DeepJEB++: 3D Data Augmentation Framework

• DeepJEB++ is a three-stage framework that augments a small CAD seed set into a vast simulation-ready 3D bracket dataset using 2D latent space techniques.
• It integrates foundation models like Stable Diffusion for image generation and TRELLIS for 3D lifting, coupled with automated finite-element simulation labeling.
• The approach achieves a 40× expansion from 380 seed brackets to over 15,000 deployable designs, showing significant improvements in IoU and Chamfer metrics.

Searching arXiv for the specified paper and closely related works to ground the article in current literature. First, I’ll look up the target paper by arXiv ID, then identify key related components such as TRELLIS and SimJEB if available on arXiv. DeepJEB++ is a three-stage, foundation-model-driven data-augmentation framework for engineering design that expands a small seed set of jet engine brackets into a large, simulation-labeled 3D dataset under constrained resources. The method addresses two coupled bottlenecks in data-driven engineering: existing 3D data augmentation techniques have limitations in preserving subtle and diverse geometric variations, and it remains difficult to automate the subsequent simulation-labeling process, where boundary conditions vary depending on the generated geometry. Its core design choice is to augment in the data-rich 2D latent space, then transfer to 3D, and finally automate boundary-condition recognition and finite-element labeling. In the reported instantiation, the pipeline expands a 380-bracket seed set into 15 360 simulation-labeled 3D jet-engine brackets under a single-GPU-per-stage budget (Yoo et al., 11 Jun 2026).

1. Problem setting and system-level design

DeepJEB++ is positioned within the problem of physics-labeled 3D engineering data generation. The motivating constraint is the lack of large-scale 3D datasets that pair geometry with physics-based performance labels. The framework therefore couples generative modeling with automated simulation in a single workflow rather than treating geometric augmentation and physics labeling as separate tasks (Yoo et al., 11 Jun 2026).

The pipeline is organized into three stages. Stage 1 performs 2D latent-space augmentation from multi-view renders of seed CAD meshes. Stage 2 lifts validated images to 3D meshes through a domain-adapted generative foundation model. Stage 3 recognizes the load and bolt interfaces on each mesh and assigns finite-element labels—mass, stress, and displacement—without manual intervention. The reported output is 15 360 “deployable” brackets obtained from a process that begins with 22 495 candidate meshes and 17 391 solved FEAs, with an overall expansion reported as $15\,360 / 381 \simeq 40\times$ (Yoo et al., 11 Jun 2026).

This decomposition is technically significant because it shifts geometric novelty generation into a modality where pretrained generative priors are stronger, then uses a separate 3D reconstruction stage and a geometry-aware labeling stage to recover engineering utility. A plausible implication is that the framework is less dependent on large native 3D corpora than end-to-end 3D generative alternatives, although the paper limits its empirical study to the jet-bracket domain.

2. Stage 1: 2D latent-space augmentation

Stage 1 begins with multi-view rendering of the seed set. Each of the 380 seed CAD meshes is rendered at 26 fixed camera poses, specified as $8$ azimuth $\times 3$ elevations plus top and bottom, yielding 9 880 images. Camera extrinsics, intrinsics, lighting, and background parameters are stored for later conditioning in Stage 2 (Yoo et al., 11 Jun 2026).

The generative backbone is Stable Diffusion v1.5, described as latent diffusion on a frozen text-encoder plus trainable U-Net and VAE. Fine-tuning uses the standard denoising diffusion objective

$$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$

with $c = \text{“A mechanical bracket design.”}$. The training configuration reported in Table 1 uses 7 800 multi-view images, corresponding to 300 brackets $\times$ 26 views, at $512^2$ resolution; full U-Net finetuning with the text encoder frozen; AdamW with $\mathrm{lr}=1\mathrm{e}{-5}$, batch size $16$, and 18 500 steps on one A100 over 1–2 days. The resulting model is reported to generate novel images that respect bracket topology—four bolt holes plus a single clevis—rather than generic “gear-like” shapes (Yoo et al., 11 Jun 2026).

Novel image synthesis is then expanded through latent-space interpolation. Two bracket images $I_1, I_2$ are encoded to latents $8$0 via the VAE encoder, followed by

$8$1

Each interpolated latent is decoded with the fine-tuned diffusion U-Net using noise strength $8$2, a UniPC sampler, 20 steps, and guidance scale $8$3. Representative-pair selection proceeds by grayscale-normalizing all 380 images, PCA-reducing each to a low-dimensional vector, scoring pairwise Euclidean distance, choosing 50 evenly spaced anchors, and forming 1 184 pairs that, over 19 $8$4-steps and 8 rendered views, produce approximately 147 000 candidates (Yoo et al., 11 Jun 2026).

