README.md

---
license: odc-by
pretty_name: "DeepJEB++"
size_categories:
  - 10K<n<100K
task_categories:
  - tabular-regression
  - graph-ml
tags:
  - engineering-design
  - finite-element-analysis
  - structural-mechanics
  - 3d
  - mesh
  - generative-design
  - foundation-model
  - jet-engine-bracket
  - surrogate-modeling
---

<div align="center">
  <img src="assets/banner_displacement.png" alt="DeepJEB++ generated brackets — displacement fields" width="100%">
</div>

<h1 align="center">DeepJEB++</h1>
<p align="center"><b>Foundation Model-Driven Large-Scale 3D Engineering Dataset via 2D Latent Space Augmentation</b></p>

<p align="center">
  <a href="https://arxiv.org/abs/2606.12994"><img src="https://img.shields.io/badge/arXiv-2606.12994-b31b1b.svg" alt="arXiv"></a>
  <a href="https://huggingface.co/datasets/KAIST-SmartDesignLab/DeepJEB-PP"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-ffb000.svg" alt="Hugging Face Dataset"></a>
  <img src="https://img.shields.io/badge/Status-Under%20Review-orange.svg" alt="Status: Under Review">
  <img src="https://img.shields.io/badge/license-ODC--By%201.0-blue.svg" alt="License: ODC-By 1.0">
</p>

<p align="center">Soyoung Yoo · Leekyo Jeong · Jinsu Ra · Dongeon Lee · Sunwoong Yang · Hyogu Jeong · Namwoo Kang &nbsp;—&nbsp; <b>KAIST SmartDesignLab</b></p>

---

## Contents

- [News](#news)
- [Overview](#overview)
- [The data, qualitatively](#the-data-qualitatively)
- [Augmentation methodology](#augmentation-methodology)
- [Dataset structure](#dataset-structure)
- [Usage](#usage)
- [Applications](#applications)
- [Citation](#citation)
- [Acknowledgements](#acknowledgements)
- [License](#license)

---

## News

- **2026-06** — DeepJEB++ released on Hugging Face: **15,360** designs with surface meshes, boundary conditions, per-load FEA surface fields, and scalar labels (incl. mass).
- **2026-06** — Preprint on arXiv ([2606.12994](https://arxiv.org/abs/2606.12994)); manuscript **under review**.

---

## Overview

> **DeepJEB++** is a large-scale dataset of **generatively-designed jet-engine brackets**, each paired with
> physics-based performance labels from an automated finite-element (FEA) pipeline. It is built by **augmenting
> the SimJEB design space inside a 2D latent space** and lifting the synthesized images to 3D with a **3D
> foundation model (TRELLIS)**, then automatically recovering boundary conditions and solving four structural
> load cases. The result couples **geometry ↔ physics** at a scale (40× SimJEB) suitable for data-driven and
> surrogate modelling in engineering design.

<div align="center">
  <img src="assets/teaser.gif" alt="A generated bracket rotating, coloured by its vertical-load displacement field" width="46%">
  <br><sub>A single design, coloured by its vertical-load displacement field (blue = clamped bolts, red = lug tip).</sub>
</div>

| | |
|---|---|
| **Designs (deployable)** | 15,360 |
| **Load cases** | vertical / horizontal / diagonal / torsional |
| **Per design** | surface mesh · boundary conditions · FEA surface fields · scalar labels (incl. mass) |
| **Material** | Ti-6Al-4V · E = 113,800 MPa · ν = 0.342 |
| **Scale** | 40× SimJEB (380) |
| **Paper** | [arXiv:2606.12994](https://arxiv.org/abs/2606.12994) |
| **License** | ODC-By 1.0 (matching upstream SimJEB / DeepJEB) |

---

## The data, qualitatively

<div align="center">
  <img src="assets/gallery.png" alt="Generated bracket variety with auto-detected interfaces" width="92%">
  <br><sub><b>Generated bracket variety + auto-detected interfaces</b> — 24 of 15,360, each with a gate-validated 4-bolt flange and lug-clevis detection (orange).</sub>
</div>

<br>

<div align="center">
  <img src="assets/fea_fields.png" alt="Four-load FEA response fields" width="92%">
  <br><sub><b>4-load FEA response fields.</b> Top: displacement (deformed ×9). Bottom: von Mises stress. Columns: vertical / horizontal / diagonal / torsional.</sub>
</div>

