NVIDIA Scionna2M: Blender-Powered Generative AI Dataset for RF Channel Models

NVIDIA Scionna2M is a generative AI project supported by NVIDIA's Academic Grant Program. It focuses on building a large-scale multimodal dataset based on Blender prompts and electromagnetic simulations for generative AI research on RF channel models. This project aims to address the limitations of traditional channel modeling methods and the challenges of RF channel data acquisition, providing high-quality training resources for generative AI applications in the wireless communication field.

Traditional RF channel modeling relies on mathematical statistical models (overly simplified) or complex electromagnetic simulations (high computational cost). Generative AI has great potential in RF channel modeling, but high-quality training data is scarce—due to high acquisition costs, limited scenarios, difficult annotation, and privacy sensitivity. The Scionna2M project was born to fill this gap by building a multimodal dataset.

Reasons for Choosing Blender

Blender has powerful 3D scene modeling, procedural generation, open-source ecosystem, and physical ray tracing capabilities, making it suitable for large-scale scene construction.

Core Workflow

1. Scene Modeling: Build 3D scenes in Blender that include geometric structures, material properties, and antenna positions;
2. Rendering: Extract visual data (images, depth maps, etc.);
3. Electromagnetic Simulation: Export the scene to professional tools to calculate channel parameters (e.g., multipath effects, fading characteristics);
4. Data Alignment: Integrate visual, geometric, and electromagnetic data to form multimodal samples.

Large-Scale Data Generation

Generate 2 million samples through strategies such as scene variations (layout, material, antenna adjustments), parameterized sampling (frequency, bandwidth), and domain randomization.

Multimodal Components

• Visual Modality: RGB images, depth maps, semantic segmentation maps, normal maps;
• Geometric Modality: Scene meshes, material properties (e.g., dielectric constant);
• Electromagnetic Modality: Channel impulse response, frequency response, path loss, MIMO matrix;
• Metadata: Scene category, simulation parameters, environmental conditions.

Scene Coverage

Includes indoor (offices, residences, commercial spaces), outdoor (urban streets, suburbs), and dynamic scenes (moving obstacles, time-varying channels).

Channel Generation Models

• Conditional Generation Models: Generate channel responses based on scene descriptions (text/images) for fast simulation and data generation;
• Image-to-Channel Models: Predict channel characteristics from input scene images to assist network planning;
• Text-to-Channel Models: Generate channel statistical characteristics via natural language descriptions.

Downstream Tasks

Supports research on channel estimation, beamforming, positioning awareness, communication system design, etc.

Blender Automation

• Procedural Scene Generation: Batch create scenes using Python scripts;
• Material Library: Contains electromagnetic properties of common materials;
• Render Farm: Distributed processing for large-scale rendering.

Electromagnetic Simulation Integration

Combines ray tracing (high frequency), finite element/time-domain finite difference (low frequency), and hybrid methods.

NVIDIA Ecosystem

Uses GPU acceleration for rendering/simulation/training, and integrates tools like Omniverse, Modulus, TensorRT, and NeMo.

• 6G Research: Millimeter wave/terahertz channels, RIS-assisted communication, integrated communication and sensing;
• Network Planning: Evaluate coverage quality, optimize base station configuration;
• V2X: Dynamic scene channel modeling, multi-vehicle collaborative communication;
• IoT: Indoor positioning, LPWAN coverage planning, industrial IoT reliability analysis.

Limitations

There is a gap between simulation accuracy and reality; scene coverage is incomplete; computational cost is high.

Future Directions

Integrate real data; expand dynamic scenes; support multi-band; develop physically constrained generative models.

Open-Source Plan

Plan to open-source part or all of the dataset, generation tools, benchmark models, and evaluation tools; encourage the community to contribute new scenes, simulation methods, and models.