Manufacturability screening is performed by a VLM-based filter built on LLaVA. The prompt is explicitly engineered to avoid the “no/negation” bias: “Describe the part’s surface and structural quality. Do not use the word ‘no’.” A negative-word vocabulary $8$5 with $8$6 defect terms is produced by an LLM, and each $8$7 is embedded into the same space as the VLM output text. For each synthesized image $8$8, LLaVA produces a description $8$9, this is embedded to $\times 3$0, and the defect score is

$\times 3$1

The top $\times 3$2 highest-$\times 3$3 images are discarded as the most defect-like. On a 636-image benchmark of 400 good and 236 corrupt images, the filter achieves 76.1% accuracy, with false-negative rate 14.0% and false-positive rate 40.7%, after adding the “no” ban. The output of this stage is 22 495 bracket images, specifically one “diagonal” view per interpolation step (Yoo et al., 11 Jun 2026).

The paper also reports an ablation labeled “VLM NWN prompt” in which accuracy improves from 56.45% to 72.17%. Taken together, these results indicate that prompt design is not incidental but part of the filtering mechanism itself.

3. Stage 2: lifting images to 3D meshes

Stage 2 converts validated images into 3D geometry using TRELLIS, a foundation model based on Structured 3D Latents (SLAT) and two sparse flow transformers. The first transformer, the SS-Generator, predicts a sparse occupancy grid $\times 3$4 from image features. The second, the SLAT-Generator, fills each occupied voxel with a $\times 3$5-dimensional feature vector encoding shape and appearance. Decoding maps SLAT to a mesh via marching cubes or 3D-Gaussian splatting, producing a watertight triangle mesh (Yoo et al., 11 Jun 2026).

The model is domain-adapted to the bracket domain rather than used as a purely off-the-shelf reconstructor. Data preparation, reported as approximately 10 hours, renders 150 views for each of 380 meshes, extracts DINO features, voxelizes each mesh to occupancy $\times 3$6, and combines DINO features with $\times 3$7 into SLAT codes. Training then separates occupancy and latent supervision: the SS-Generator is conditioned on 25 random views per shape and optimized with supervised occupancy loss using binary cross-entropy, whereas the SLAT-Generator uses single-view conditioning and a supervised latent-reconstruction loss, specifically $\times 3$8 on SLAT features. At inference time, the best single diagonal view per shape is selected, followed by mesh repair, non-manifold fix, alignment, and normalization (Yoo et al., 11 Jun 2026).

Reported quality gains on an 80-shape held-out SimJEB test are substantial. Single-view IoU improves from $\times 3$9 to $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$0, an increase of $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$1, and Chamfer decreases from $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$2 to $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$3, a reduction of $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$4. The ablation summary restates this as “SD fine-tuning: IoU +107%, Chamfer –39% vs. no adaptation.” A separate multi-view comparison reports 6-view CD = 0.0032 versus single diagonal CD = 0.0031, which the paper characterizes as saturation at 6 views (Yoo et al., 11 Jun 2026).

The physical plausibility priors are explicit. The sparse-occupancy prior imposed by the SS-Generator enforces closed, manifold geometry; SLAT features are regularized via $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$5 to remain near a pretrained manifold learned from 500 K+ shapes; and post-processing mesh repair enforces watertightness and removes self-intersections. This suggests that plausibility is distributed across the representation, the learned prior, and deterministic repair, rather than delegated to a single generative component.

4. Stage 3: automated boundary-condition recognition and finite-element labeling

Stage 3 assigns engineering labels by recognizing attachment and loading interfaces directly from the reconstructed meshes. The target topology is a bracket with four bolt holes and a clevis load interface comprising two holes. Bolt-hole detection begins by finding connected components of vertices whose normals are perpendicular to the bolt axis. For each component, the method fits the angular sweep of normals and accepts the component if the sweep is at least $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$6; it then merges coaxial components, keeps the smallest-diameter ring, and discards counterbores. Clevis detection sections the mesh with planes perpendicular to the pin axis so that each prong yields a closed loop, fits a circle using Kåsa’s method through the least-squares form $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$7, clusters loop centers in 3D, and selects the pair with equal radius and separation at least 20 mm. Rigid registration to a fixed SimJEB bolt template is then performed using Umeyama’s method to establish a consistent global frame (Yoo et al., 11 Jun 2026).

The system also includes explicit geometric gating. For SimJEB meshes, acceptance requires $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$8, $$L(\theta) = E_{z_0\sim q(z),\, t,\, \epsilon\sim\mathcal{N}(0,I)} \left[ \|\epsilon - \epsilon_\theta(z_t, t; c)\|^2 \right]$$9 mm, and $c = \text{“A mechanical bracket design.”}$0 mm. For generated meshes, these thresholds are relaxed to $c = \text{“A mechanical bracket design.”}$1, $c = \text{“A mechanical bracket design.”}$2 mm, and $c = \text{“A mechanical bracket design.”}$3 mm (Yoo et al., 11 Jun 2026).