The hero banner shows real brackets coloured by their per-case vertical-load displacement field. The same
brackets, as raw geometry:

<div align="center">
  <img src="assets/banner_geometry.png" alt="Generated bracket meshes (geometry)" width="100%">
</div>

---

## Augmentation methodology

The core idea is **2D latent-space augmentation**: instead of perturbing 3D meshes directly, new designs are
synthesized by **interpolating between SimJEB seed brackets in the latent space of a fine-tuned diffusion
model**, then reconstructed in 3D by a foundation model and labelled by FEA.

| # | Step | What happens |
|---|------|--------------|
| 1 | **Seed pairs** | Pairs of SimJEB bracket renders chosen as interpolation endpoints. |
| 2 | **2D latent interpolation** | Fine-tuned Stable Diffusion mixes the two VAE latents (ratio 0→1) → frames IS00–IS18. |
| 3 | **Image → 3D** | A single diagonal view drives TRELLIS (SimJEB-finetuned) image-to-3D, 25-step. |
| 4 | **Automatic BC** | 4-bolt flange + lug-clevis detected and validated by a calibrated gate. |
| 5 | **FEA labels** | Four load cases solved → displacement, von Mises, mass per design. |

<div align="center">
  <img src="assets/framework.png" alt="End-to-end framework" width="92%">
  <br><sub><b>End-to-end framework</b> — generation (latent interpolation + foundation-model lifting) → automatic labelling.</sub>
</div>

<br>

<div align="center">
  <img src="assets/interp_2d.png" alt="2D latent interpolation" width="80%">
  <br><sub><b>2D latent interpolation</b> — a smooth transition between two parent brackets (IS00 → IS18).</sub>
</div>

> **Why 2D-latent augmentation?** Interpolating in a learned image latent space produces smooth, valid,
> manufacturable-looking new brackets that span the design space between real examples — far easier than
> perturbing 3D meshes directly — while a 3D foundation model guarantees consistent, watertight geometry ready
> for FEA. A key finding: increasing the diffusion sampling steps raised valid BC-detection from **16% → 96%**.

---

## Dataset structure

Distributed as per-component `.tar.gz` archives + a CSV. Every design shares one `<case>` id
(e.g. `012-015-diag_xz_mm_IS02`) across all modalities.

```
DeepJEB-PP/
├── 1_surface_meshes.tar.gz       # 15,360 × <case>.obj
├── 2_boundary_conditions.tar.gz  # 15,360 × <case>.npz
├── 3_fea_fields.tar.gz           # 15,360 × <case>.npz
├── deepjebpp_labels.csv          # scalar labels (15,360 rows)
└── metadata.json                 # material / loads / units / schema
```

**Modalities**

| Modality | File | Content |
|---|---|---|
| Geometry | `1_surface_meshes/<case>.obj` | input surface mesh, native ~50k verts |
| Boundary conditions | `2_boundary_conditions/<case>.npz` | `bolt_idx` (clamped), `lug_idx` (loaded), `bolt_holes` — indices into the 25k FEM `surface_points` |
| FEA surface fields | `3_fea_fields/<case>.npz` | `surface_points` (N,3), `surface_faces` (M,3), and per load `{ver,hor,dia,tor}_U` (N,3) + `_vm` (N,) |
| Scalar labels | `deepjebpp_labels.csv` | `mass_g`, `vol_mm3`, per-load `max|u|`, `p95 von Mises`, … |

**FEA specification**

| | |
|---|---|
| Material | Ti-6Al-4V · E = 113,800 MPa · ν = 0.342 (yield 903 MPa / 131 ksi, reference) |
| Vertical (ver) | force (0, 0, 1) · 35,600 N |
| Horizontal (hor) | force (−1, 0, 0) · 37,800 N |
| Diagonal (dia) | force (−0.669, 0, 0.743) · 42,300 N |
| Torsional (tor) | moment (0, 1, 0) · 565,000 N·mm |
| Solver | tetgen + conjugate-gradient, 25k node budget |

---

## Usage

**Download & extract**

```bash
huggingface-cli download KAIST-SmartDesignLab/DeepJEB-PP --repo-type dataset --local-dir DeepJEB-PP
cd DeepJEB-PP && for f in *.tar.gz; do tar -xzf "$f"; done
```

