Predictive Shape Adaptation in RIS-Assisted Flexible Pinching-Antenna Systems for 6G Networks: A Deep Learning and Manifold Optimization Approach

Abstract

Flexible Pinching-Antenna Systems (F-PASS) have recently emerged as a transformative, energy-efficient alternative to conventional rigid massive MIMO arrays. By dynamically adapting its geometric shape, F-PASS can optimize spatial degrees of freedom and multiplexing gains in real-time. However, existing F-PASS architectures operate in a purely reactive mode, where unavoidable mechanical reconfiguration latency fundamentally limits their real-time channel-tracking capabilities in highly dynamic, high-mobility environments. In this paper, we propose a novel proactive F-PASS architecture assisted by Reconfigurable Intelligent Surfaces (RIS) for next-generation 6G wireless networks. By synergistically combining the coarse physical shape adaptation of the F-PASS transceiver with the fine-grained, instantaneous wavefront manipulation of the RIS, we significantly mitigate the physical latency penalty. To completely overcome mechanical delay, we introduce a predictive deep learning framework based on dilated Temporal Convolutional Networks (TCNs) to accurately anticipate rapid spatial-channel variations. This intelligence enables the F-PASS to proactively reconfigure its geometry, ensuring perfect spatial alignment at the predicted channel realization. We rigorously formulate a comprehensive joint optimization framework to maximize global energy efficiency (EE). The resulting mixed-integer non-linear programming (MINLP) problem is solved via a robust alternating optimization (AO) algorithm leveraging Dinkelbach’s fractional programming and Riemannian Conjugate Gradient (RCG) manifold optimization. Extensive numerical simulations, backed by rigorous analytical theorems, demonstrate that the proposed Predictive RIS-FPASS achieves excellent results, specifically delivering a remarkable 43.5% improvement in energy efficiency and a 38.2% increase in spectral efficiency over conventional reactive baselines.