Flexible RISs: Learning-based Array Manifold Estimation and Phase-shift Optimization

1. Problem Statement

Reconfigurable Intelligent Surfaces (RISs) are critical for next-generation wireless networks, offering programmable control over signal propagation. However, most existing research assumes RISs are mounted on planar (flat) surfaces. In practical deployments, RISs are often installed on non-planar structures such as curved building facades, structural columns, or vehicle bodies.

• The Core Challenge: If an RIS is physically curved but its phase shifts are optimized based on a planar assumption, the geometric mismatch causes significant performance degradation (e.g., reduced data rates and sensing accuracy).
• Limitations of Existing Solutions:
  • Analytical solutions for curved RISs exist but are typically restricted to specific, pre-defined geometries (e.g., only cylindrical or only parabolic).
  • Treating the geometry estimation as a non-linear, non-convex optimization problem is computationally prohibitive for real-time implementation.
  • Existing "flexible RIS" literature often assumes element locations are known and controllable, whereas this paper addresses scenarios where the RIS is conformal to a static but unknown surface.

2. Methodology

The authors propose a Deep Learning (DL) framework that estimates the unknown surface geometry and optimizes phase shifts using only sparse received power measurements, avoiding the need for complex Channel State Information (CSI) or phase synchronization.

A. Parametric Surface Modeling

To make the geometry estimation tractable, the authors introduce a low-dimensional parametric model based on a Taylor series approximation.

• Instead of estimating the coordinates of every RIS element ($N$ variables), the surface curvature is modeled as a quadratic function of the $y$ and $z$ coordinates:
  $$x = g(y, z) \approx a_{yy}y^2 + a_{zz}z^2 + a_{yz}yz + a_y y + a_z z$$
• This reduces the estimation problem from $N$ dimensions to just 5 unknown coefficients ($a_{yy}, a_{zz}, a_{yz}, a_y, a_z$).

B. Data Collection Strategy

The system operates in a downlink scenario with a Base Station (BS), a non-planar RIS, and $K$ Mobile Users (MUs).

• Input Data: The system collects received signal power measurements ($P_{k,m}$) at $K$ distinct user locations. For each location, the RIS cycles through $M$ different phase configurations.
• Measurement Type: The method relies exclusively on received power (magnitude squared) rather than phase information. This is chosen because power measurements are robust and do not require complex synchronization hardware, unlike phase measurements.
• Sampling Density: The paper derives a heuristic for the minimum spacing between measurement points to ensure distinct spatial information, based on the decay of the sinc function beam pattern.

C. Neural Network Architecture

A Feed-Forward Multi-Layer Perceptron (MLP) is trained to perform the following mapping:

• Input: A vector of received power measurements corresponding to various user locations and RIS configurations.
• Output: The 5 geometric coefficients ($a_{yy}, a_{zz}, a_{yz}, a_y, a_z$) that define the surface curvature.
• Training: The network is trained using a dataset generated via simulation (Non-Linear Least Squares optimization) to minimize the Mean Squared Error (MSE) between predicted and actual coefficients.
• Inference: Once trained, the NN performs real-time inference. Once the coefficients are estimated, the "known" geometry allows for the analytical calculation of optimal phase shifts for any target user location.

3. Key Contributions

1. General Parametric Model: Introduction of a Taylor-series-based model that captures arbitrary surface curvatures (cylindrical, parabolic, etc.) while significantly reducing the dimensionality of the estimation problem.
2. DL-Based Geometry Estimation: A novel neural network approach that implicitly learns the mapping from sparse received power measurements to surface geometry parameters, bypassing the need for iterative numerical optimization during runtime.
3. Robustness to Unknown Geometry: The framework enables real-time phase-shift optimization for RISs on static, unknown non-planar surfaces, a scenario not previously addressed in literature using NNs.
4. Performance Analysis: Comprehensive evaluation of the algorithm's convergence, performance gain over planar designs, and robustness against user location estimation errors.

4. Simulation Results

The authors evaluated the proposed method using a 28 GHz mmWave system with a $40 \times 10$ RIS element array.

• Convergence: The NN training converged smoothly, with both training and validation losses decreasing steadily over 500 epochs.
• Performance vs. Curvature Variance ($\sigma$):
  • Low Variance ($\sigma \leq 0.25$): The conventional planar design performs slightly better or comparably, as the surface is nearly flat.
  • High Variance ($\sigma > 0.25$): The planar design degrades significantly. The proposed DL-based method maintains high performance, demonstrating a substantial gain in Signal-to-Noise Ratio (SNR) over the planar assumption.
  • Generalization: The algorithm showed robustness even when tested on curvature variances outside the training distribution (unobserved data range), though with a gradual performance drop.
• Robustness to Location Error:
  • The proposed method outperforms the planar design even when the Mobile User's location has estimation errors ($\epsilon$).
  • Interestingly, the planar design is less sensitive to location errors because its beam is already misaligned due to the geometric mismatch; thus, further location errors do not drastically change the already poor received power.

5. Significance and Conclusion

This paper addresses a critical gap between theoretical RIS models and practical deployment realities. By shifting from rigid planar assumptions to a learning-based approach for flexible, non-planar RISs, the authors demonstrate that:

• Accurate beamforming is possible even when the physical surface geometry is unknown.
• Deep learning can effectively replace computationally expensive iterative optimization for real-time RIS configuration.
• The proposed method significantly enhances system reliability and data rates in scenarios involving curved architectural elements, paving the way for the integration of RISs into complex urban and vehicular environments.

The work validates that a data-driven approach using sparse power measurements is a viable and robust solution for the "array manifold estimation" problem in next-generation wireless networks.