jp dZddlmZddlZddlZddlmZddlZejdddl m Z ddl Z ddlmZedz Zd9d Zd:d Zgd Zd;d Zdd2Zd?d3Zd4d5 d@d6ZdAd7Zed8k(reeyy)Bz>Build a model-by-model input, fusion, and execution dashboard.) annotationsN)PathAgg)ROOTzresults/models/dashboardc ttdz jd}g}|D]}tj|d5}|j t |djtj|d|dgtj|d|d gd dddtj|Dcgc]}|d jc}}|tj|}tjt|d d dt|zfd\}}tj|}t!||D]a\} }| j#|d dd|j%|j'} t)dD]} t)dD]} |d | | f} tj| sdn| d|d| | frdnd}tj| r| tj*|kDrdnd}| j-| | |ddd|| j/d d gd!d"g| j1t)dt)d d#D cgc]} | d$c} | j3d%| j5|d&d'|j7| | d(d)d*+d|j9t:d,z d-.tj<|y#1swYxYwcc}wcc} w)/Nzdata_processed/pairedz*_cad_ply_targets.npzF) allow_picklerun_id heater_top_cheater_bottom_cheater_top_imputedheater_bottom_imputed)runvaluesimputedrrg@Tfigsizeconstrained_layoutautoinferno)aspectcmapvminvmax—z.0fr*blackwhitecenter)havafontsizecolorrupperlower02dz heater zoneru2 XLS heater input (°C; * = linearly interpolated)u°Cg?gQ?)axlabelfractionpadthermal-process-input.pngdpi)sortedrglobnploadappendstritemstack concatenateravelisfinitepltsubplotslen atleast_1dzipimshowminmaxrange nanmediantext set_yticks set_xticks set_xlabel set_titlecolorbarsavefigOUTclose)pairedrowspathdatarowrfinitefigaxesr-imagelayerzonevaluer.r(s models/build_dashboard.pythermal_previewr`s T3399:QR SF D WWT .$ KK4>..01((D$8$?P:Q#RS88T*>%?F]A^$_` / .^^dCdsS]002dC DF BKK' (F SYBc$i3H]abIC == DtT?C #h-YVZZ\`f`j`j`l m1XEb H eTk2%'[[%7c{RUV_R`afhlalRm3suJv=w#%;;u#5%",,vBV:V\ceUxHqX]^ " q!fw01 eBiE!RL!ILDT#J-L!IJ m$bllc%j\As3t&u Ur4 H#KK11sK;IIcN// . D"JsA&K%K2" K7 %K/ c i}|dD]&}|j|dgj|(i}|jD]\}}|Dcgc] }|ddk(s |}}|Dcgc]}|d }}|Dcgc]}|d }}|Dcgc]%}|jds|jd'} }|s|ddnd t |t |Dchc]}|jd c}|sd n2t |d k(r |dd dnt |d dt|d d|| |djddd||<|Scc}wcc}wcc}wcc}wcc}w)Nmodelsfamilystatusforward_backward_smokeforward_latency_ms_cpu output_shapesourcerzforward/backward passedparams unavailablerz.1f msu–reasonr )rdchecksrilatencyshapessourcesrl) setdefaultr9itemsgetrBr5rFrG) reportgroupedrWresultrcrTpassedrnrorps r_ _summariesrx0s%'Gh3x="-44S9 F  !%S#X:R)R#Srrc3JK|]}d|jddvs|yw)cad_surface_pointsrhr Nrs).0rWs r_ z"comparison_rows..Ys)hJS:NRURYRYZbdfRg:gSJs##rrc&|jddS)N actual_nodesrr)rWs r_z!comparison_rows..