Description

AI-RAN AODT Digital Twin Solution

1. Solution Overview

The AI-RAN AODT Digital Twin Solution is a turnkey wireless-network simulation and validation platform built around NVIDIA Aerial Omniverse Digital Twin, or AODT. It enables researchers, operators, network vendors and engineering teams to create physically accurate digital representations of 5G and 6G radio environments, ranging from a single radio site to city-scale deployments.

The platform combines GPU-accelerated electromagnetic simulation, realistic 3D environments, antenna and mobility models, a software-defined RAN simulation stack, AI/ML interfaces and geospatial visualization. It can be used to design radio networks, generate site-specific channel data, develop AI-native RAN algorithms and validate network behavior before installing physical infrastructure.

The commercial solution is based on the AODT 1.5 architecture, which separates the system into three principal components:

• AODT Client: configures scenarios and controls simulation execution.
• AODT Worker: performs GPU-accelerated EM and RAN simulation.
• AODT Viewer: visualizes scenes, radio units, user equipment, mobility and propagation rays through a web browser.

The platform supports both on-premises and cloud deployment, as well as colocated and distributed client-worker configurations.

---

2. Reference System Architecture

A typical AODT deployment follows this processing chain:

GIS and 3D Scene Data → AODT World Model → Mobility and Antenna Models → GPU-Accelerated EM Solver → RAN L1/L2 Simulation → AI/ML Applications → Results Database and 3D Viewer

The architecture can also be incorporated into NVIDIA’s broader AI-RAN development workflow:

Train → Simulate → Deploy

• Train: develop wireless algorithms and AI models with NVIDIA Sionna, NVIDIA Aerial and accelerated-computing systems.
• Simulate: evaluate the algorithms in physics-based RF environments using AODT.
• Deploy: transfer validated algorithms to an NVIDIA Aerial RAN Computer or an Aerial hardware testbed for live-network operation.

AODT can operate as a universal simulation service. A lightweight client communicates with the centralized GPU worker through gRPC, allowing integration with Python, C++ and other simulation or AI-development environments.

---

3. Hardware Configuration

3.1 AODT Simulation Worker Server

Recommended Qualified On-Premises Configuration

Dell PowerEdge R750 AODT Simulation Server

Component	Recommended specification
CPU	Intel Xeon Gold 6336Y
CPU Cores	24 cores / 48 threads
GPU	NVIDIA L40S or NVIDIA RTX 6000 Ada
System Memory	512 GB DDR4
Storage	2 TB enterprise SSD or NVMe storage
Operating System	Ubuntu Server 22.04
Deployment Role	AODT Worker, EM Solver, RAN Simulator and GIS Processing

This configuration corresponds to NVIDIA’s qualified on-premises AODT worker platform. The GPU executes the electromagnetic propagation calculations, antenna processing, channel generation and RAN simulation workloads.

Supported GPU Options

AODT 1.5 worker systems support the following accelerator classes:

• NVIDIA RTX 6000 Ada
• NVIDIA L40
• NVIDIA L40S on qualified systems
• NVIDIA A100
• NVIDIA H100

The appropriate accelerator depends on scene size, ray count, antenna configuration, simulation duration and the required turnaround time.

---

3.2 AODT Client Workstation

The client workstation is used to define simulations, build YAML configurations, invoke the AODT service, retrieve results and perform post-processing.

Component	Recommended specification
CPU	Modern Intel, AMD or ARM multicore processor
Memory	32 GB or higher
Storage	1 TB SSD
GPU	Not required for the AODT client
Operating System	Linux, Windows or macOS
Network	Gigabit Ethernet minimum; 10GbE recommended for large datasets

The official AODT architecture does not require a GPU on the client computer. The client and worker may run on the same Linux host or on different systems connected through a network.

---

3.3 Visualization Terminal

The AODT Viewer runs through a modern web browser and displays:

• 3D terrain and buildings
• Radio-unit positions
• User-equipment positions
• Propagation rays
• Mobility trajectories
• Simulation time progression
• Entity and antenna properties
• Geospatial network layouts

The current viewer uses a CesiumJS-based interface designed for responsive visualization of large geospatial scenes, 3D Tiles and quantized-mesh terrain.

