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

Model	Output domain	Process fusion	Parameters	CPU forward	Synthetic loss, step 1→2	Weak-target training result
2D CNN/U-Net	pixel	FiLM	37,505	136.397 ms	0.038875 → 0.030681 (+21.1%)	—
FNO2d	pixel/field	FiLM	229,273	57.361 ms	0.028175 → 0.027498 (+2.4%)	—
Dense 3D U-Net	voxel	FiLM	27,249	3.557 ms	0.070244 → 0.065694 (+6.5%)	—
Sparse 3D U-Net	sparse voxel	none	46,432	5.364 ms	—	—
PointNet	point	none	26,561	2.317 ms	0.0072413 → 0.0010095 (+86.1%)	—
DGCNN	point	none	9,025	5.239 ms	0.0088493 → 0.0078437 (+11.4%)	—
Mesh GNN	CAD node	edge-message	37,777	113.922 ms	0.01316 → 0.0091457 (+30.5%)	—
Risk GNN	CAD node + run	edge-message	18,694	29.426 ms	0.5644 → 0.53942 (+4.4%)	loss 0.5600; IoU 0.306; Dice 0.469
UV-Net	B-Rep face	none	1,601	59.354 ms	0.24004 → 0.23371 (+2.6%)	—
Tabular MLP	run	direct concat	11,457	0.030 ms	0.049162 → 0.042526 (+13.5%)	—

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

prepared geometryimage / voxel / point / graph / face

→

geometry encoderz_geo [B,S,C]

Process

XLS input50 values + observed/imputed masks

→

process MLP150→64→32

→

repeat embeddingz_proc [B,S,32]

Fusion / output

concatenate[z_geo ; z_proc]

→

fusion MLP/convC+32→C

→

spatial headsrisk + deformation

→

global pool + headdefect type

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.

image

2D CNN / U-Net

forward/backward passed

Route: STEP → six 17-channel orthographic images + 17 masks → 2D CNN encoder; 50 XLS + 6 CAD-summary values and 56 masks → single linear projection → FiLM at bottleneck → pixel field [view, 1, H, W]

Geometry

image input34×128×128

→

Conv2d ×234→16→16

→

MaxPool2d128→64

→

Conv2d ×216→32→32

Process

process input112

→

Linear112→64

→

splitscale 32 + bias 32

Fusion / output

FiLM bottleneck32×64×64

→

ConvTranspose2d32→16

→

concat skip16+16

→

Conv2d ×232→16→16

→

1×1 head16→1 pixel

Current code: Thermal/process values generate scale and bias for the deepest image features; skip features and the modulated bottleneck enter the decoder.
Actual checks: 1
Parameters: 37,505
CPU forward: 136.4 ms
Output shape(s): [[12, 1, 128, 128]]

Execution check only: no validated pixel risk mask exists.

image / field

FNO2d

forward/backward passed

Route: aligned 17-channel image/field + 17 masks → 1×1 lift + Fourier blocks; 50 XLS + 6 CAD-summary values and 56 masks → single linear projection → FiLM before spectral blocks → pixel/field grid

Geometry

field input34×128×128

→

1×1 lift34→24

Process

process input112

→

Linear112→48

→

splitscale 24 + bias 24

Fusion / output

FiLM lifted field24×128×128

→

spectral + skip ×324 channels

→

1×1 head24→1 pixel

Current code: The current input is the CAD image tensor. A physical thermal operator still needs paired input and target fields over time.
Actual checks: 1
Parameters: 229,273
CPU forward: 57.4 ms
Output shape(s): [[12, 1, 128, 128]]

Architecture execution only; this is not a learned temperature field.

voxel

Dense 3D U-Net

forward/backward passed

Route: STEP → 12-channel dense surface-voxel tensor + 12 masks → 3D convolution encoder; 50 XLS + 6 CAD-summary values and 56 masks → single linear projection → FiLM at 3D bottleneck → voxel field [run, 1, X, Y, Z]

Geometry

voxel input24×32×4×16

→

Conv3d ×224→8→8

→

MaxPool3d16×2×8

→

Conv3d ×28→16→16

Process

process input112

→

Linear112→32

→

splitscale 16 + bias 16

Fusion / output

FiLM bottleneck16×16×2×8

→

ConvTranspose3d16→8

→

concat skip8+8

→

Conv3d ×216→8→8

→

1×1 head8→1 voxel

Current code: The process branch modulates the deepest 3D features before the skip-connected decoder.
Actual checks: 1
Parameters: 27,249
CPU forward: 3.6 ms
Output shape(s): [[2, 1, 32, 4, 16]]

The voxel input represents sampled surfaces, not filled material volume.

sparse voxel

Sparse 3D U-Net

forward/backward passed

Route: STEP → occupied voxel coordinates + features → Minkowski sparse encoder; XLS branch proposed → append/broadcast process embedding at sparse bottleneck → occupied-voxel field

