⚡ electrical engineering

NYUSIM: A Roadmap to AI-Enabled Statistical Channel Modeling and Simulation

This paper presents the migration of the NYUSIM channel modeling framework from MATLAB to Python, introducing new 6G spectrum models and a standardized 3D antenna format while rigorously validating statistical consistency to establish a scalable foundation for AI-enabled wireless channel simulation.

Original authors: Isha Jariwala, Xinquan Wang, Bridget Meier, Guanyue Qian, Dipankar Shakya, Mingjun Ying, Homa Nikbakht, Daniel Abraham, Theodore S. Rappaport

Published 2026-02-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Isha Jariwala, Xinquan Wang, Bridget Meier, Guanyue Qian, Dipankar Shakya, Mingjun Ying, Homa Nikbakht, Daniel Abraham, Theodore S. Rappaport

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a self-driving car, but you can't test it on real streets yet because it's too dangerous or expensive. Instead, you need a perfect video game simulation that mimics real traffic, weather, and road conditions so the car's AI can learn how to drive.

In the world of wireless communication (like your phone connecting to 5G or the future 6G), NYUSIM is that video game. It's a super-advanced simulator that predicts how radio waves travel through cities, buildings, and forests.

This paper is about a major upgrade to NYUSIM. Here is the breakdown in simple terms:

1. The Big Move: From "Excel" to "Python"

For years, NYUSIM was built using a programming language called MATLAB. Think of MATLAB as a powerful, but slightly old-fashioned, calculator that works great for one person at a time. It's like a single-lane road.

The researchers moved the entire system to Python.

The Analogy: Imagine moving from a single-lane country road to a massive, multi-lane superhighway.
Why it matters: Python is the language of Artificial Intelligence (AI). By switching to Python, the simulator can now talk directly to AI brains. It can also run on thousands of computers at once (parallel processing), generating millions of "what-if" scenarios in minutes instead of years. This is crucial for training the AI that will power 6G networks.

2. The New "Antenna Glasses" (Ant3D)

Radio waves don't just travel in a straight line; they bounce off buildings, trees, and cars. To understand this, you need to know exactly what the "lens" (the antenna) looks like.

The Old Way: Previously, the simulator assumed antennas were perfect, simple shapes (like a flat sheet of light).
The New Way (Ant3D): The researchers introduced a new format called Ant3D.
The Analogy: Imagine you used to wear simple sunglasses that only blocked light from the front. Now, you've been given 360-degree VR goggles that show you exactly how light hits you from every angle, including the sides and top.
The Result: The simulator can now use real-world antenna designs (like the ones on your phone or a cell tower) to see exactly how they distort radio waves, making the simulation incredibly realistic.

3. Exploring New Territories (FR3 and FR1(C))

Wireless signals use different "frequencies" (like different radio stations).

The Past: NYUSIM was great at simulating high-frequency signals (like 28 GHz and up), which are fast but travel short distances.
The New Frontier: The researchers have now mapped out the "Upper Mid-Band" (frequencies around 6.75 GHz and 16.95 GHz).
The Analogy: Think of the old frequencies as sprinters (fast but get tired quickly). The new frequencies are like marathon runners (they can go further and carry more data than older signals, but aren't as fast as the sprinters).
Why it matters: These new frequencies are the "Goldilocks zone" for 6G. They offer a perfect balance of speed and distance. The researchers went out into New York City, measured these waves in real life, and taught the simulator how to mimic them perfectly.

4. The "Twin" Test (Verification)

When you move a complex system from one language to another, you worry you might break something.

The Process: The team didn't just guess; they ran the old MATLAB version and the new Python version side-by-side millions of times.
The Result: They used statistical tests (like a "lie detector" for data) to prove that the new Python version produces exactly the same results as the old one.
The Analogy: It's like taking a classic recipe written in a handwritten notebook and translating it into a digital app. They tasted the dish made by the app and compared it to the original. If they taste identical, the translation is a success.

Why Should You Care?

This paper isn't just about code; it's about preparing for the future.

AI Needs Data: To build AI that manages our future networks, we need massive amounts of realistic data. The new Python version can generate this data instantly.
Better 6G: By accurately simulating these new "mid-band" frequencies and using realistic antenna shapes, engineers can design cell towers and phones that work better, faster, and with fewer dropped calls.

In a nutshell: The researchers took their famous radio-wave simulator, upgraded its engine to run on AI-friendly fuel (Python), gave it 3D glasses to see the real world better, and mapped out new territories for the next generation of internet. It's a robust, verified foundation for the 6G revolution.

1. Problem Statement

The evolution of wireless communication toward 6G requires accurate channel models that can support higher frequencies (including the upper mid-band/FR3), wider bandwidths, and complex spatial antenna patterns. However, existing channel simulators face three critical limitations hindering the integration of Artificial Intelligence (AI) and Machine Learning (ML):

Restricted Data Access: Proprietary or compiled code (e.g., legacy MATLAB versions) often embeds measurement datasets and parameter generation logic, making it difficult to export large-scale data for AI training.
Limited Scalability: Traditional simulators are not designed for parallel execution or cloud computing, preventing the generation of the millions of labeled/unlabeled samples required for modern deep learning models.
Limited Interoperability: Closed systems do not integrate cleanly with modern AI frameworks (PyTorch, TensorFlow, JAX), creating barriers for embedding learned models directly into simulation workflows (e.g., for digital twins or AI-native PHY/MAC co-design).

