Deep Learning-Driven MA-RIS Co-Design for Near-Field ISAC: Joint Antenna Positioning, Phase Shifting and Beamforming Optimization

Movable antennas (MA) and reconfigurable intelligent surfaces (RIS) are recognized as the pivotal technologies for integrated sensing and communication (ISAC) system. Although MAs overcome the gain limitations of fixed-position antennas via exploiting spatial degrees of freedom, RIS enhances signal propagation through programmable phase manipulation. However, existing ISAC research predominantly relies on far-field channel models and conventional optimization methods, which suffer from suboptimal performance in high-frequency near-field scenarios and exhibit high computational complexity when handling coupled continuous-discrete variables. To address these limitations, this paper proposes a novel MA-enabled RIS (MA-RIS) architecture under near-field spherical wave propagation conditions. The proposed system jointly optimizes the weighted sum rate of communication users while ensuring sensing performance through Cramér-Rao bound constraints. Unlike traditional alternating optimization frameworks, we develop a deep learning-based solution including four specialized neural modules: 1) a feature extractor for channel state information, 2) an antenna position vector optimizer, 3) a beamforming matrix optimizer, and 4) a RIS phase matrix optimizer. This integrated architecture enables joint optimization of spatial antenna configurations, beamforming parameters, and RIS phase shifts while accounting for near-field distance-dependent channel characteristics. The simulation results reveal that the proposed scheme achieves significant performance enhancements of 5%∼10% compared to conventional fixed-position RIS configurations, while maintaining comparable effectiveness with the benchmark alternating optimization (AO) approach, which demonstrates our method’s superior capability in balancing system adaptability with optimization efficiency.