AIMold: An Autonomous AI-based Pipeline for Complex Mold Design¶
Conference: ECCV2026
Paper: Official paper page
Full Text: Official PDF
Code: https://github.com/tb2-sy/AIMold
Area: 3D Vision
Keywords: Mold assembly generation, Manufacturing-aware CAD, Demolding orientation, Structured latents, Geometry post-processing
TL;DR¶
AIMold combines orientation selection, coarse voxel assembly generation, latent geometry refinement, and CAD post-processing, raising auxiliary-component COV from the VecSet baseline's 21.24% to 57.29% on the MoldCAD test split while still requiring downstream manufacturing verification.
Background & Motivation¶
An injection mold is not simply a negative copy of the desired object: the finished part must also be removable when the mold opens. Side holes, undercuts, and re-entrant features can lock a part inside a conventional two-piece mold, requiring auxiliary components that retract in other directions. Traditional automation often addresses orientation selection, undercut detection, or parting surfaces separately, although these decisions constrain one another. Changing the opening direction can change both the required auxiliary components and the boundary between the upper and lower molds. The target here is therefore a geometrically coordinated assembly, not an isolated object that merely looks like CAD.
Learning this relationship is difficult because paired engineering data is scarce and a part can admit several valid mold designs. A general-purpose shape prior does not automatically encode manufacturing practice, while implicit representations may lose input–output alignment when trained on a small specialized dataset. The authors commission professional designs for MoldCAD and retain a shared voxel coordinate system before introducing fine geometry. Core idea: learn orientation and assembly layout from expert mold designs, then combine detailed generation with contact-aware geometric reconstruction to produce structured drafts for CAD/CAM verification and revision.
Method¶
Overall Architecture¶
The input is a single-body CAD solid or watertight mesh; the output includes auxiliary components, upper and lower molds, and a reconstructed parting surface. Plane-candidate orientation establishes the working frame, and sequential voxel generation predicts auxiliary components before conditioning upper/lower mold generation on them. Structured-latent refinement converts the coarse structures into detailed meshes, followed by contact-constrained CAD reconstruction and STEP/STP export. “Autonomous” describes the chained design procedure, not a language-agent architecture or a formal guarantee of demoldability.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Single-body part"] --> B["Plane-candidate orientation"]
B --> C["Sequential voxel generation<br/>Auxiliaries → upper/lower molds"]
C --> D["Structured-latent refinement"]
D --> E["Contact-constrained<br/>CAD reconstruction"]
E --> F["Mold assembly<br/>Downstream verification"]
Key Designs¶
1. Plane-candidate orientation: classify geometric candidates instead of regressing a continuous pose
The opening direction determines which features become undercuts, but directly regressing a direction vector from a high-dimensional shape representation performs poorly in the reported experiment. AIMold first detects planar regions and constructs a finite set of candidate orientations from them. Candidates whose rotation axes align with the annotated orientation are positive examples; non-aligned candidates are negatives. The classifier uses the voxel representation and encoder architecture employed by the generation system, replacing the decoder with a lightweight binary head. It selects an orientation for the subsequent generation stages rather than searching freely through continuous pose space.
This design lets geometry narrow the decision before learning begins. Both positive and negative classes are augmented to four times the scale of the original dataset to address imbalance. However, matching an annotated orientation does not establish that every generated component can subsequently retract without collision. The main paper does not separately evaluate the full behavior when plane detection misses useful candidates or a part has few prominent planar regions.
2. Sequential voxel generation: predict side components before assigning upper/lower mold occupancy
The number of auxiliary components varies by part, making a fixed collection of output slots awkward. Instead, the first generator takes the part alone and predicts the binary occupancy of the union of its auxiliary components in a common voxel frame. The authors rely on the observation that auxiliary components rarely intersect, so their union can subsequently be separated into individual parts. The exact separation algorithm is not specified in the main text; it should not be embellished into an additional learned instance-segmentation module. The voxel encoder and decoder use a Transformer architecture resembling a 3D VAE, with the shared coordinates preserving spatial correspondence to the input.
Upper/lower mold generation then takes the joint voxelized representation of the original part and the generated auxiliaries. Its output has three labels: upper-mold occupancy, lower-mold occupancy, and empty space. This shared partition prevents a voxel from being assigned to both mold halves, rather than predicting two unrelated binary volumes. The ordering matters because auxiliary components occupy regions that the main molds must accommodate. Nevertheless, consistency on a coarse voxel grid is not equivalent to precise assembly tolerances or collision-free motion.
