Ordinary Differential Equation-based CNN for Channel Extrapolation over RIS-assisted Communication

Meng Xu, Shun Zhang, Caijun Zhong, Jianpeng Ma, Octavia A. Dobre

Introduction

Reconfigurable Intelligent Surfaces (RIS) are a cornerstone technology for 6G wireless networks, offering the ability to manipulate the propagation environment through arrays of passive reflecting elements. However, the potential of RIS is often bottlenecked by the overhead required to acquire accurate Channel State Information (CSI). In a typical RIS-assisted system, the number of elements $L$ can reach hundreds or thousands, making full channel estimation prohibitively expensive in terms of pilot signal consumption.

To address this, researchers have proposed channel extrapolation: measuring only a small subset of RIS elements and using deep learning to infer the full channel map. This paper introduces a novel Convolutional Neural Network (CNN) architecture inspired by the mathematics of Ordinary Differential Equations (ODEs) to achieve high-fidelity extrapolation with minimal overhead.

System Model and Problem Statement

The system consists of a base station (BS) with $M$ antennas, an RIS with $L$ elements, and a single-antenna user. The goal is to estimate the cascaded channel—the link from the BS to the RIS and then to the user—without measuring every element individually.

Figure 1: RIS-assisted communication system model showing the BS, RIS with L elements, and the user.

By sub-sampling the RIS (e.g., activating only 1/4 or 1/8 of the elements), the system obtains a sparse "sketch" of the channel. The challenge lies in accurately "filling in the blanks" to reconstruct the full $L$-element channel matrix, a task similar to image super-resolution.

ODE-Based CNN Architecture

Standard deep networks like ResNets can be mathematically interpreted as a discrete approximation of a first-order ODE using the Euler method. This paper improves upon this by employing higher-order numerical solvers to design the network blocks, specifically the Runge-Kutta (RK3) method.

The RK3-Block

The proposed architecture replaces standard cascaded layers with specialized RK3-Blocks. While a standard residual connection simply adds the input to the output of a layer, the RK3-Block implements a third-order approximation scheme. It uses a three-stage structure where intermediate features are combined using specific linear coefficients (such as $1/2, 2, 4, -1,$ and $1/6$) to minimize truncation error and better capture the latent spatial correlations of the wireless channel.

Figure 2: (a) Overall network architecture; (b) Detailed structure of the proposed RK3-Block; (c) The internal convolution and nonlinear operation G.

By utilizing these mathematically rigorous connections, the network can characterize complex relationships between data layers more effectively than traditional architectures, leading to faster convergence and higher accuracy.

Experimental Results and Insights

The model was evaluated using the DeepMIMO dataset (indoor scenario I1) with a 64-element RIS. Key findings include:

  • Superior Accuracy: The ODE-RK3 network consistently outperformed standard CNNs and LeapFrog-based architectures. At a 1/4 sampling rate, the ODE-RK3 achieved a Normalized Mean Squared Error (NMSE) of -33.77 dB, significantly better than the -26.78 dB achieved by a standard CNN.
  • Efficiency Gains: Remarkably, the ODE-RK3 at a 1/4 sampling rate outperformed a standard CNN at a 1/2 sampling rate. This implies that the ODE-based architecture can cut pilot overhead by half compared to baseline deep learning methods while maintaining better performance.
  • Robustness: The model demonstrated stability across different sampling rates (down to 1/16) and remained effective even when there was a frequency gap between the measured and extrapolated channels, a common scenario in Frequency Division Duplex (FDD) systems.

Conclusion

By bridging the gap between numerical analysis and deep learning, the ODE-based CNN provides a powerful tool for RIS channel extrapolation. This approach significantly reduces the pilot overhead required for 6G systems, making massive RIS deployments more practical. Future work will explore optimizing the sub-sampling patterns and testing the architecture on even larger RIS arrays.

Summary updated

BibTeX
@article{xu2021ordinary,
  author = {Meng Xu and Shun Zhang and Caijun Zhong and Jianpeng Ma and Octavia A. Dobre},
  title = {{Ordinary Differential Equation-based CNN for Channel Extrapolation over RIS-assisted Communication}},
  journal = {IEEE Communications Letters},
  year = {2021},
  doi = {10.1109/lcomm.2021.3064596},
}
Figure