[ssww~q7Q)key)NNNdg-q=rrdrirf loss_finaltraining_threshold_ioutraining_threshold_dice) modelrdrirnfusiondomainsynthetic_initialsynthetic_finalsynthetic_reduction weak_lossioudice)COMPARISON_MODELSnextrGrsabsr9)rtweakrWractualoutputdisplayrc synth_namerr candidatesselectedsynthinitialfinal reductiontraineds r_comparison_rowsrRs/5k/BV/BhRUoX#/BIV H F F7H3VV%+GVcs8}/FcV G K KhJhjtuvjwxH / /:+QRH!!}H-7 j)T$4! "8_NGUw/#c'lE2JJI@@ (:hllS[F\||$<=[a!(U[d/6l+D5<401$7>D23D   8I, M3WHs DD D#D#c\|Dcgc]}|d }}tjt|}tjdddd\}}tj |Dcgc]}|dr|dntj c}t}tj |Dcgc]}|d|dntj c}t}tj |Dcgc]}|d |d ntj c}t}|d j||d |d jd |d jd|d jd|dj|tj|dd |djd |djd|djd|dj||d |djd|djd|D]7} | j||| j| jdd9|dj!d dd|j#d|j%t&dz d !tj(|ycc}wcc}wcc}wcc}w)"Nrr)Trrirnrrz#4472c4)r(log Parameterszcount (log scale)g{Gz?z#ed7d31zActual-input CPU forwardzmilliseconds (log scale)rz#2f9e75zTwo-step synthetic loss changez reduction (%)xg?)axisalphaz#555g?)r( linewidthuEArchitecture comparison — execution and synthetic optimization onlyzmodel-comparison.pngr2r3)r7arangerBr@rAasarraynanfloatbarh set_xscalerNrMmaximumrK invert_yaxisgridaxvlinesuptitlerPrQrR) rTrWnamesyrYrZrirnrr-s r_comparison_previewros%) *TcS\TE * #d)A Q7tLIC ZZtTt#h-XRVVCtTV[ \Fjj^bc^bWZC N,F#i.BFFR^bcejkG x|}x|qt#>S:T:`C 56fhflfllx|}@EFIGLLF)L,d1g.@.@.GaIZIZ[gIhjnopjqj|j|~QkRGLLBJJw,IL>Q@R@RSX@Y[_`a[b[l[lnH\IKOPQKRK]K]^xKyGLLIYL/a1B1BCc1dfjklfmfxfxzIgJ a!2BGGCG4PGOOAVrO2LLXYKK,,#K6IIcN +Uc}s J !J !J$ !J)cg}|D]}|ddn|ddd|ddd|ddd }|d dnd |d d d |ddd|dd}|ddn|ddd}|ddn|dd}|jdtj|ddtj|ddtj|dd|d|d|d|ddj|S)Nrrz.5g → rz (rz+.1fz%)rzloss z.4fz; IoU r.3fz; Dice rrnrkri,zrz rz rz r )r9htmlescapejoin)rTbodyrWrrrnris r_comparison_tablers D !45=EcJ]F^_bEcchilm~iADiEEGHKLaHbcgGhhjDk K(0uc+>Ns=SSYZ]^cZdehYiipqtu{q|~AqB7Cy>1%#i.9MS7Qh-/H a7H ht{{3w<894;;sS[}C]B^^ghlhshstwyAuBiChDDMNTMUU^_f^ggpqzp{{DEIDJJTU V  774=r)uimage input|34×128×128uConv2d ×2|34→16→16uMaxPool2d|128→64uConv2d ×2|16→32→32)process input|112uLinear|112→64zsplit|scale 32 + bias 32)uFiLM bottleneck|32×64×64uConvTranspose2d|32→16zconcat skip|16+16uConv2d ×2|32→16→16u1×1 head|16→1 pixel)geometryprocessrufield input|34×128×128u1×1 lift|34→24)ruLinear|112→48zsplit|scale 24 + bias 24)uFiLM lifted field|24×128×128uspectral + skip ×3|24 channelsu1×1 head|24→1 pixel)uvoxel