Multi-Frequency Channel Modeling for Millimeter Wave and THz Wireless Communication via Generative Adversarial Networks

TL;DR

This paper introduces a neural network-based generative adversarial network (GAN) approach to model joint multi-frequency millimeter wave and Terahertz wireless channels, capturing complex statistical relationships for improved simulation.

Contribution

It presents a novel GAN-based methodology for multi-frequency channel modeling that captures joint distributions with minimal assumptions, applicable to mmWave and THz systems.

Findings

Model accurately captures joint statistical relationships across frequencies.

Method demonstrates effectiveness on urban micro-cellular links at 28 and 140 GHz.

Generates realistic multi-cluster channel profiles for simulation.

Abstract

Modern cellular systems rely increasingly on simultaneous communication in multiple discontinuous bands for macro-diversity and increased bandwidth. Multi-frequency communication is particularly crucial in the millimeter wave (mmWave) and Terahertz (THz) frequencies, as these bands are often coupled with lower frequencies for robustness. Evaluation of these systems requires statistical models that can capture the joint distribution of the channel paths across multiple frequencies. This paper presents a general neural network based methodology for training multi-frequency double directional statistical channel models. In the proposed approach, each is described as a multi-clustered set, and a generative adversarial network (GAN) is trained to generate random multi-cluster profiles where the generated cluster data includes the angles and delay of the clusters along with the vectors of random received powers, angular, and delay spread at different frequencies. The model can be readily applied for multi-frequency link or network layer simulation. The methodology is demonstrated on modeling urban micro-cellular links at 28 and 140 GHz trained from extensive ray tracing data. The methodology makes minimal statistical assumptions and experiments show the model can capture interesting statistical relationships between frequencies.