GNN-enabled Precoding for Massive MIMO LEO Satellite Communications

TL;DR

The paper tackles energy-efficient downlink precoding for massive-MIMO LEO satellites under stringent on-board power constraints and rapidly varying channels. It proposes a dual approach: (i) an end-to-end graph neural network (GNN) for low-complexity precoding and (ii) a deep-unfolding framework that combines the Dinkelbach algorithm with WMMSE, augmented by Taylor expansion to approximate matrix inverses. The key contributions are the end-to-end GNN design, a deep-unfolding GNN that preserves algorithmic structure while accelerating convergence, and a Taylor-expansion-based inverse approximation, together yielding dramatic reductions in computational complexity and enhanced robustness. The results show substantial EE improvements and real-time feasibility over state-of-the-art methods, highlighting the practical potential of AI-driven precoding for 6G LEO networks.

Abstract

Low Earth Orbit (LEO) satellite communication is a critical component in the development of sixth generation (6G) networks. The integration of massive multiple-input multiple-output (MIMO) technology is being actively explored to enhance the performance of LEO satellite communications. However, the limited power of LEO satellites poses a significant challenge in improving communication energy efficiency (EE) under constrained power conditions. Artificial intelligence (AI) methods are increasingly recognized as promising solutions for optimizing energy consumption while enhancing system performance, thus enabling more efficient and sustainable communications. This paper proposes approaches to address the challenges associated with precoding in massive MIMO LEO satellite communications. First, we introduce an end-to-end graph neural network (GNN) framework that effectively reduces the computational complexity of traditional precoding methods. Next, we introduce a deep unfolding of the Dinkelbach algorithm and the weighted minimum mean square error (WMMSE) approach to achieve enhanced EE, transforming iterative optimization processes into a structured neural network, thereby improving convergence speed and computational efficiency. Furthermore, we incorporate the Taylor expansion method to approximate matrix inversion within the GNN, enhancing both the interpretability and performance of the proposed method. Numerical experiments demonstrate the validity of our proposed method in terms of complexity and robustness, achieving significant improvements over state-of-the-art methods.

Figures (13)

• Figure 1: This diagram illustrates the architecture of an end-to-end GNN.
• Figure 2: Downlink LEO satellite communication system with $K$ single-antenna UTs and its graph representation.
• Figure 3: Detailed process of the edge $(n,k)$.
• Figure 4: The diagram illustrates the architecture of a deep unfolding model, which appears to be used for iterative optimization or inference tasks.
• Figure 5: Downlink LEO satellite communication system and its graphical representation with five antennas as an example.