---

3.4 Storage and Simulation Database

The standard local deployment includes:

• MinIO S3-compatible object storage
• Nessie Iceberg catalog
• Parquet simulation-result tables
• Optional H5 waveform files
• Containerized AODT services

For cloud deployment, Amazon S3 and AWS Glue can replace the local object store and catalog.

Recommended commercial storage configuration:

Component	Recommended specification
Active simulation storage	2–8 TB NVMe
Long-term dataset storage	8 TB or higher
File formats	Parquet, H5 and AODT scene assets
Object storage	MinIO or Amazon S3
Catalog	Nessie/Iceberg or AWS Glue

---

3.5 Cloud Deployment Option

AODT may also be deployed on AWS. NVIDIA lists the following qualified reference configuration:

Component	AWS reference configuration
Instance	g6e.xlarge
GPU	NVIDIA L40S
vCPU	4
System Memory	32 GB
Operating System	Ubuntu Server 22.04
Deployment	Colocated client and worker

Cloud deployment is suitable for temporary projects, parallel experiments, remote collaboration and simulation workloads that do not justify maintaining permanent on-premises infrastructure.

---

3.6 Optional AI-RAN Development Hardware

The commercial suite may be expanded with:

• NVIDIA DGX or DGX Spark for AI-model training
• NVIDIA Aerial CUDA-Accelerated RAN server
• Aerial Testbed with commercial O-RU
• NVIDIA Aerial RAN Computer for field deployment
• RF channel emulator
• Network measurement and calibration equipment
• High-capacity storage server
• Additional AODT worker nodes

These components create an integrated workflow in which models are trained on accelerated systems, evaluated in AODT and then deployed on a real-time Aerial RAN platform.

---

4. Software Stack

The standard software package includes:

• NVIDIA Aerial Omniverse Digital Twin
• GPU-accelerated AODT EM Solver
• AODT antenna engine
• AODT mobility model
• AODT RAN L1/L2 simulation stack
• AODT GIS and scene-generation pipeline
• AODT Python and C++ client libraries
• gRPC-based client-server interface
• CesiumJS AODT Viewer
• Docker and NVIDIA Container Toolkit
• MinIO object storage
• Nessie and Apache Iceberg catalog services
• Parquet and H5 data-processing utilities

AODT 1.5 adopts a cloud-native, headless-first architecture. Simulations can be configured programmatically, executed in batches and integrated directly into automated development or continuous-validation pipelines.

---

5. Main Functions

5.1 Physics-Based Electromagnetic Simulation

The AODT EM Solver generates site-specific radio-channel information from a three-dimensional representation of the physical environment.

It can model:

• Radio propagation through complex urban environments
• Building and surface materials
• Terrain effects
• Vegetation and canopy transmission losses
• Reflection and scattering
• Diffraction
• Outdoor-to-indoor propagation
• Directional antennas and antenna arrays
• Dynamic scatterers
• Mobile radio units and user equipment

The primary outputs include channel impulse responses, channel frequency responses, propagation paths, path power, delay, angle of arrival and angle of departure.

---

5.2 High-Fidelity 3D Scene Generation

AODT can convert geographic and engineering data into simulation-ready three-dimensional environments.

Supported scene sources include:

• OpenStreetMap
• CityGML and other GML-based building data
• Digital terrain models
• Public elevation datasets
• Customer-provided building models
• Material and vegetation data
• Site-specific antenna and radio-unit information

The GIS pipeline generates simulation assets, Cesium 3D Tiles and terrain data that can be stored in S3-compatible storage and accessed by both the simulation engine and viewer.

---

5.3 Radio-Network Planning

The platform can be used to evaluate:

• Radio-site placement
• Antenna height
• Antenna azimuth and mechanical tilt
• Beam configuration
• Coverage distribution
• Inter-cell interference
• Indoor coverage
• Outdoor-to-indoor penetration
• Massive-MIMO configuration
• Mobility and handover scenarios
• Network densification strategies

Because AODT uses site-specific geometry and material properties, it provides a more deterministic environment than conventional statistical channel models.

---

5.4 RAN System Simulation

AODT includes a batched RAN mode that combines the EM Solver with a 5G RAN Layer 1 and Layer 2 transmit-and-receive chain.

