rdz## Open resultsrdz"- [Interactive report](index.html)zx- [All risk models compared on AC28 and ME1a](model_ranking/index.html) and [full metrics CSV](model_ranking/report.csv)zS- [ME1a preprocessing gallery](../../experiments/ME1a_DT/data_processed/index.html)zK- [ME1a graph viewer](../../experiments/ME1a_DT/results/graphs/viewer.html)zh- [ME1a all-model training gallery](../../experiments/ME1a_DT/results/models/cad_predictions/index.html)zW- [ME1a weak GNN gallery](../../experiments/ME1a_DT/results/models/weak_gnn/index.html)z- [ME1a conditional diffusion gallery](../../experiments/ME1a_DT/gen_models/outputs/index.html) and [report JSON](../../experiments/ME1a_DT/gen_models/outputs/report.json)zt- [Graph representation CSV](representation_sensitivity.csv) and [per-resolution training CSV](graph_fit/report.csv)z- [AC28 preprocessing gallery](../../data_processed/index.html) and [AC28 all-model training gallery](../models/cad_predictions/index.html)rdzThe 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.rdz README.md zutf-8)encodingabody{font:15px/1.5 system-ui,sans-serif;background:#f5f7fb;color:#1a2532;max-width:1450px;margin:auto;padding:24px}h1,h2{line-height:1.2}a{color:#125a93}section{background:white;border:1px solid #dce3ec;border-radius:12px;padding:20px;margin:20px 0}.lead{font-size:18px}.warning{background:#fff2cf;border-left:5px solid #d29a13;padding:14px}.scroll{overflow:auto}table{border-collapse:collapse;min-width:850px;width:100%}th,td{border-bottom:1px solid #e1e6ec;padding:8px;text-align:right;white-space:nowrap}th:first-child{text-align:left}thead th{background:#edf3f8}.grid{display:grid;grid-template-columns:repeat(auto-fit,minmax(310px,1fr));gap:14px}.grid img,.wide{width:100%;border:1px solid #e0e6ed;border-radius:6px}.kpis{display:flex;flex-wrap:wrap;gap:12px}.kpis span{background:#eaf1f7;border-radius:8px;padding:10px 14px}small{color:#546474}@media(max-width:700px){body{padding:12px}.grid{display:block}}zME1a_DT + AC28 results

ME1a_DT with AC28: preprocessing and model results

a2

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.

z' STEP mesh vertices resolutionsz' finest graph nodesz B-Rep facesz' fused PLY verticesz( architecture checksz. prediction-model familiesa diffusion variants

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.

a
Requested nodesAC28 actualME1a actualAC28 P95 mmME1a P95 mmAC28 image IoUME1a image IoUAC28 voxel JaccardME1a voxel JaccardAC28 weak-label transfer IoUME1a weak-label transfer IoU

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 budgetRepresentation quality versus node countWeak label and training fit by node count

zrepresentation_sensitivity.csvzRepresentation metrics CSVu · zgraph_fit/report.csvzAll graph-resolution fits CSVz4../../experiments/ME1a_DT/results/graphs/viewer.htmlzME1a interactive graph viewerz

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 z4 mm. The automatic rigid pose has capped 80 mm RMSE a mm, so residual/deformation values remain diagnostic.

Six CSV defect marks on merged imageFused and component PLY scansCAD and fused PLY fit with diagnostic targets

z?../../experiments/ME1a_DT/results/inspection/target_review.jsonzAnnotation and pose review JSONzA../../experiments/ME1a_DT/results/inspection/component_audit.jsonzAll PLY audit JSONz3../../experiments/ME1a_DT/data_processed/index.htmlzME1a preprocessing galleryaQ

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 z.

z
ModelEpochsME1a lossME1a APME1a IoUME1a DiceCAD map coverageAC28 APAC28 IoU
All ME1a weak-fit model predictions on one CAD

zC../../experiments/ME1a_DT/results/models/cad_predictions/index.htmlz(ME1a model image gallery and checkpointszD../../experiments/ME1a_DT/results/models/cad_predictions/report.jsonzAll model metrics JSONz<../../experiments/ME1a_DT/results/models/weak_gnn/index.htmlzConstrained GNN trainingz4../../experiments/ME1a_DT/results/models/report.jsonzArchitecture execution reportz$../models/cad_predictions/index.htmlzAC28 model resultsa 

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.

a6
VariantNodes/runBest lossDeformation MAE mmWeak risk APWeak risk IoU
Four ME1a diffusion variantsGraph baseline and TAG U-Net on ME1a

z7../../experiments/ME1a_DT/gen_models/outputs/index.htmlz3Diffusion gallery, generated clouds and checkpointsz8../../experiments/ME1a_DT/gen_models/outputs/report.jsonzFive-variant report JSONz#../../gen_models/outputs/index.htmlzAC28 diffusion galleryaC

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.