3. Structured-latent refinement: retain coarse spatial support while recovering local geometry
Coarse occupancy is useful for layout but loses curves, sharp edges, and thin CAD structures, making direct geometric post-processing difficult. The second stage retains active voxel locations and generates feature latents associated with those locations before decoding a mesh. For training representations, multiview DINOv2 features are aggregated and compressed by a 3D VAE encoder. An object is thus represented by active coordinates and their local features rather than an unanchored global vector. The geometry flow model uses a sparse-structure Transformer adapted from the referenced structured-3D-latent framework, removing its original text/image conditions and using the voxel structure to guide generation.
Rectified Flow learns a velocity field between noise and data latents through a Flow Matching objective. The mesh decoder predicts FlexiCubes parameters and signed distances at voxel vertices to reconstruct the surface. The stage primarily refines geometry on the supplied spatial support; it is not a second independent decision about where auxiliary components should exist. This separates learning the assembly layout from expressing local detail. Equations (1) and (2) are corrupted in the local text extraction, so this note retains the supported mechanism without reconstructing the loss expression or the expanded decoder parameterization.
4. Contact-constrained CAD reconstruction: correct the draft against the original part
A plausible generated mesh is insufficient when mating surfaces must closely follow the target CAD geometry. Post-processing begins with screened Poisson reconstruction, then analyzes the part near the predicted component to estimate a dominant axis and separate embedded from external regions. For the embedded region, it extracts an intersection polyline or curve from the target cavity and sweeps a mating volume along that axis. For the external region, RANSAC fits standard geometric primitives, after which Boolean composition joins the regions. The original part provides the geometric reference, reducing the need for the network to reproduce every contact point precisely.
Upper and lower molds are also rebuilt rather than simply exported unchanged. A plane or quadratic parting surface is fitted to their predicted interface and used to split an inflated assembly bounding volume. The original part and optimized auxiliary components are then subtracted from the respective blocks. ShapeUp optimization, quad remeshing, and NURBS surface conversion precede STEP/STP export. These are non-learned operations: AIMold combines learned drafts with geometric reconstruction rather than directly generating an exact B-Rep end to end. Despite the reconstruction step intended to improve watertightness, the authors still report missing-face and non-watertight failures; CAD importability is not universal manufacturing success.
A Worked Example¶
Consider a single-body part with a side hole as an illustrative input. Plane detection supplies candidate orientations, and the classifier selects a frame compatible with the primary opening direction. In that frame, the side hole may remain inaccessible to the upper and lower molds alone, so the auxiliary generator predicts occupancy in the relevant lateral region. That occupancy becomes part of the condition for the mold-half generator, informing it which space the auxiliaries require. Refinement replaces coarse voxel steps with more detailed meshes, while CAD reconstruction uses the actual hole boundary to form the mating volume and subtract it from the mold blocks. This is an explanation of module interaction, not an additional reported success case; retraction feasibility still requires downstream checks.
Loss & Training¶
MoldCAD selects single-body parts from ABC and BRepNet, filtering for basic properties such as watertightness and suitable wall thickness while retaining undercuts and side holes. Professional engineers create paired assemblies with annotations for mold halves, parting surfaces, auxiliary instances, and demolding orientation. The abstract reports 4,934 CAD models, over 3,850 mold assemblies, and more than 23,000 individual models in total; these distinct counting units should not be collapsed into 4,934 complete paired assemblies. Section 5.1 uses a 9:1 training/test split, with 10% held out for evaluation. The first stage uses AdamW with a learning rate of \(10^{-4}\); structured-latent encoding renders 12 views per CAD model, and experiments run on NVIDIA A100 GPUs. The second-stage training follows the referenced structured-latent setup. The main text does not fully specify the first-stage occupancy loss, orientation tolerance, inference step count, or all architecture hyperparameters, so they are not supplied here.