Geometry

sparse input403 coords, 24 feat

→

MinkowskiConv24→16

→

stride-2 down16→32

→

mid Conv32→32

Process

XLS branchnot wired

Fusion / output

transpose Conv32→16

→

concat sparse skip16+16

→

sparse head32→1 cell

Current code: Geometry-only sparse implementation. Process fusion is not yet wired because MinkowskiEngine is unavailable on this machine.
Actual checks: 1
Parameters: 46,432
CPU forward: 5.4 ms
Output shape(s): [[403, 1]]

point

PointNet segmentation

forward/backward passed

Route: STEP → 8,192 surface points + normals/features → shared point MLP + global max pool; XLS MLP proposed → append global process embedding to every point → point risk/deformation field

Geometry

point input8,192×C

→

shared Conv1dC→64→128

→

global max128

→

repeat + concat128+128

Process

XLS MLPproposed, not wired

Fusion / output

point head256→64→1

→

outputone scalar / point

Current code: The present PointNet smoke is geometry-only. The diagram shows the intended process-fusion extension.
Actual checks: 3
Parameters: 25,921, 26,561
CPU forward: 1.7–2.3 ms
Output shape(s): [[1, 8192, 1], [1, 8192, 1], [1, 8192, 1]]

CAD points are valid pre-process input; scan points belong to the post-process inspection task.

point

DGCNN segmentation

forward/backward passed

Route: points → dynamic feature-space kNN → two EdgeConv stages + global pool; XLS MLP proposed → append process embedding to local/global point features → point field

Geometry

point input512×C

→

dynamic kNNk=12

→

Edge MLP 12C→32→32

→

new kNN + Edge MLP 264→32→32

Process

XLS MLPproposed, not wired

Fusion / output

local + local + global32+32+32

→

point head96→32→1

Current code: The present DGCNN smoke is geometry-only and capped at 512 points for CPU memory. Process fusion remains a proposed extension.
Actual checks: 3
Parameters: 8,385, 9,025
CPU forward: 2.1–5.2 ms
Output shape(s): [[1, 512, 1], [1, 512, 1], [1, 512, 1]]

graph

MeshGraphNet-style GNN

forward/backward passed

Route: STEP → nodes, mesh edges, curvature/boundary features → node/edge message passing; 112-value process/CAD-summary vector → global encoder → global embedding appended to every edge message → CAD-node field

Geometry

node inputN×32

→

node encoder32→48

→

edge inputE×8

Process

global input112

→

Linear + GELU112→48

Fusion / output

edge message ×348+48+8+48→48→48

→

sum at destinationN×48

→

GRU update ×396→48

→

node head48→1

Current code: This is the most direct variable-node CAD representation. The same weights executed on all seven graph resolutions.
Actual checks: 7
Parameters: 37,777
CPU forward: 4.6–113.9 ms
Output shape(s): [[658, 1], [23912, 1], [2786, 1], [658, 1], [5812, 1], [11850, 1], [1268, 1]]

trained diagnostic

Physics-constrained risk GNN

forward/backward passed

Route: 11,956-node STEP graph → message passing; 50 XLS + observed/imputed masks → global encoder → process embedding in every edge message → node risk + signed/magnitude deformation; run defect type

Geometry

node input2×11,956×32

→

node encoder32→32

→

edge input2×72,176×8

Process

weak-train input2×150

→

Linear + GELU150→32

→

smoke alternative2×112→32

Fusion / output

edge message ×232+32+8+32→32→32

→

GRU update ×264→32

→

raw head32→(3 + classes)

→

node headsrisk, signed, magnitude

→

mean poolrun defect type

Current code: Four output heads are active. Magnitude ≥ |signed deformation| is enforced by construction; risk/deformation graph smoothness are soft losses.
Actual checks: 1
Parameters: 18,694
CPU forward: 29.4 ms
Output shape(s): [{"risk_logit": [23912], "signed_normal_deformation": [23912], "deformation_magnitude": [23912], "defect_type_logits": [2, 3]}]

Weak two-run training diagnostic: loss 1.962 → 0.560, IoU 0.306, Dice 0.469. Type learning is unavailable because both labels are 찍힘.

B-Rep

UV-Net-style face model

forward/backward passed

Route: STEP faces → trimmed 8×8 UV grids + face adjacency → per-face 2D CNN + face graph messages; XLS MLP proposed → append run embedding before face-message head → face risk/field

Geometry

face UV grids6,564×12×8×8

→

Conv2d ×212→8→8

→

mean UV pool6,564×8

→

face edges36,090

Process

XLS MLPproposed, not wired

Fusion / output

face message8+8→8

→

sum at destination6,564×8

→

face head8→1

Current code: The current UV model is geometry-only. Process fusion and a face-to-surface decoder remain proposed.
Actual checks: 1
Parameters: 1,601
CPU forward: 59.4 ms
Output shape(s): [[6564, 1]]

tabular

Tabular MLP surrogate

forward/backward passed

Route: STEP summary: bbox, area, node/triangle counts → feature scaling + mask append; 50 XLS channels + validity masks → concatenate all tabular inputs → one run-level value

Geometry

CAD summary6 values + masks

Process

XLS process50 values + masks

→

concatenate112 total

Fusion / output

Linear + GELU112→64

→

Linear + GELU64→64

→

run head64→1

Current code: This baseline directly concatenates process and CAD summary features before the MLP; it cannot localize a risk area.
Actual checks: 1
Parameters: 11,457
CPU forward: 0.0 ms
Output shape(s): [[2, 1]]