Finite-element preprocessing decimates each surface to approximately 25 000 nodes and remeshes to tetrahedra with radius–edge $c = \text{“A mechanical bracket design.”}$4 and minimum dihedral $c = \text{“A mechanical bracket design.”}$5. The material model is Ti-6Al-4V with $c = \text{“A mechanical bracket design.”}$6 GPa, $c = \text{“A mechanical bracket design.”}$7, and yield strength 903 MPa. Boundary conditions clamp all bolt nodes with $c = \text{“A mechanical bracket design.”}$8 and apply distributed nodal forces at the clevis for four cases: vertical load $c = \text{“A mechanical bracket design.”}$9 kN, horizontal load $\times$0 kN, diagonal load $\times$1 kN at $\times$2, and torsional moment $\times$3 kN·mm about the pin axis. The solver is linear elasticity,

$\times$4

For each case, the labels are the maximum displacement magnitude $\times$5 and $\times$6, where

$\times$7

A mass label is additionally computed as $\times$8 (Yoo et al., 11 Jun 2026).

This stage is central to the dataset’s engineering relevance. The framework does not merely synthesize geometry; it operationalizes geometry into simulation-ready instances with automatically assigned boundary conditions and labels.

5. Evaluation protocol and reported results

DeepJEB++ evaluates quality along three intrinsic axes: manufacturability, label fidelity, and distributional consistency. Manufacturability is defined as the fraction of images passing the VLM filter plus the fraction of meshes with valid four-bolt and clevis detection. Label fidelity compares FEA on generated meshes against SimJEB ground truth on the same CAD geometry. Distributional consistency considers mass distributions, stress and displacement histograms normalized by dataset medians, and appearance coverage computed by PCA on Stable-Diffusion VAE embeddings of diagonal-view renders (Yoo et al., 11 Jun 2026).

For label fidelity, the paper uses a node-level six-DOF least-squares fit,

$\times$9

together with

$512^2$0

The reported median $512^2$1 values are 0.94 for vertical loading, 0.94 for horizontal loading, 0.88 for diagonal loading, and 0.91 for torsional loading. For mass, the paper reports $512^2$2 versus SimJEB CAD and a median error of 0.06% (Yoo et al., 11 Jun 2026).

For distributional consistency, mass distributions are described as mesh-robust, while stress and displacement are assessed through histograms normalized by dataset medians. In appearance space, the reported PCA analysis indicates that DeepJEB++ envelops SimJEB and DeepJEB support. This suggests that the generated set is not merely larger than the seed data but also broader in latent appearance coverage, while still remaining connected to the original distribution (Yoo et al., 11 Jun 2026).

The quantitative yield is reported explicitly: 380 seed brackets lead to 22 495 candidate meshes, then 17 391 solved FEAs, and finally 15 360 deployable instances with valid single-body mass and boundary-condition detection. Compute requirements are also itemized: one A100 GPU per stage, with Stage 1 taking 1–2 days, Stage 2 approximately 2 days, and Stage 3 using a CPU cluster for FEA (Yoo et al., 11 Jun 2026).

6. Interpretation, limitations, and scope

The paper identifies several limitations. The VLM quality filter, reported at 72%–76% accuracy, remains coarse, and future work is stated to include fine-tuning VLMs on engineered defect labels. The 2D-to-3D lifting stage can introduce mesh artifacts such as thin walls and noise, motivating improved mesh repair or learned denoisers. Boundary-condition recognition relies on geometric heuristics, and the authors note that a learning-based segmentation network might generalize better to novel topologies. Portability beyond the jet-bracket domain is also unproven, because Stage 3 heuristics would need re-validation for other mechanical parts (Yoo et al., 11 Jun 2026).

The paper further notes that absolute stress and displacement are mesh-resolution dependent, so current comparisons focus on distributional shape rather than absolute values. This addresses a common misconception that the resulting simulation labels should be interpreted as mesh-independent ground truth. Within the reported framework, label fidelity is strong but still conditioned on reconstruction quality, meshing choices, and the specific FEA discretization pipeline (Yoo et al., 11 Jun 2026).

A second plausible misconception is that large-scale 3D engineering augmentation must be performed directly in 3D. DeepJEB++ provides an alternative formulation: augment first in a pretrained 2D latent space, then lift to 3D through a domain-adapted foundation model, and finally automate boundary-condition recognition and simulation. The paper does not claim universal superiority for this strategy, but its reported results suggest that it is an effective method under constrained academic compute for the specific bracket domain examined (Yoo et al., 11 Jun 2026).

The downstream impact on surrogate modeling or topology optimization is explicitly identified as not yet quantified. Accordingly, DeepJEB++ should be understood as an infrastructure contribution: it establishes a method for scaling from hundreds to tens of thousands of physics-labeled engineering designs, and the dataset is to be made publicly available to support reproducible engineering-AI research.