**Load one design**

```python
import numpy as np, pandas as pd, trimesh

case  = "012-015-diag_xz_mm_IS02"
mesh  = trimesh.load(f"1_surface_meshes/{case}.obj")
bc    = np.load(f"2_boundary_conditions/{case}.npz")   # bolt_idx, lug_idx
field = np.load(f"3_fea_fields/{case}.npz")            # ver_U, ver_vm, hor_U, ...
label = pd.read_csv("deepjebpp_labels.csv").set_index("case").loc[case]

clamped = field["surface_points"][bc["bolt_idx"]]      # clamped bolt nodes (mm)
vm_ver  = field["ver_vm"]                              # vertical-load von Mises (MPa)
```

**PyTorch dataloader** (geometry + fields + scalar targets)

```python
import os, glob, numpy as np, pandas as pd, torch
from torch.utils.data import Dataset

class DeepJEBPP(Dataset):
    """Per-case surface points, BC masks, per-load fields, and scalar labels."""
    LOADS = ["ver", "hor", "dia", "tor"]

    def __init__(self, root, load="ver"):
        self.root, self.load = root, load
        self.cases = sorted(os.path.splitext(os.path.basename(f))[0]
                            for f in glob.glob(f"{root}/3_fea_fields/*.npz"))
        self.labels = pd.read_csv(f"{root}/deepjebpp_labels.csv").set_index("case")

    def __len__(self):
        return len(self.cases)

    def __getitem__(self, i):
        c = self.cases[i]
        fld = np.load(f"{self.root}/3_fea_fields/{c}.npz")
        bc  = np.load(f"{self.root}/2_boundary_conditions/{c}.npz")
        pts = fld["surface_points"].astype("float32")
        n   = len(pts)
        bolt = np.zeros(n, "float32"); bolt[bc["bolt_idx"]] = 1.0   # clamped mask
        lug  = np.zeros(n, "float32"); lug[bc["lug_idx"]]  = 1.0    # loaded mask
        row  = self.labels.loc[c]
        return {
            "case": c,
            "points": torch.from_numpy(pts),                       # (N,3) mm
            "bc": torch.from_numpy(np.stack([bolt, lug], 1)),      # (N,2)
            "U":  torch.from_numpy(fld[f"{self.load}_U"].astype("float32")),   # (N,3)
            "vm": torch.from_numpy(fld[f"{self.load}_vm"].astype("float32")),  # (N,)
            "y":  torch.tensor([row["mass_g"],
                                row[f"{self.load}_p95vm"],
                                row[f"{self.load}_maxu"]], dtype=torch.float32),
        }

# ds = DeepJEBPP("DeepJEB-PP", load="ver"); print(len(ds), ds[0]["points"].shape)
```

---

## Applications

- **Surrogate modelling** — learn geometry → performance (mass, p95 von Mises, peak displacement, or full
  nodal fields) with point-cloud / mesh-GNN / implicit models; a 40× larger training corpus than SimJEB.
- **Field prediction** — predict per-node displacement and stress fields under each of the four load cases.
- **Generative & inverse design** — benchmark generators on a labelled, BC-aware bracket design space; close
  the loop with the released solver-input meshes.
- **Design optimisation** — data-driven optimisation / constraint screening using the mass and stress labels.
- **Cross-dataset transfer** — pre-train on DeepJEB++ and transfer to the smaller real SimJEB / DeepJEB sets.

---

## Citation

```bibtex
@article{deepjebpp2026,
  title   = {DeepJEB++: Foundation Model-Driven Large-Scale 3D Engineering
             Dataset via 2D Latent Space Augmentation},
  author  = {Yoo, Soyoung and Jeong, Leekyo and Ra, Jinsu and Lee, Dongeon
             and Yang, Sunwoong and Jeong, Hyogu and Kang, Namwoo},
  journal = {arXiv preprint arXiv:2606.12994},
  year    = {2026}
}
```

---

## Acknowledgements

DeepJEB++ builds on the **SimJEB** dataset (Whalen et al.) and the original **DeepJEB**, both derived from the
**GE Jet Engine Bracket Challenge** geometry, and uses the **TRELLIS** 3D foundation model for image-to-3D
generation. Developed at **KAIST SmartDesignLab**.

---

## License

Released under the **Open Data Commons Attribution License (ODC-By v1.0)**, matching the upstream
SimJEB / DeepJEB datasets. Derived from the SimJEB dataset (GE Jet Engine Bracket Challenge geometry).