Datasets:

cmudrc
/

OpenSeeSimE-Structural

dataset_info:
  features:
    - name: file_name
      dtype: string
    - name: source_file
      dtype: string
    - name: question
      dtype: string
    - name: question_type
      dtype: string
    - name: question_id
      dtype: int32
    - name: answer
      dtype: string
    - name: answer_choices
      list: string
    - name: correct_choice_idx
      dtype: int32
    - name: image
      dtype: image
    - name: video
      dtype: video
    - name: media_type
      dtype: string
  splits:
    - name: test
      num_bytes: 187015546578
      num_examples: 102678
  download_size: 175022245655
  dataset_size: 187015546578
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
license: mit
task_categories:
  - visual-question-answering
language:
  - en
size_categories:
  - 100K<n<1M

OpenSeeSimE-Structural: Engineering Simulation Visual Question Answering Benchmark

Dataset Summary

OpenSeeSimE-Structural is a large-scale benchmark dataset for evaluating vision-language models on structural analysis simulation interpretation tasks. It contains over 100,000 question-answer pairs across parametrically-varied structural simulations including stress analysis, and deformation patterns.

Purpose

While vision-language models (VLMs) have shown promise in general visual reasoning, their effectiveness for specialized engineering simulation interpretation remains largely unexplored. This benchmark enables:

Statistically robust evaluation of VLM performance on engineering visualizations
Assessment across multiple reasoning capabilities (captioning, reasoning, grounding, relationship understanding)
Evaluation using different question types (binary classification, multiple-choice, spatial grounding)

Dataset Composition

Statistics

Total instances: 102,678 question-answer pairs
Simulation types: 5 structural models (Dog Bone, Hip Implant, Pressure Vessel, Beams, Wall Bracket)
Parametric variations: 1,024 unique instances per base model (4^5 parameter combinations)
Question categories: Captioning, Reasoning, Grounding, Relationship Understanding
Question types: Binary, Multiple-choice, Spatial grounding
Media formats: Both static images (1920×1440 PNG) and videos (Originally Extracted at: 200 frames, 29 fps, 7 seconds)

Simulation Parameters

Each base model varies across 5 parameters with 4 values each:

Dog Bone: Length, Thickness, Diameter, Axial Load, Bending Load
Hip Implant: Beam Length, Beam Diameter, Ball Diameter, Axial Load, Bending Load
Pressure Vessel: Length, Thickness, Diameter, Material, Pressure
Thermal Beam: Thickness, Bending Load, Young's Modulus, Tensile Yield Strength, Cross Section Shape
Wall Bracket: Length, Width, Height, Thickness, Bending Force

Question Distribution

Binary Classification: 40% (yes/no questions about symmetry, stress types, uniformity, etc.)
Multiple-Choice: 30% (4-option questions about deformation direction, stress dominance, magnitude ranges, etc.)
Spatial Grounding: 30% (location-based questions with labeled regions A/B/C/D)

Data Collection Process

Simulation Generation

Base models sourced from Ansys Mechanical tutorial files
Parametric automation via PyMechanical and PyGeometry interfaces
Systematic variation across 5 parameters with 4 linearly-spaced values
All simulations solved using finite element analysis with validated convergence settings

Ground Truth Extraction

Automated extraction eliminates human annotation costs and ensures consistency:

Statistical Analysis: Direct queries on result arrays (max, min, mean, std)
Distribution Analysis: Threshold-based classification using coefficient of variation
Physics-Based Classification: Stress tensor analysis and mechanics principles
Spatial Localization: Color-based region generation with computer vision algorithms

All ground truth derived from numerical simulation results rather than visual interpretation.