input|24×32×4×16uConv3d ×2|24→8→8uMaxPool3d|16×2×8uConv3d ×2|8→16→16)ruLinear|112→32zsplit|scale 16 + bias 16)uFiLM bottleneck|16×16×2×8uConvTranspose3d|16→8zconcat skip|8+8uConv3d ×2|16→8→8u1×1 head|8→1 voxel)z sparse input|403 coords, 24 featuMinkowskiConv|24→16ustride-2 down|16→32umid Conv|32→32zXLS branch|not wired)utranspose Conv|32→16zconcat sparse skip|16+16usparse head|32→1 cell)upoint input|8,192×Cushared Conv1d|C→64→128zglobal max|128zrepeat + concat|128+128zXLS MLP|proposed, not wiredupoint head|256→64→1zoutput|one scalar / point)upoint input|512×Czdynamic kNN|k=12uEdge MLP 1|2C→32→32u!new kNN + Edge MLP 2|64→32→32zlocal + local + global|32+32+32upoint head|96→32→1)unode input|N×32unode encoder|32→48uedge input|E×8zglobal input|112uLinear + GELU|112→48)u%edge message ×3|48+48+8+48→48→48usum at destination|N×48uGRU update ×3|96→48unode head|48→1)unode input|2×11,956×32unode encoder|32→32uedge input|2×72,176×8)uweak-train input|2×150uLinear + GELU|150→32usmoke alternative|2×112→32)u%edge message ×2|32+32+8+32→32→32uGRU update ×2|64→32uraw head|32→(3 + classes)z"node heads|risk, signed, magnitudezmean pool|run defect type)uface UV grids|6,564×12×8×8uConv2d ×2|12→8→8umean UV pool|6,564×8zface edges|36,090)uface message|8+8→8usum at destination|6,564×8uface head|8→1zCAD summary|6 values + maskszXLS process|50 values + maskszconcatenate|112 total)uLinear + GELU|112→64uLinear + GELU|64→64urun head|64→1) 2D CNN / U-Netr}rrPointNet segmentationDGCNN segmentationMeshGraphNet-style GNNPhysics-constrained risk GNNUV-Net-style face modelTabular MLP surrogatec g}t|D]k\}}|jdd\}}|r|jd|jdtj|dtj|dmdj |S)N|ru$→z
zz
r ) enumeratesplitr9rrr)rrrindexr;r.shapes r_ _layer_chainrs F ' tzz#q) u  MM@ A 3DKK4F3GGWX\XcXcdiXjWkkyz{ ( 776?rc t|}dtj|dt|ddt|ddt|dd S) Nz@
Geometry
rzY
Process
rz`
Fusion / output
rz
) STRUCTURESrrr)rrTs r_ structurers| e DOPTP[P[\aPbOcdFFRSWXbScFdEefSS_`den`oSpRqrZZfgkltguZvYwx  rr )notec tjtj| j dgd} dj d| j dgDxsd} tj| j dd }| j d rd tj| d d nd }| rdtj| d nd }d j gdtj|dtj|dtj|d|d|d|dtj|dtj|dtj|dtj|dtj|dtj|dt |dtj| d| j ddd| dtj| j d dd!| d"||d#S)$NroF) ensure_asciiz, c3$K|]}|d yw)rN)rr^s r_rzcard..sJ1I%|1Isrirrdz not executedrlz

