Overcoming the Phase Discontinuity Problem in RIS Optimization: A Robust and Scalable Phase‐Aware Deep Regression Framework via Hybrid CNN–LSTM Architecture

Reconfigurable intelligent surfaces (RIS) are crucial for 6G connectivity. However, data‐driven RIS optimizers suffer from the phase discontinuity problem, where estimation collapses near boundaries due to topological mismatches between the circular phase manifold and Euclidean output spaces. To overcome this, we propose a robust, phase‐aware hybrid CNN–LSTM framework that maps the cyclic phase‐space onto continuous Euclidean coordinates via orthogonal sine–cosine projections. Supported by formal embedding proofs, the framework is evaluated on 40,000 channel realizations. It achieves a root‐mean‐square error, outperforming established benchmarks by 21%. It yields a 0.08 ms inference latency ( faster than particle swarm optimization). With linear complexity, it maintains a 0.64 ms delay for elements, meeting URLLC requirements. Ablation studies isolate 29.9% and 20.9% accuracy gains from sine–cosine encoding and LSTM spatial recurrence, respectively. Robustness tests confirm a sustained 15% margin over classical solvers under 10% channel estimation impairment. This scalable architecture offers a superior alternative to quadratic‐complexity () Transformer models, representing one of the first principled deployments of sine–cosine encoding for RIS optimization.