The RAN simulation produces:

• Channel impulse and frequency responses
• Per-UE throughput
• Block error rate
• Proportional-fair scheduling metrics
• MCS and link-adaptation information
• Frequency-time grid data
• Receiver waveform dumps
• Slot-level RAN telemetry

The RAN mode follows a slot-based simulation timeline to preserve 5G processing behavior.

---

5.5 Massive-MIMO and Multiuser Simulation

The AODT RAN environment supports:

• Up to 64 antennas at a radio unit
• Downlink MU-MIMO
• Uplink MU-MIMO
• SRS transmission
• SRS channel estimation
• SRS-based beamforming
• Configurable target BLER for individual UEs
• Custom antenna patterns
• Per-element antenna characteristics

These functions allow engineers to evaluate beamforming, channel estimation, scheduling and adaptive-transmission algorithms under site-specific propagation conditions.

---

5.6 User and Vehicle Mobility

AODT generates time-indexed positions, orientations and speeds for UEs and moving scatterers.

Supported mobility modes include:

• Outdoor 2D mobility
• Outdoor 3D mobility
• Indoor mobility
• Manually defined waypoint routes
• GPX-based measurement or drive-test routes
• SUMO-based urban vehicle and pedestrian movement
• UAV and altitude-aware scenarios
• Procedurally generated UE distributions
• Dynamic vehicle scattering

The mobility model uses GPU-accelerated pathfinding to generate collision-free routes around terrain and buildings.

---

5.7 Digital-Twin Calibration

The AODT calibration workflow aligns the simulation more closely with real-world RF measurements.

Calibration can optimize or recover:

• Building material parameters
• Surface material properties
• Vegetation parameters
• Radio-unit orientation
• UE orientation along measurement routes
• Radio-unit beam codebooks
• Site-specific propagation characteristics

Measurement data collected from an actual network can therefore be used to improve the predictive accuracy of the digital twin.

---

5.8 AI and Machine-Learning Integration

AODT can generate physically grounded datasets for training and validating AI-RAN algorithms.

Typical applications include:

• Neural channel estimation
• Neural receivers
• Beam selection
• Beam prediction
• AI-assisted link adaptation
• RAN scheduling
• Mobility prediction
• Radio-resource optimization
• Coverage optimization
• Interference prediction
• Integrated sensing and communications
• Reinforcement-learning-based RAN control

AODT has demonstrated AI-based PUSCH and SRS channel-estimation workflows and can expose position and channel information to external models.

---

5.9 Interactive and Batched Operation

Two principal simulation workflows are available:

Batched Mode

The complete scenario is executed first, after which the results are exported for analysis. This mode is suitable for dataset generation, benchmarking, automated testing and parameter sweeps.

Interactive Mode

The client requests channel information at individual simulation steps. This mode is suitable for closed-loop AI training, external RAN simulators and applications that must respond to the evolving channel during execution.

---

5.10 API and Workflow Automation

AODT provides programmatic interfaces for:

• Scenario creation
• RU and UE configuration
• Antenna configuration
• Mobility generation
• EM simulation
• RAN simulation
• Calibration
• Result extraction
• Channel-data access
• Automated experiment execution
• CI/CD-style validation

The AODT client provides C++ and Python libraries, while the gRPC service architecture allows external tools and simulation chains to connect to the GPU-accelerated worker.

---

6. Performance and Scalability

6.1 Electromagnetic Simulation Scale

AODT 1.5 supports:

Performance parameter	AODT capability
Maximum ray emission	Up to 40 million rays per RU
Reference simulation area	Up to 25 km²
Reference GPU	NVIDIA L40S
Maximum RU antenna count	Up to 64 antennas
Simulation scope	Single radio site to city scale
MIMO capability	Uplink and downlink MU-MIMO
Deployment	Single-GPU qualified worker supported

The 40-million-ray and 25-square-kilometer figures are documented for an NVIDIA L40S worker. Actual capacity and simulation time depend on scene geometry, ray-interaction depth, antenna count, number of RUs and UEs, mobility, simulation length and output settings.

---

6.2 Fast EM Mode

AODT 1.5 includes a Fast EM mode for workloads that prioritize simulation turnaround.

It applies:

• Plane-wave approximation for MIMO panels
• Steering-vector derivation from the array-center propagation path
• Doppler phase-shift approximation for slot-based simulations
• Reduced repeated ray tracing across OFDM symbols

Fast EM mode reduces computational workload but may introduce an accuracy tradeoff for electrically large arrays, near-field propagation or rapidly changing channels.

