AI Boosts Wireless Data Modelling to 142GHz

Accurate wireless channel modelling is crucial for the development and deployment of future 6G communication systems. Isha Jariwala, Xinquan Wang, and Bridget Meier, along with colleagues from New York University, Tandon School of Engineering, present NYUSIM, a significantly enhanced statistical channel modelling and simulation framework. This work builds upon the established NYUSIM platform, migrating it from MATLAB to Python and incorporating new statistical models generated from recent field measurements at 6.75GHz and 16.95GHz. The researchers rigorously validated the Python implementation against the original MATLAB version, ensuring statistical consistency and reproducibility. Furthermore, NYUSIM now supports a standardised 3D antenna data format, Ant3D, and is designed for integration with modern artificial intelligence workflows, providing a robust and extensible foundation for AI-enabled channel modelling and large-scale data generation.

Accurate wireless communication relies on detailed understanding of signal behaviour in the real world. A new framework, NYUSIM, offers a powerful tool for creating realistic digital twins of wireless environments. This advance promises to accelerate the development of next-generation mobile networks and artificial intelligence applications. Scientists are reshaping wireless channel modelling with a new, artificially intelligent framework designed to generate remarkably realistic data for future 6G networks.

This work addresses a critical need for large, accurate datasets to train and validate models that can predict how radio signals behave in diverse and changing environments. Building upon the established NYUSIM simulator, originally introduced in 2016, researchers have transitioned the entire system from MATLAB to Python, significantly improving its scalability and adaptability for advanced artificial intelligence workflows.

The updated platform now incorporates extensive measurements from the 6.75GHz (FR1(C)) and 16.95GHz (FR3) frequency bands, key spectrum allocations for upcoming 6G technologies. A novel three-dimensional antenna data format, termed Ant3D, has been developed to accurately represent complex antenna patterns, moving beyond simplified models and enabling more precise simulations of signal propagation.

This standardized format allows for the incorporation of both commercially available and custom-measured antenna characteristics into the channel modelling process. The migration to Python underwent rigorous validation using Kolmogorov-Smirnov (K-S) tests and moment analysis, ensuring the new version faithfully reproduces the statistical properties of the original MATLAB implementation.

At its core, this research establishes a verified and extensible foundation for AI-enabled channel modelling. Once limited by closed-source code and scalability issues, channel simulation is now poised to benefit from the power of generative and discriminative AI. By leveraging machine learning techniques, the system can learn from real-world measurements and synthesize realistic channel environments, opening doors to innovations in areas like 6G digital twins and AI-native physical layer design.

The team anticipates this work will support the development of more efficient and reliable wireless communication systems for years to come. With over 100,000 downloads and citations in more than 3,300 publications as of 2024, the original NYUSIM has already become a cornerstone of wireless research. This updated Python version builds upon that legacy, offering a modern, open-source platform for researchers worldwide. By seamlessly integrating with modern AI workflows and enabling large-scale parallel data generation, the new NYUSIM promises to accelerate the development of the next generation of wireless technologies.

Statistical Validation of Python NYUSIM Implementation Against MATLAB Equivalent

Migration of the NYUSIM framework from MATLAB to Python has been rigorously validated, demonstrating statistical consistency with the original implementation. Kolmogorov-Smirnov (K-S) tests and moment analysis confirmed that the Python version accurately reproduces spatio-temporal channel statistics. End-to-end testing, utilising unified randomness control, further verified this consistency alongside the open-source MATLAB NYUSIM v4.0.

These tests assessed spatial consistency, a key requirement for accurate channel modelling. The Python implementation matched the statistical behaviour of the MATLAB version, ensuring reliable data generation for future research. The system employs the MT19937 random number generator, mirroring the MATLAB setup with a 53-bit uniform mapping and consistent handling of various random distributions.

Detailed comparisons of formulas, array shapes, and indexing methods were undertaken to minimise discrepancies. Executing both versions on identical inputs revealed comparable runtime and memory usage for single runs. However, Python’s scalability allows for parallel simulations across multiple machines using tools like multiprocessing, Ray, or Dask, eliminating licensing costs associated with MATLAB and enabling the generation of the large datasets needed for artificial intelligence training.

NYUSIM Python incorporates a 3D antenna data format, termed Ant3D, providing a standardised way to define antenna patterns. Each antenna pattern is defined on a spherical grid, covering azimuth angles from 0 to 360 degrees and elevation angles from -90 to 90 degrees. The Ant3D format stores antenna gain values in dBi within a matrix G(θ, φ), accommodating isotropic, horn, dipole, and phased-array antennas.

