Planar Diffractive Neural Networks Empowered Communications: A Spatial Modulation Scheme

Xiaokun Teng, Yanqing Ren, Weicong Chen, Wankai Tang, Xiaoli Li, Shi Jin

Abstract

Diffractive neural networks, where signal processing is embedded into wave propagation, promise light-speed and energy-efficient computation. However, existing three-dimensional structures, such as stacked intelligent metasurfaces (SIMs), face critical challenges in implementation and integration. In contrast, this work pioneers planar diffractive neural networks (PDNNs) empowered communications, a novel architecture that performs signal processing as signals propagate through artificially designed planar circuits. To demonstrate the capability of PDNN, we propose a PDNN-based space-shift-keying (PDNN-SSK) communication system with a single radio-frequency (RF) chain and a maximum power detector. In this system, PDNNs are deployed at both the transmitter and receiver to jointly execute modulation, beamforming, and detection. We conduct theoretical analyses to provide the maximization condition of correct detection probability and derive the closed-form expression of the symbol error rate (SER) for the proposed system. To approach these theoretical benchmarks, the phase shift parameters of PDNNs are optimized using a surrogate model-based training approach, which effectively navigates the high-dimensional, non-convex optimization landscape. Extensive simulations verify the theoretical analysis framework and uncover fundamental design principles for the PDNN architecture, highlighting its potential to revolutionize RF front-ends by replacing conventional digital baseband modules with this integrable RF computing platform.

Introduction

As the telecommunications industry pivots toward 6G, the requirement for terabit-per-second data rates and massive antenna arrays is straining traditional wireless architectures. Conventional tasks such as modulation, beamforming, and detection are typically handled by digital signal processors (DSPs) and high-resolution data converters. However, these components consume significant power, creating a bottleneck for future high-speed, energy-efficient networks.

Recent research has explored "computing-by-propagation," utilizing 3D Stacked Intelligent Metasurfaces (SIMs) to process signals directly in the electromagnetic (EM) wave domain. While promising, 3D SIMs are bulky, sensitive to physical misalignment, and difficult to integrate into the planar form factors of modern user equipment. This paper introduces Planar Diffractive Neural Networks (PDNNs) as a 2D alternative that confines wave propagation and processing within standard printed circuit boards (PCBs), offering a scalable and integrable solution for RF computing.

PDNN-SSK System Architecture

The proposed system utilizes a PDNN-based space-shift-keying (PDNN-SSK) architecture. Unlike traditional systems requiring multiple RF chains for spatial multiplexing, this system employs a single RF chain.

Hardware Implementation

The PDNN architecture replaces 3D free-space diffraction with a 2D network of microstrip transmission lines. The core components include:

  • Planar Diffractive Coupling: Implemented via branch-line couplers that allow signals to interact across different paths.
  • Phase Shifting: Controlled via delay-line phase shifters to steer and shape the wavefront.
  • Nonlinear Activation (Optional): Power detectors can provide nonlinearities if required for specific processing tasks.
The PDNN architecture showing the layered structure (a) and the detailed circuit implementation (b) including coupling, phase shifting, and detection components.

Signal Processing Flow

  1. Transmitter: Information is encoded using Space-Shift Keying (SSK), where the input port selection represents the transmitted symbol. The transmitting PDNN passively shapes the signal for beamforming through the wireless channel.
  2. Receiver: The receiving PDNN acts as an analog RF demodulator, focusing incoming energy onto a specific output port. A non-coherent power detector identifies the port with the maximum energy to recover the symbol, bypassing complex digital demodulation.
Detailed view of the Branch-Line Coupler and Delay-Line Phase Shifter components used within the PDNN layers.

Optimization via Surrogate Modeling

Configuring the phase shifters in a PDNN is a high-dimensional, non-convex optimization problem. Traditional mathematical methods like projected gradient ascent often fail due to local minima. To address this, the authors developed a differentiable surrogate model of the physical RF system. This allows the use of machine learning optimization techniques, specifically the Adam optimizer with backpropagation, to tune phase shifts and maximize the effective sum-rate.

Key Insights and Design Principles

  • Spatial Scaling Efficiency: In PDNN-SSK, increasing the modulation order $M$ (number of ports) improves energy efficiency per bit. Unlike QAM, where higher orders crowd the signal space and require more power, spatial dimensions allow for higher throughput without a severe SNR penalty.
  • Depth vs. Width: Research indicates a fundamental trade-off: wider networks (more ports) require deeper layers and stronger internal coupling to effectively route signals. Shallow networks lack the physical connectivity to utilize a large number of ports.
  • Planar Advantages: Confining waves to 2D transmission lines significantly reduces propagation losses compared to 3D diffractive networks, enabling deeper and more complex computing structures.

Performance Evaluation

Simulations using Rayleigh-fading channels demonstrate the efficacy of the PDNN-SSK system:

  • Energy Efficiency: To achieve a Symbol Error Rate (SER) of $10^{-3}$, a system with $M=64$ requires an SNR ($E_b/N_0$) of approximately 5 dB, whereas $M=4$ requires 8.5 dB, proving excellent scalability.
  • Detector Performance: The non-coherent power detector performs remarkably well, showing only a 1-3 dB SNR penalty compared to the computationally intensive Maximum Likelihood (ML) detector.
  • Optimization Accuracy: The surrogate model training approach yielded SER curves that closely match the theoretical interference-free lower bounds, outperforming traditional gradient-based methods which often hit an error floor.

Conclusion

PDNNs represent a significant step toward integrating RF computing into standard wireless hardware. By performing complex signal processing at the speed of light within a PCB, PDNNs offer a path to ultra-low-latency and energy-efficient 6G front-ends. Future work will explore discrete phase shift resolutions, multi-user MIMO scenarios, and broadband applications to further mature this technology for real-world deployment.

Summary updated

Abstract

Diffractive neural networks, where signal processing is embedded into wave propagation, promise light-speed and energy-efficient computation. However, existing three-dimensional structures, such as stacked intelligent metasurfaces (SIMs), face critical challenges in implementation and integration. In contrast, this work pioneers planar diffractive neural networks (PDNNs) empowered communications, a novel architecture that performs signal processing as signals propagate through artificially designed planar circuits. To demonstrate the capability of PDNN, we propose a PDNN-based space-shift-keying (PDNN-SSK) communication system with a single radio-frequency (RF) chain and a maximum power detector. In this system, PDNNs are deployed at both the transmitter and receiver to jointly execute modulation, beamforming, and detection. We conduct theoretical analyses to provide the maximization condition of correct detection probability and derive the closed-form expression of the symbol error rate (SER) for the proposed system. To approach these theoretical benchmarks, the phase shift parameters of PDNNs are optimized using a surrogate model-based training approach, which effectively navigates the high-dimensional, non-convex optimization landscape. Extensive simulations verify the theoretical analysis framework and uncover fundamental design principles for the PDNN architecture, highlighting its potential to revolutionize RF front-ends by replacing conventional digital baseband modules with this integrable RF computing platform.
BibTeX
@misc{teng2025planar,
  author = {Xiaokun Teng and Yanqing Ren and Weicong Chen and Wankai Tang and Xiaoli Li and Shi Jin},
  title = {{Planar Diffractive Neural Networks Empowered Communications: A Spatial Modulation Scheme}},
  year = {2025},
  eprint = {2512.00847},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.00847},
}
Figure