DeepJEB

Overview

DeepJEB is a large-scale synthetic dataset of jet engine brackets generated through deep generative models and automated finite element analysis (FEA) simulation pipelines. It addresses the challenge of limited sample sizes in traditional engineering datasets by producing diverse, high-quality 3D geometries with corresponding structural simulation results.

The dataset was derived from the GE Jet Engine Bracket Challenge and enriched using deep generative models applied to the SimJEB dataset. Models trained on DeepJEB demonstrated significant improvements in surrogate model performance — achieving up to a 23% increase in the coefficient of determination and over 70% reduction in mean absolute percentage error compared to traditional datasets.

Applications

• Surrogate model training for structural performance prediction
• 3D data processing and machine learning tasks requiring large data volumes
• Performance prediction using deep learning (GNNs, CNNs)
• Inverse problem resolution in structural engineering
• Design optimization and generative design research
• Benchmarking data-driven surrogate models

Load Cases

Each bracket is structurally analyzed under four distinct load scenarios, following the original GE challenge specifications:

• Vertical — Static, 8,000 lbs upward
• Horizontal — Static, 8,500 lbs horizontal out.
• Diagonal — Static, 42° from vertical, 9,500 lbs
• Torsional — Static torsional, 5,000 lb-in at the horizontal plane

In addition, two natural frequency analysis results are provided per sample.

Dataset Contents

Each of the 2,138 brackets includes the following files:

File Type	Format	Description
CAD Model	STEP	3D solid geometry for each bracket design
Volume Mesh	VTK	Tetrahedral volume mesh for FEA
Surface Mesh	STL	Triangulated surface representation
Analysis Results	CSV	Nodal stress/displacement results from 4 load cases + modal analysis
Solver Input	FEM	Complete FEA solver input file
Multi-view Image	PNG	Images of the bracket from various angles

Cluster Organization

Brackets are automatically classified into 20 groups using unsupervised 3D clustering technology based on shape similarity, each containing 50 curated data entries. This organization helps researchers select representative subsets for training or evaluation.

Key Features

• Data augmented 10x using deep learning from the original SimJEB dataset
• High-accuracy analysis with 2nd-order tetrahedral elements
• Signed von Mises Stress included (distinguishing tensile and compressive stresses)
• Model body information: center of gravity, moment of inertia
• Natural frequency analysis with mode shapes
• Multi-view rendered images for each bracket (26 views per sample)

Citation

If you use this dataset in your research, please cite: