DeepJEB++: Foundation Model-Driven Large-Scale 3D Engineering Dataset via 2D Latent Space Augmentation

2026-06-11 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Engineering & Applied Sciences · Depth: Expert, quick

Summary

DeepJEB++ is a foundation-model-driven data-augmentation framework designed to generate large-scale, simulation-labeled 3D engineering datasets from small seed sets. It addresses limitations in existing 3D augmentation techniques and automates the simulation-labeling process. The framework expands a small set of jet engine brackets (fewer than 400 designs) into 15,360 simulation-labeled 3D brackets, achieving a 40x expansion using a single GPU per stage. Its core methodology involves augmenting data in a 2D latent space before transferring to 3D. This process includes fine-tuning a 2D latent diffusion model, synthesizing novel views with a vision-language-model (VLM) quality filter, lifting validated images to 3D meshes, and automatically assigning finite-element labels like mass, stress, and displacement.

Key takeaway

For Machine Learning Engineers or Research Scientists constrained by limited 3D engineering data or manual simulation labeling, DeepJEB++ offers a scalable solution. You should consider adopting this 2D-to-3D augmentation strategy to rapidly generate large, labeled 3D datasets. This approach significantly reduces resource demands and manual intervention for complex engineering simulations, enabling faster iteration in data-driven design.

Key insights

2D latent space augmentation efficiently expands small 3D engineering datasets with physics-based labels.

Principles

• Augmenting in 2D latent space is efficient for 3D data.
• VLM filtering ensures manufacturable design quality.
• Automated pipelines streamline physics-based labeling.

Method

Fine-tune a 2D latent diffusion model, synthesize views with VLM filtering, lift 2D images to 3D meshes via a generative foundation model, then automate finite-element label assignment.

In practice

• Expand 3D engineering datasets 40x with single GPU.
• Automate physics-based performance label generation.
• Filter generated designs for manufacturability.

Topics

• 3D Data Augmentation
• Foundation Models
• Latent Diffusion Models
• Engineering Design
• Finite Element Analysis
• Jet Engine Brackets

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Dreaming in Cubes — A VQ-VAE and Transformer pipeline can generate semantically consistent 3D voxel terrain from a compressed latent space.
• Latent Anchor-Driven Test Generation for Deep Neural Networks — Latte employs anchor-guided, seed-centric latent space exploration to generate diverse, fault-revealing DNN test cases with low semantic drift.
• AnimateAnyMesh++: A Flexible 4D Foundation Model for High-Fidelity Text-Driven Mesh Animation — AnimateAnyMesh++ is a 4D foundation model for high-fidelity, text-driven mesh animation.
• A Unified Latent Space Disentanglement VAE Framework with Robust Disentanglement Effectiveness Evaluation — Robustly evaluate VAE latent space disentanglement and interpretability without ground-truth generative factors using novel assessment tools.
• Latent Generative Solvers for Generalizable Long-Term Physics Simulation — LGS uses latent space, generative modeling, and an uncertainty knob for stable, generalizable, and efficient long-term PDE simulation.
• Building Generative Image & Video models at Scale - Sander Dieleman (Veo and Nano Banana) — Diffusion models generate high-quality audiovisual data through iterative denoising of learned compressed latent representations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.