z

r z

z)

z

z

z>
Prepared input preview for u���
Prepared geometry input · open source gallery

Route: rz; z

z$
Current code
z
Actual checks
rmrz
Parameters
z
CPU forward
rnz'
Output shape(s)
z
z
)rrjsondumpsrsrr)rgrouppreviewgalleryrencoderrrrcurrentrvrrorirdrl note_htmls r_cardrs[[FJJx$<5Q RF YYJHb1IJ J SeF [[Hn= >FJP**U]J^"4;;vh/?#@"A FdfF>B"4;;t#4"5T:I  ) )-U);(< =2 26++e2D1E EM NRkkZ_N`Ma bB CIBJ J Y + ,3) 4V W[VaVabgVhUi j0  15 H0E/F  GL  MQKKX_L`Ka  bd  eidodopwdxcy  z  @D@K@KLR@ST  TY  Z^ZeZeflZmYn  n  $ %)KK$8#9 :Y Z`YcYcdlnoYpXq rN OUNV Vs txtt@F@J@JKTV[@\t]s^ ^E FLEM M h k    rctjddtjtdz j }tjtdz j }t |}tt||}t|d}tdd|dz|d zd d d d dd|dd tdd|dz|d zddd ddd|dd tdd|dz|dzd d!d d"d#d$|d%d& td'd(|d)z|dzd*d+d,d-d.d/|d01 td2d3|d4z|d5zd6d7d8d9d:d;|d<d= td>d3|d?z|d5zd@dAd8dBdCdD|dE1 tdFdG|dHz|dIzdJdKdLdMdNdO|dP1 tdQdR|dSz|dTzdUdVdWdXdYdZ|d[d\|d]d^d_|d`d^da|dbd^dc|ddd^de  tdfdg|dhz|dizdjdkd8dldmdn|do1 tdpdqdr|dszdtdudvdwdxdy|dz1 g }d{t|d|td}d~gdtgddtgdddj|d }tdz }|j|d|S)NT)parentsexist_okzresults/models/report.jsonz#results/models/weak_gnn/report.jsonz ../../../rr[z:data_processed/images/visualizations/cad_multiview_128.pngz data_processed/images/index.htmlu6STEP → six 17-channel orthographic images + 17 masksz2D CNN encoderz*50 XLS + 6 CAD-summary values and 56 masksu/single linear projection → FiLM at bottleneckzpixel field [view, 1, H, W]zThermal/process values generate scale and bias for the deepest image features; skip features and the modulated bottleneck enter the decoder.ryz:Execution check only: no validated pixel risk mask exists.) rrrrrrrrrrrvrr}z image / fieldz9results/representations/feature_channel_contact_sheet.pngz)aligned 17-channel image/field + 17 masksu1×1 lift + Fourier blocksu8single linear projection → FiLM before spectral blockszpixel/field gridz|The current input is the CAD image tensor. A physical thermal operator still needs paired input and target fields over time.zEArchitecture execution only; this is not a learned temperature field.rrz=data_processed/voxels/visualizations/cad_surface_dense_32.pngz data_processed/voxels/index.htmlu9STEP → 12-channel dense surface-voxel tensor + 12 masksz3D convolution encoderu2single linear projection → FiLM at 3D bottleneckzvoxel field [run, 1, X, Y, Z]zWThe process branch modulates the deepest 3D features before the skip-connected decoder.rzHThe voxel input represents sampled surfaces, not filled material volume.rrz>data_processed/voxels/visualizations/cad_surface_sparse_64.pngu.STEP → occupied voxel coordinates + featureszMinkowski sparse encoderzXLS branch proposedz7append/broadcast process embedding at sparse bottleneckzoccupied-voxel fieldz|Geometry-only sparse implementation. Process fusion is not yet wired because MinkowskiEngine is unavailable on this machine.r) rrrrrrrrrrrvrrz;data_processed/points/visualizations/cad_surface_points.pngz data_processed/points/index.htmlu0STEP → 8,192 surface points + normals/featuresz"shared point MLP + global max poolzXLS MLP proposedz.append global process embedding to every pointzpoint risk/deformation fieldzeThe present PointNet smoke is geometry-only. The diagram shows the intended process-fusion extension.rz_CAD points are valid pre-process input; scan points belong to the post-process inspection task.rz3results/representations/point_sample_comparison.pngu$points → dynamic feature-space kNNz!two EdgeConv stages + global poolz7append process embedding to local/global point featuresz point fieldz~The present DGCNN smoke is geometry-only and capped at 512 points for CPU memory. Process fusion remains a proposed extension.rrgraphzresults/graphs/comparison.pngz data_processed/graphs/index.htmlu7STEP → nodes, mesh edges, curvature/boundary featuresznode/edge message passingu7112-value process/CAD-summary vector → global encoderz/global embedding appended to every edge messagezCAD-node fieldzsThis is the most direct variable-node CAD representation. The same weights executed on all seven graph resolutions.rrztrained diagnosticz'results/models/weak_gnn/predictions.pngz"results/models/weak_gnn/index.htmlz11,956-node STEP graphzmessage passingu250 XLS + observed/imputed masks → global encoderz'process embedding in every edge messagez9node risk + signed/magnitude deformation; run defect typeuFour output heads are active. Magnitude ≥ |signed deformation| is enforced by construction; risk/deformation graph smoothness are soft losses.rz'Weak two-run training diagnostic: loss loss_initialrrrz, IoU rz, Dice ru>. Type learning is unavailable because both labels are 찍힘.rzB-Repz/data_processed/brep/visualizations/faces_uv.pngzdata_processed/brep/index.htmlu5STEP faces → trimmed 8×8 UV grids + face adjacencyz%per-face 2D CNN + face graph messagesz-append run embedding before face-message headzface risk/fieldzdThe current UV model is geometry-only. Process fusion and a face-to-surface decoder remain proposed.rrtabularr1z!data_processed/tabular/index.htmlz.STEP summary: bbox, area, node/triangle countszfeature scaling + mask appendz 50 XLS channels + validity maskszconcatenate all tabular inputszone run-level valueztThis baseline directly concatenates process and CAD summary features before the MLP; it cannot localize a risk area.ru Model input and fusion dashboard

All predictions on CAD · Weak-GNN result · Local response diagnostics · All processed inputs · Execution JSON

Model inputs, process fusion, and outputs

Each card separates the geometry representation from the XLS process branch and shows where they meet. Execution results prove tensor compatibility and gradients. A separate weak-training diagnostic maps every feasible model back onto common CAD nodes.