Preprocessing and Data Format

Image Processing

Resolution: 1920×1440 pixels
Format: PNG with lossless compression
Standardized viewing orientations: front, back, left, right, top, bottom, isometric
Consistent color mapping: rainbow gradients (red=maximum, blue=minimum)
Automatic deformation scaling (1.5× relative to maximum dimension)

Video Processing

200 frames at 29 fps (7 seconds duration)
Maximum deformation at frame 100 (temporal midpoint)
H.264 compression at 1920×1440 resolution
Uniform frame sampling for model input (32 frames)

Data Fields

{
    'file_name': str,           # Unique identifier
    'source_file': str,         # Base simulation model
    'question': str,            # Question text
    'question_type': str,       # 'Binary', 'Multiple Choice', or 'Spatial'
    'question_id': int,         # Question identifier (1-20)
    'answer': str,              # Ground truth answer
    'answer_choices': List[str], # Available options
    'correct_choice_idx': int,  # Index of correct answer
    'image': Image,             # PIL Image object (1920×1440)
    'video': Video,             # Video frames
    'media_type': str          # 'image' or 'video'
}

Labels

All labels are automatically generated from simulation numerical results:

Binary questions: "Yes" or "No"
Multiple-choice: Single letter (A/B/C/D) or descriptive option
Spatial grounding: Region label (A/B/C/D) corresponding to labeled visualization locations

Label generation employs domain-specific thresholds:

Uniformity: CV ≤ 0.2 (20%)
Symmetry: 60% of node pairs within 10% tolerance (structural)
Spatial matching: 50-pixel separation for region placement

Dataset Splits

Test split only: 102,678 instances
No train/validation splits provided (evaluation benchmark, not for model training)
Representative sampling across all simulation types and question categories

Intended Use

Primary Use Cases

Benchmark evaluation of vision-language models on engineering simulation interpretation
Capability assessment across visual reasoning dimensions (captioning, spatial grounding, relationship understanding)
Transfer learning analysis from general-domain to specialized technical visual reasoning

Out-of-Scope Use

Real-time engineering decision-making without expert validation
Safety-critical applications without human oversight
Generalization to simulation types beyond structural mechanics

Limitations

Technical Limitations

Objective tasks only: Excludes subjective engineering judgments requiring domain expertise
Single physics domain: Structural mechanics only (see OpenSeeSimE-Fluid for fluid dynamics)
Ansys-specific: Visualizations generated using Ansys Mechanical rendering conventions
Static parameters: Fixed material properties and boundary conditions per instance
2D visualizations: All inputs are 2D projections of 3D simulations

Known Biases

Color scheme dependency: Questions exploit default rainbow gradient conventions
Geometry bias: Selected simulation types may not represent full diversity of structural analysis applications
View orientation bias: Standardized camera positions may not capture all critical simulation features

Ethical Considerations

Responsible Use

Models evaluated on this benchmark should NOT be deployed for safety-critical engineering decisions without expert validation
Automated interpretation should augment, not replace, human engineering expertise
Users should verify that benchmark performance translates to their specific simulation contexts

Data Privacy

All simulations contain no proprietary or confidential engineering data
No personal information collected
Publicly available tutorial files used as base models

Environmental Impact

Dataset generation required significant computational CPU resources
Consider environmental cost of large-scale model evaluation on this benchmark

License

MIT License - Free for academic and commercial use with attribution

Citation

If you use this dataset, please cite:

@article{ezemba2024opensesime,
  title={OpenSeeSimE: A Large-Scale Benchmark to Assess Vision-Language Model Question Answering Capabilities in Engineering Simulations},
  author={Ezemba, Jessica and Pohl, Jason and Tucker, Conrad and McComb, Christopher},
  year={2025}
}

AI Usage Disclosure

Dataset Generation

Simulation automation: Python scripts with Ansys PyMechanical interface
Ground truth extraction: Automated computational protocols (no AI involvement)
Quality validation: Expert oversight of automated extraction procedures
No generative AI used in dataset creation, labeling, or curation

Visualization Generation

Ansys Mechanical rendering engine (deterministic, physics-based)
Standardized color mapping and camera controls
No AI-based image generation or enhancement

Contact

Authors: Jessica Ezemba (jezemba@andrew.cmu.edu), Jason Pohl, Conrad Tucker, Christopher McComb
Institution: Department of Mechanical Engineering, Carnegie Mellon University

Acknowledgments

Ansys for providing simulation software and tutorial files
Carnegie Mellon University for computational resources
Reviewers and domain experts who validated the automated extraction protocols

Version: 1.0
Last Updated: December 2025
Status: Complete and stable