---

6.3 GPU-Accelerated Channel Access

The service architecture centralizes the complex EM calculations on the AODT worker. The client can access channel outputs in memory, while GPU inter-process communication enables efficient transfer of channel information into external simulators and AI pipelines.

This architecture is suitable for:

• Closed-loop AI-RAN simulation
• Large channel-dataset generation
• Link-level simulator integration
• System-level simulator integration
• Hardware-independent RAN validation
• Automated scenario benchmarking

---

6.4 Data Output Performance

AODT stores batched results as Parquet tables in S3-compatible storage and registers them through an Iceberg catalog. Optional H5 waveform files allow detailed inspection of received time-frequency I/Q data.

Available output categories include:

• Simulation setup
• Channel impulse responses
• Channel frequency responses
• Ray paths
• Antenna and entity configuration
• Mobility data
• RAN telemetry
• Throughput
• BLER
• Proportional-fair metrics
• Optional received waveform grids

---

6.5 Performance Qualification

AODT does not have a single universal “real-time factor” applicable to every simulation. Execution time changes substantially with:

• Number of emitted rays
• Number of reflections, diffractions and scattering interactions
• Physical area and scene complexity
• Number of RUs and UEs
• Antenna-array dimensions
• Number of MIMO links
• Mobility sampling interval
• Number of time steps
• Full RAN versus EM-only operation
• Standard versus Fast EM mode

For commercial projects, system acceptance should therefore use a customer-specific benchmark scene and a defined simulation configuration rather than a generic runtime figure.

---

7. Typical Applications

The AODT solution is suitable for:

• AI-RAN research and development
• 5G Advanced and 6G system design
• Network digital-twin construction
• Radio-site planning
• Coverage and capacity analysis
• Massive-MIMO algorithm development
• Neural-receiver dataset generation
• Beamforming and precoding research
• Integrated sensing and communications
• RAN software validation
• Autonomous-network training
• Urban, indoor and campus-network simulation
• UAV communication research
• Digital drive testing
• Predeployment network validation
• Commercial network optimization

---

8. Delivery Scope

A complete commercial AODT package can include:

• Qualified AODT GPU simulation server
• NVIDIA L40S or RTX 6000 Ada GPU
• AODT client workstation
• High-capacity simulation storage
• Network and management accessories
• NVIDIA AODT installation
• Docker and NVIDIA Container Toolkit deployment
• Local S3-compatible storage and catalog configuration
• AODT client and worker configuration
• CesiumJS viewer deployment
• Sample 3D wireless scene
• OSM or CityGML scene-import workflow
• EM simulation example
• RAN simulation example
• Mobility and dynamic-scattering example
• AI/ML interface example
• Calibration workflow preparation
• API and automation scripts
• System documentation
• Developer training
• Remote technical support
• Optional on-site installation

---

9. Key Advantages

Physics-Based Network Modeling

AODT calculates radio propagation from three-dimensional geometry, terrain, materials, antennas and moving objects rather than relying exclusively on simplified statistical models.

End-to-End Digital-Twin Workflow

The same environment can support scene creation, EM simulation, RAN simulation, AI-model development, visualization and data analysis.

City-Scale GPU Acceleration

AODT uses NVIDIA GPUs to process large 3D environments and high ray counts, supporting simulations from individual radio sites to city-scale areas.

AI-Native Architecture

Channel and RAN data can be integrated directly into AI-training and inference workflows for neural receivers, channel estimation, beam management and network optimization.

Open Integration Interfaces

Python, C++, gRPC, YAML configuration and structured output formats simplify integration with customer applications and third-party simulation tools.

Cloud and On-Premises Deployment

The platform can run on a qualified local GPU server or in an AWS environment, depending on data-security, scalability and operational requirements.

---

Product Summary

The AI-RAN AODT Digital Twin Solution provides a scalable, GPU-accelerated platform for constructing realistic digital representations of wireless networks. By integrating site-specific 3D environments, physics-based electromagnetic simulation, mobility, RAN processing and AI-development interfaces, the solution enables communication systems to be designed, trained and validated before they are introduced into a physical network.

It provides the simulation bridge between wireless algorithm development and commercial AI-RAN deployment.