AI-based CSI Feedback with Digital Twins: Real-World Validation and Insights

TL;DR

This paper validates the use of digital twins for training deep learning models in real-world CSI feedback scenarios, demonstrating the necessity of dedicated DTs despite online learning benefits.

Contribution

It provides the first real-world validation of digital twin-based CSI feedback models and assesses online learning's impact on real-world performance.

Findings

Digital twins can generate realistic site-specific data for training.

Models trained in DTs require dedicated DTs for satisfactory real-world performance.

Online learning enhances model adaptation but does not replace the need for dedicated DTs.

Abstract

Deep learning (DL) has shown great potential for enhancing channel state information (CSI) feedback in multiple-input multiple-output (MIMO) communication systems, a subject currently under study by the 3GPP standards body. Digital twins (DTs) have emerged as an effective means to generate site-specific datasets for training DL-based CSI feedback models. However, most existing studies rely solely on simulations, leaving the effectiveness of DTs in reducing DL training costs yet to be validated through realistic experimental setups. This paper addresses this gap by establishing a real-world (RW) environment and corresponding virtual channels using ray tracing with replicated 3D models and accurate antenna properties. We evaluate whether models trained in DT environments can effectively operate in RW scenarios and quantify the benefits of online learning (OL) for performance enhancement. Results show that a dedicated DT remains essential even with OL to achieve satisfactory performance in RW scenarios.