Hardware-Conditioned Generative Channel Modeling: A Diffusion-Based Approach for Location and Hardware-Aware Wireless Dataset Synthesis

Nitish Deshpande, Sanjay Ganapathy, Viraj Shah

Keywords: Wireless Networks, Channel Generation, Conditional Diffusion Models

Abstract: The design of next-generation wireless communication systems is increasingly reliant on data-driven machine learning models. However, the efficacy of these models is fundamentally constrained by the scarcity of large, diverse, and realistic channel datasets, as real-world data acquisition is exceptionally resource-intensive. Generative AI, particularly diffusion models, has emerged as a promising solution for synthetic data generation. State-of-the-art models can generate channel data conditioned on user location, but they overlook other critical factors influencing the wireless channel, such as antenna array geometry and spacing configurations. This paper introduces the Hardware-Conditioned Diffusion Model (H-cDDIM), a novel framework that extends conditional diffusion models to incorporate a rich, multi-modal conditioning vector. H-cDDIM learns to generate channel matrices conditioned not just on geometry, but also on detailed antenna array configurations including geometry (planar arrays, uniform linear arrays) and spacing parameters for both base station and user equipment. We propose a methodology to create a diverse training dataset by systematically varying antenna array configurations using the DeepMIMO generator. Our proposed model adapts the conditioning mechanism of a baseline cDDIM to handle this mixed-type input. The resulting H-cDDIM is capable of generating high-fidelity, site-specific channel data for a wide range of antenna configuration scenarios, thereby accelerating the research and deployment of AI-enabled wireless technologies. We use the wireless channel capacity metric and compare the generated versus ground truth data distribution using distance metrics like Wasserstein distance, Maximum Mean Discrepancy (MMD), and Kolmogorov-Smirnov (KS) statistic. Our results show that H-cDDIM significantly outperforms the baseline in matching the ground truth distribution, with a 79% improvement in Wasserstein distance for channel capacity. Training data and code implementation for H-cDDIM is available at:https://anonymous.4open.science/r/h-cddim-FCB3.

Submission Number: 289