Additionally, there is a lack of standardized, measurement-based channel models for the emerging FR3 band (7–24 GHz) and FR1(C) (6.75 GHz), which are critical for 6G deployment. Existing datasets for these bands often lack sufficient bandwidth, angular coverage, or environmental diversity.

2. Methodology

The authors propose a comprehensive migration and enhancement of the NYUSIM channel simulator, originally developed by NYU WIRELESS. The methodology involves three core pillars:

A. Migration from MATLAB to Python

Architecture: The entire NYUSIM codebase (v4.0) was ported from MATLAB to a modular, open-source Python framework.
Design Philosophy: The code was reorganized into deterministic classes (fixed physics like path loss, atmospheric attenuation) and stochastic classes (randomness like shadow fading, delay spreads).
Reproducibility: To ensure the Python version is statistically identical to the MATLAB baseline, the authors:
- Implemented the same MT19937 pseudo-random number generator (replacing NumPy's default PCG64) to ensure identical random sequences.
- Matched all mathematical formulas, constants, and array indexing conventions (handling MATLAB's 1-based vs. Python's 0-based indexing).
- Applied a unified randomness control mechanism.

B. Introduction of Ant3D and AntPat Modules

Ant3D Format: A new standardized, full-sphere data format was developed to define 3D antenna patterns. It stores gain matrices over azimuth ( $\phi$ ) and elevation ( $\theta$ ) angles, supporting canonical, commercial, or measured patterns.
AntPat Module: A reconstruction pipeline allows users to input vendor-provided 2D cuts (Vertical and Horizontal) and catalog peak gains. The module synthesizes a full 3D radiation pattern using a multiplication method in the decibel scale:
$G(\theta, \phi) \approx G_H(\phi) + G_V(\theta) - G_{max}$
This enables the simulation of realistic beamforming, sidelobes, and polarization mismatches.

C. Expansion to FR3 and FR1(C) Frequencies

The framework incorporates new statistical models derived from the world's first comprehensive measurement campaigns at 6.75 GHz (FR1(C)) and 16.95 GHz (FR3) in New York City.
These models utilize a Time Cluster Spatial Lobe (TCSL) framework, capturing spatio-temporal channel behavior including path loss, delay spreads, and angular spreads.

3. Key Contributions

NYUSIM Python Framework: A fully verified, open-source Python implementation of the NYUSIM simulator that preserves the physical rigor of the original MATLAB version while enabling seamless integration with AI/ML ecosystems.
Ant3D Standard: A novel, standardized 3D antenna data format that supports full-sphere gain matrices, facilitating the use of realistic commercial and measured antenna patterns in statistical channel modeling.
FR3/FR1(C) Channel Models: The integration of extensive field measurement data for the 6.75 GHz and 16.95 GHz bands, filling a critical gap in 6G channel modeling literature.
Rigorous Validation: The migration was validated through:
- Function-to-Function Testing: Element-wise comparison of deterministic outputs.
- Statistical Validation: Kolmogorov–Smirnov (K–S) tests and moment analysis on over 10,000 Monte Carlo realizations, confirming that the Python version reproduces MATLAB statistics with <1% variance in means and variances.

4. Results

Statistical Fidelity: The Python implementation successfully reproduced the spatio-temporal statistics of the MATLAB v4.0 version across all standardized environments (Urban Micro/Macro, Rural Macro, Indoor Factory/Hotspot) and frequencies (28–142 GHz).
FR3 Validation:
- Simulations at 16.95 GHz and 6.75 GHz were compared against empirical measurements.
- Directional RMS Delay Spread (DS): The cumulative distribution functions (CDFs) of simulated vs. measured directional RMS DS showed strong alignment.
- Quantitative Metrics: Table I in the paper shows that the mean ( $\mu$ ) and standard deviation ( $\sigma$ ) of the log-scaled RMS DS for both LOS and NLOS conditions in the simulation closely matched field measurements (e.g., at 16.95 GHz LOS, Measured $\mu$ =1.23 vs. Simulated $\mu$ =1.30).
Scalability: The Python architecture enables parallel processing via tools like multiprocessing, Ray, or Dask, allowing for the generation of massive datasets required for AI training without licensing costs.

5. Significance

AI-Native Channel Modeling: By moving to Python, NYUSIM becomes a "measurement-driven learning pipeline." It allows researchers to train generative AI models (GANs, Diffusion models, VAEs) directly on channel data and integrate discriminative models for parameter prediction, bridging the gap between physical measurements and AI synthesis.
6G Readiness: The inclusion of FR3 (7–24 GHz) models and 3D antenna capabilities addresses the specific needs of 6G systems, which will rely heavily on these mid-bands and complex antenna arrays.
Reproducibility and Transparency: The open-source nature of the Python framework promotes transparency in propagation research, allowing the community to validate, extend, and learn from the models, moving away from "black box" proprietary simulators.
Foundation for Digital Twins: The ability to integrate with network simulators (like ns-3) and AI workflows makes NYUSIM a critical component for developing 6G digital twins and self-evolving communication systems.

In conclusion, this paper establishes NYUSIM Python as a robust, verified, and extensible foundation for the next generation of AI-enabled wireless research, effectively transitioning channel modeling from static statistical tools to dynamic, data-driven platforms.