RadioDiff-Turbo: Lightweight Generative Large Electromagnetic Model for Wireless Digital Twin Construction

In this paper, we propose a novel and practical real-time wireless digital twin (WDT) construction approach, which can demonstrate and update the spatial distribution of the wireless channel features in the digital space in real-time whenever the features of the environment or the wireless transmitter are varied. Different from traditional computing costly electromagnetic (EM) computation based methods, such as EM ray tracing, a diffusion-based large generative EM model is proposed for the inference of spatial distribution of the wireless channel features in the digital space. Furthermore, addressing the computational efficiency bottleneck of traditional large generative models, in this paper a series of turbo learning methods are used to lighten the RadioDiff model, which is the previous state-of-the-art (SOTA) EM model. First of all, through the analysis of the properties of the stochastic differential equation (SDE) corresponding to the sampling process, it is revealed that the number of sampling steps can be significantly reduced by more than 100x by adjusting the upper limit of the integral. By removing structure-group parameters, we find that the computational efficiency of each sampling step can be further improved. Experimental results show that the proposed RadioDiff-Turbo method can improve the inference efficiency of RadioDiff by 10x, achieving an inference speed of the order of tens of milliseconds, thus providing a practical construction method for real-time WDT. Remarkably, we find that the lightened RadioDiff-Turbo, without any fine-tuning or post-training from RadioDiff, reaches a new SOTA performance, which provides an insightful direction for the research of wireless large models.