a9
RunIndividual APIndividual IoUJoint APJoint IoUHeld-out APHeld-out IoU

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 resultsIndividual, pooled and held-out GNN prediction maps

zmodel_ranking/index.htmlz!Risk GNN versus every other modelzgnn_joint/index.htmlz!Four-size architecture experimentzgnn_joint_hard/index.htmlz(Optimized 12k GNN and pooling experimentzgnn_joint_hard/report.csvz!All optimized per-run metrics CSVu_

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.

z:../../experiments/ME1a_DT/data_processed/images/index.htmlImagesz:../../experiments/ME1a_DT/data_processed/voxels/index.htmlVoxelsz:../../experiments/ME1a_DT/data_processed/points/index.htmlPointsz;../../experiments/ME1a_DT/data_processed/tabular/index.htmlz Process tablez8../../experiments/ME1a_DT/data_processed/brep/index.htmlzB-Repz../../data_processed/index.htmlz AC28 inputszFull written summaryz

z index.html)r OUTMEPROJECTrOappendritemsgetrrsumlenjoinminmaxstr write_textrprint),rep graph_fittrained ac28_trainedsmokeweak diffusiongnn_softgnn_hardtargetscanmanifestbrepgraphsbudgets graph_rowsamfafm model_rowsnamerowbaselinediffusion_rowsmetricrPgnn_runsgnn_rowsrNsinglerUheldr%p95 smoke_passsummarymdbasevalueskeycsspager&s, @r mainrs( S2 23I>C3../ :I @@ AGgJJKL r.. /E b66 7D2667I #-- .v 6H #22 3F ;H :: ;F b:: ;D "<< =H b44 5D 55 6F 3GJ T#T T W#W W [9[ [ ^9^ ^HVAJi.8I!7LIVWXfVghiUjk!4567yJbHcAd@ef!67896!LfJgChBij!67896!LfJgChBij"FGHIJ"FGHI U V JH%++-Sh'++D1HT[[%6$7yXP#l+,-YvcBU>V7W6XY#o./0 &EUAV:W9XY#01234 21 55J4P\K]^_J`a'(+,lQ wx ! ""A#iPZF[B\A]^Z  [  072;B \]_ ZB DB  T#T T W#W W [9[ [ ^9^ ^ $uzzfQZAn4Ea3HQ~M^_`La 1-. /q9Q7R0S 1/0 1&;U9V2W 1/0 1&;U9V2W 2?@ A 2?@ A #CDD EI I J 2lmo+B `adehaijm`nnqruvyrz{~qC D A&I_B`atBuvyAz{ @ 26:S3TUV3WXj3klo2pqstz{UtVsWWt uvx0 ZLUVYZ_`kZlVmUnoe f OPTUcPdehOiinosuApBCFoGG`aef~a@C`DDKLPQjLkloKppm n Z[]u7 9 9BH%++-S ( # ' ' - $uzz4CM(:6#lBS;T 3*+ ,VC4H-I 3'( )&5G1H*I15F4+, -5-1F4( )u #>??@D D E. **B$9~b! $uzz3y>S5I!4LfUXYdUeNf 6(# $VF6N4,@%A&PVXdIeBf#hiiimn o% //B(37 (3w8W(3#'78: $uzz3"o(nASmC#Sm(n"oppquuv rrB E EB _  2 @ `C h h r h svrw h wL h LPKKX_L`Ka h b h ##45a8 h 9` h aggt`uvw`xzHaIJK`L h Ls h txyEtFGHsI h Ii h jrs@jABPjQRSiT h T{ h |F{G h Go h pst{|DtEpFoG h Gu h vyzCDNzOvPuQ h Qu  h v x v } v } ~ H v I u J h J ] h ^bcCD`^a]b h bf h gklBCbgcfd h dh h imndeDiEhF h FE h FLMcFdexFyz}E~ h ~r h syzSsTUVsWXjsklorp h py h z ~  @Abz cy d h dh h imnqrFiGhH h HL h MQRGHdMeLf h fC h DHIaDbcfCg h gB h C E C J C J K U C V B W h W E h F J K P Q { F | E } h } A h BFGMNfBgAh h hl h mqrpqKmLlM h MQ h RVWMNmRnQo h os h txy_`ttusv h v~ h AFFGUV~W h WM  h N R S LMBN CM D h DH h IMNHIcIdHe h ei h jnoTUmjnio h ou h vxv}v}~F vG uH  h H A h BFGabEBFAG h GK h LPQghKLLKM h MQ h RVWrs]R^Q_ h _c h dhiDEhdicj h jQ h RVWST\R]Q^ h ^b h cghdemcnbo h os h txyuv~ts@  h @ D  h E I J G H W E X D Y  h Y ]  h ^ b c ] ^ e ^ f ] g  h g k  h l p q R S ` l a k b  h b f  h g k l w x N g O f P  h P e  h D !!$!8 #l s rsjB"  X ((+   ((9gLSl Zr