Ordinary Differential Equation-Based CNN for Channel Extrapolation Over RIS-Assisted Communication

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

TL;DR

Estimating the channel for every element in a Reconfigurable Intelligent Surface (RIS) is a massive overhead nightmare. This paper proposes a deep learning approach that only measures a fraction of the RIS elements and uses a novel Convolutional Neural Network (CNN) architecture—inspired by the mathematics of Ordinary Differential Equations (ODEs)—to accurately extrapolate the full channel information, significantly boosting spectral efficiency.

What Problem Does This Paper Address?

Reconfigurable Intelligent Surfaces (RIS) are a "hot" 6G technology consisting of arrays of passive elements that can "steer" wireless signals. However, to use them effectively, the system needs to know the Channel State Information (CSI) for the link between the base station, the RIS, and the user.

The catch? RIS arrays often have hundreds of elements. Measuring the channel for every single one requires a massive amount of "pilot" signals, which eats up the very bandwidth the RIS is supposed to save. Traditional statistical models struggle with the rapidly changing, complex scattering of real-world environments. While standard deep learning (DL) can help, deep networks often suffer from vanishing gradients or slow convergence when trying to map partial channel data to a full, high-resolution channel map.

Methods

The authors propose a Channel Extrapolation framework. Instead of measuring the whole array, they sub-sample the RIS (turning on only a fraction of elements) to create a sparse "sketch" of the channel. They then use a specialized CNN to "fill in the blanks."

The innovation lies in the ODE-based CNN structure. Most modern neural networks (like ResNets) can be viewed as a discrete approximation of a first-order ODE (specifically, the Euler method). The authors argue that we can do better by using more sophisticated numerical solvers:

  • LeapFrog (LF) Method: A second-order approximation scheme that connects the $(n+1)$-th layer with the $(n-1)$-th layer.
  • Runge-Kutta (RK3) Method: A third-order approximation that uses a "3-stage" block structure. It involves complex linear combinations and coefficients between layers to minimize "truncation error"—the math version of making sure the network doesn't lose detail as it gets deeper.

By replacing standard cascaded layers with these RK3-Blocks, the network gains a more powerful way to characterize the latent relationships between different data layers, leading to more accurate extrapolation.

Datasets

The researchers used the DeepMIMO dataset, specifically the indoor scenario I1.

  • Generation: This data is generated using ray-tracing software (Wireless InSite) to simulate a realistic indoor environment with a 4-antenna base station and a 64-element RIS (8x8 array).
  • Scale: They simulated 20,100 users, splitting them into a 4:1 ratio for training and testing.
  • Parameters: The simulation operated at 2.4 GHz with 64 subcarriers (OFDM).

Key Insights and Ideas

  • Math-Informed Architecture: Instead of just adding more layers (which leads to diminishing returns and training difficulties), the authors used numerical analysis (ODEs) to rethink how layers should talk to each other.
  • The "Super-Resolution" Analogy: The paper treats channel extrapolation similarly to image super-resolution. Just as a CNN can guess the missing pixels in a blurry photo by looking at surrounding patterns, this model guesses the missing channel data by exploiting the spatial correlation between nearby RIS elements.
  • Efficiency Gains: A major "aha" moment is that the ODE-RK3 network at a 1/4 sampling rate actually outperforms a standard CNN at a 1/2 sampling rate. This means you can cut your measurement overhead in half again just by switching the network architecture.

Key Results

  • Superior Accuracy: In terms of Normalized Mean Squared Error (NMSE), the ODE-RK3 consistently beat both the LeapFrog version and the standard cascaded CNN across all sampling rates (1/2, 1/4, 1/8, and 1/16).
  • Concrete Numbers: At a 1/4 sampling rate, the ODE-RK3 achieved an NMSE of -33.77 dB, compared to only -26.78 dB for the standard CNN.
  • Faster Convergence: The training loss for the ODE-based networks dropped much faster and reached a lower stable level than the baseline, meaning the models are easier and quicker to train.
  • Frequency Robustness: The model remained stable even when there was a frequency gap between the measured and extrapolated channels (common in FDD systems where uplink and downlink operate on different bands).

Limitations and Future Work

  • Sampling Patterns: The authors used a simple uniform sampling scheme (picking elements at regular intervals). They acknowledge that the pattern of which elements you choose to measure likely has a huge impact on performance, but they left that optimization for future research.
  • Hardware Complexity: While the number of parameters was kept the same for a fair comparison, the internal linear calculations within an RK3-Block are slightly more complex than a standard skip connection, which might impact real-time inference latency on low-power hardware.
  • Scaling Up: The study focused on a 64-element RIS. Testing the limits of this extrapolation on massive surfaces (e.g., 256 or 1024 elements) remains an open area.

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