Prepared representations6 families
Forward/backward checks19 passed
Independent physical runs2
Validated prediction accuracyUnavailable

Every model prediction mapped onto CAD

These fields use the 18h01 weak target for a training-set representation comparison. Image, voxel and point outputs use stored source triangles and barycentric coordinates; graph outputs are already node-aligned; UV-Net uses a distance-marked face approximation; the tabular scalar is uniform. Gray regions exceed the 10 mm mapping-support threshold.

Weak-trained prediction fields from every model displayed on common CAD nodes

All-model result comparison

Parameter count and CPU latency were measured in the same environment, but each model processes a different representation size and output domain, so latency is an operational measurement rather than a fair speed ranking. The two-step losses use a synthetic zero target and differ by output domain; only the within-model decrease checks gradient/optimizer behavior. The final column is the only loss computed from the supplied weak CAD/PLY targets.

Model parameter, CPU latency and synthetic loss-reduction comparison
u
ModelOutput domainProcess fusionParametersCPU forwardSynthetic loss, step 1→2Weak-target training result

Task comparison: the risk GNN is currently the closest match to STEP + XLS → risk area + deformation because it produces CAD-node fields and consumes the process vector. Its IoU/Dice are training-set diagnostics from two unvalidated NG runs. CNN/FNO/3D U-Net already accept process conditioning but have no aligned spatial labels; PointNet, DGCNN, UV-Net and Sparse U-Net still need their XLS branch wired.

Shared XLS thermal/process branch

The plot shows the 34 heater values supplied to process-aware models. The complete vector has 50 XLS channels; values, observed masks and heater-imputation masks are retained separately.

Upper and lower 17-zone heater matrices for both runs

Requested MLP + encoder-feature concatenation

This is the explicit late-fusion variant to test beside the current FiLM models. The process embedding is repeated across pixels, voxels, points, graph nodes or B-Rep faces before concatenation.

Geometry
z6prepared geometry|image / voxel / point / graph / facezgeometry encoder|z_geo [B,S,C]zU
Process
)z,XLS input|50 values + observed/imputed masksuprocess MLP|150→64→32z repeat embedding|z_proc [B,S,32]z\
Fusion / output
)zconcatenate|[z_geo ; z_proc]ufusion MLP/conv|C+32→Cz spatial heads|risk + deformationzglobal pool + head|defect typea

For the current two-run diagnostic, the GNN injects its process embedding into every edge message. CNN/FNO/3D U-Net currently use a single Linear-to-FiLM branch. PointNet, DGCNN, UV-Net and Sparse U-Net currently have no process branch.

r z
Post-process PLY scans are target/inspection data. They are excluded from pre-process inference inputs. Proposed fusion blocks are explicitly marked where the current baseline is still geometry-only.
z index.htmlzutf-8)encoding)rQmkdirrloadsr read_textrxr`rrrrrr write_text)rtrsummary comparisonrelcardsdocumentrUs r_buildrsIIdTI* ZZ <<GGI JF ::tCCNNP QD G .Jz" C #7C@|<|GJKmGmNXhs_hYbdmtuBmCJF G 7/3?z;zEHIkEkAKgr^gaj|T]del]mt{ | #7C@<JMNpJpQ[s~jsgpOnw~@IxJQ[ \ $NCIIEISVWySyFPjuJSLUkT]der]s t *'3HEDEORSuOuHRvAS\LUs|FMNeFfmN O 'wDy@yDGHjDjr#sD"  u ''<( :$}WV 01DEWq {WR *+c x12,.IJ 1246NO T&(@AF dix% }12\  4435LMX]3 l$& #-0:=HK *-.Nb z %'Nr