All antenna data is normalised to the maximum gain, ensuring consistency across frequencies and scenarios. Data blocks within the Ant3D structure include frequency, angular grids, gain matrix, polarisation, and normalisation reference, preserving directional gain information. Inside the AntPat module, vendor antenna patterns can be imported and reconstructed.

The module reconstructs full 3D radiation patterns from principal cuts, a vertical cut GV(θ) at φ= 0° and a horizontal cut GH(φ) at θ= 0°, along with catalogue peak gain and half-power beamwidth (HPBW) values. Using the multiplication method, the 3D gain in decibels is approximated as GH(φ) + GV(θ) − Gmax, where Gmax is the catalogue peak gain. This reconstruction procedure ensures fidelity to the original vendor measurements, allowing users to simulate channels with realistic antenna characteristics.

NYUSIM porting to Python and statistical validation for AI-driven channel modelling

A migration of the NYUSIM wireless channel simulator from MATLAB to Python underpins this work, driven by the need for scalable datasets to support artificial intelligence research. NYUSIM generates realistic, spatio-temporal channel data using measurements taken between 28 and 142GHz, and the Python version incorporates new statistical model generation capabilities for the 6.75GHz and 16.95GHz frequency bands.

This transition was undertaken to improve scalability and integrate with modern AI workflows, allowing for large-scale parallel data generation. The choice of Python allows for easier integration with machine learning tools and avoids licensing costs associated with MATLAB when creating the extensive datasets required for training AI models. Rigorous validation confirmed the statistical consistency between the MATLAB and Python versions of NYUSIM.

Function-to-function testing compared deterministic functions element-wise, ensuring numerical equivalence, while stochastic functions underwent statistical validation using over 10,000 Monte Carlo (MC) realizations. The Kolmogorov-Smirnov (K-S) test and moment-based analysis verified that distributions of key channel characteristics, delay spread, angular spread, and shadow fading, were statistically indistinguishable between the two implementations.

These tests were performed across a range of environments, including urban microcell, rural macrocell, and various indoor settings, under both line-of-sight and non-line-of-sight conditions. Achieving statistical equivalence required careful attention to random number generation. MATLAB employs MT19937, a deterministic pseudo-random generator, while NumPy uses PCG64.

To address this, the same MT19937 generator was implemented in both versions, alongside an identical 53-bit uniform mapping, ensuring all random draws followed consistent procedures. Beyond random number control, formulas, constants, array shapes, and indexing conventions were meticulously matched to account for differences between the platforms. Furthermore, the inclusion of a standardised 3D antenna data format, Ant3D, promises to streamline integration with diverse ray tracing tools and modelling approaches. By providing a consistent structure for antenna representation, Ant3D enhances the accuracy and versatility of spatial channel simulations.

Python-based channel emulation facilitates scalable wireless AI and 6G development

Scientists have delivered a vital upgrade to a foundational tool for wireless research, migrating the NYUSIM channel emulator from MATLAB to Python. This isn’t merely a porting exercise; it’s a strategic realignment to meet the demands of artificial intelligence and the looming arrival of 6G networks. For years, building realistic wireless models has relied on painstakingly collected measurements and complex software, creating a bottleneck for those seeking to train AI algorithms that can predict and optimise signal behaviour.

Accurate channel models are essential, yet generating them at scale has proven exceptionally difficult, demanding considerable computational resources and specialised expertise. Now, with NYUSIM recast in Python, a more accessible and scalable platform emerges. Beyond the code itself, the inclusion of a standardised 3D antenna data format, Ant3D, promises to streamline integration with diverse ray tracing tools and modelling approaches.

This move addresses a long-standing problem of interoperability, where different software packages often struggle to share antenna information effectively. Validation tests confirm the Python version faithfully reproduces the statistical characteristics of the original MATLAB code, but the true benefit lies in its potential to unlock AI-driven channel modelling.

Simply having data isn’t enough. The challenge now shifts to developing AI algorithms that can effectively learn from this data and generalise to unseen environments. Existing methods, while promising, still require substantial refinement to handle the complexities of real-world wireless propagation. Looking ahead, we can anticipate a surge in research exploring generative models, techniques that can synthesise realistic channel data, supplementing and extending the capabilities of NYUSIM. This isn’t the finish line, but a powerful step towards a future where AI proactively shapes the design and operation of wireless networks.