Key Experimental Results¶
Main Results¶
The following selected entries come from Tables 1 and 2 on the MoldCAD test split described above. COV measures coverage of reference shapes by the generated set, MMD measures reference-to-generated matching distance, and JSD measures distributional discrepancy. Higher COV and lower MMD/JSD are preferred; none is an assembly-success rate. The paper points to Appendix A for calculation details, but that appendix is absent from the available 21-page local text, leaving the exact distance, sampling, and normalization protocol unverified.
| Generation task | Method | COV (%) ↑ | MMD ↓ | JSD ↓ |
|---|---|---|---|---|
| Auxiliary components, Table 1 | VecSet-based | 21.24 | 0.02900 | 0.105 |
| Auxiliary components, Table 1 | AIMold | 57.29 | 0.0147 | 0.050 |
| Upper/lower molds, Table 2 | VecSet-based | 9.21 | 0.0252 | 0.277 |
| Upper/lower molds, Table 2 | AIMold | 49.01 | 0.0091 | 0.061 |
The COV gains are 36.05 percentage points for auxiliaries and 39.80 points for mold halves relative to the reported VecSet-based implementation. These results support explicit spatial correspondence at the current specialized dataset scale, not a universal ranking against every general-purpose generator under matched pretraining and tuning budgets. Separately, Table 3 reports orientation accuracy of 91.17%, versus 12.66% for direct regression; this is not the success rate of the complete mold-generation pipeline.
Ablation Study¶
Table 4 evaluates auxiliary generation under the same MoldCAD test setting; Section 5.3 identifies the reduced-view setting as four rendered views.
| Configuration | COV (%) ↑ | MMD ↓ | JSD ↓ |
|---|---|---|---|
| Encoder without pretraining | 52.82 | 0.0173 | 0.0682 |
| Reduced to 4 rendered views | 55.44 | 0.0146 | 0.050 |
| First stage only | 55.68 | 0.0140 | 0.054 |
| Full AIMold | 57.29 | 0.0147 | 0.050 |
Key Findings¶
- Removing pretraining lowers COV by 4.47 percentage points and worsens both MMD and JSD, supporting the usefulness of general 3D priors.
- Refinement is not uniformly better: compared with the first stage alone, the full model gains 1.61 COV points and lowers JSD, but MMD rises from 0.0140 to 0.0147.
- Increasing from 4 to 12 views mainly improves COV in this table; the four-view variant has slightly lower MMD and identical JSD, so an across-the-board metric improvement would be inaccurate.
Highlights & Insights¶
- Professional paired assemblies make manufacturing decisions learnable rather than merely increasing the number of isolated CAD shapes. They expose relationships among auxiliaries, primary molds, and the target part.
- Explicit coordinates preserve alignment in a small-data setting while generative modeling supplies local detail. This is a task-specific representational choice rather than a claim that larger latent models are always worse.
- Contact reconstruction uses the precise input as a geometric reference. The division between learned drafts and deterministic correction is useful when output plausibility and exact mating geometry have different requirements.
Limitations & Future Work¶
- Author-reported: thin or small structures are resolution-limited, missing faces can produce non-watertight outputs, the dataset remains small, and multiview encoding trades view count against captured occluded structure.
- Explicit task boundary: generation does not enforce demolding constraints; it relies on implicit manufacturing priors and requires downstream CAD/CAM verification and revision.
- Reader assessment: distribution metrics and orientation accuracy do not establish collision-free assembly motion, tolerance compliance, or injection-molding quality. Those measurements would strengthen the industrial claim more directly than COV alone.
- Evidence gap: Appendix A is unavailable locally and core equations are extraction-corrupted. This note does not claim a complete reproduction protocol or equate CAD export with physical manufacturing validation.
Related Work & Insights¶
- Versus traditional orientation and undercut analysis: earlier methods use visibility, geometric rules, or search for local manufacturing conditions. AIMold learns assembly patterns from expert designs without providing strict feasibility guarantees.
- Versus VecSet-style generation: shared voxel coordinates preserve input–output correspondence here. The reported advantage is specific to the MoldCAD comparison, not a universal ordering of representation families.
- Versus structured 3D latent generation: AIMold reuses sparse structural latents and mesh decoding while conditioning on coarse mold geometry. Its contribution centers on the manufacturing task, paired data, and hybrid pipeline rather than a new general flow-matching theory.
Rating¶
- Novelty: 4/5. The complex-mold task and professional paired data are distinctive, while many generation and geometry components are established.
- Experimental Thoroughness: 3/5. Two generation tasks, orientation comparisons, and ablations are useful, but end-to-end manufacturing feasibility remains insufficiently quantified.
- Writing Quality: 3/5. The workflow and task boundary are clear; metric implementation and several important training details depend on information outside the main text.
- Value: 4/5. A meaningful dataset and system baseline for manufacturing-aware CAD, best viewed as design assistance rather than review-free production output.