Transformer‐Enhanced Multi‐Satellite IRS Coordination for Next‐Generation LEO Communications

The utilisation of large‐scale low Earth orbit (LEO) satellite constellations to enable sixth‐generation (6G) communications and services is hampered by significant technical challenges due to the need to optimise ground‐based intelligent reflecting surfaces (IRS) for multi‐satellite coordination in the presence of time‐varying orbital motion with velocities of > 7.5 km/s, inter‐satellite interference, and the need to support real‐time joint optimisation. State‐of‐the‐art approaches have thus far followed an alternating optimisation paradigm, wherein separate intermediate parameter spaces are jointly optimised without directly maximising the overall end‐to‐end communication performance, which in turn limits the resulting IRS coordination and leads to prohibitively high computational costs. In this paper, we present a transformer‐based IRS coordination framework for LEO satellite communications that aims to fundamentally address these open issues through a physics‐informed transformer architecture for IRS coordination in LEO that includes a physics‐informed neural network to encode orbital dynamics, multi‐head self‐attention to capture short‐ and long‐term temporal dependencies, and cross‐attention layers to learn inter‐agent coordination; all of which are optimised end‐to‐end in rate. This is accomplished through a physics‐informed loss function that enforces Doppler coherence, geometric consistency, and orbital mechanics constraints, and we showcase via extensive numerical results at Ku‐band (12 GHz) under a range of LEO settings (inclination angle and altitude) that the proposed framework: (i) significantly outperforms alternating optimisation state‐of‐the‐art methods in terms of convergence speed and achievable performance by up to 9.6% in peak throughput (1.95 bps/Hz vs 1.78 bps/Hz) and demonstrates robustness over a wide range of channel conditions (Rician K‐factor 0–20 dB), and (ii) is capable of real‐time operation with computational time under 50 ms. The proposed framework enables the practical application of IRSs in LEO constellations, which is of fundamental importance for future space‐terrestrial integrated 6G networks.