Distributed Electromagnetic Neural Networks for Task-Oriented Semantic Communications

Abstract: Semantic communications (SemCom) is a promising paradigm that prioritizes the transmission of task-relevant information, thereby enabling superior communication efficiency over traditional bit-centric systems. However, most existing SemCom systems face critical limitations in computational efficiency and spatial flexibility. To overcome these limitations, we propose a novel unmanned aerial vehicles (UAV)-enabled distributed electromagnetic neural network (EMNN) for a task-oriented SemCom system. Specifically, the proposed distributed EMNN is composed of multiple UAV-mounted stacked intelligent metasurfaces (SIM) and a ground receiving station (GRS), where multiple SIMs collaboratively encode image semantics in the wave domain, and the GRS performs decoding based on the received power distribution. Moreover, we employ a temperature-adaptive gradient optimization algorithm to train the distributed EMNN, which mitigates gradient vanishing and enhances learning stability. Finally, the numerical simulation results demonstrate the effectiveness of distributed EMNN in image recognition task-oriented SemCom, achieving an average $8\%$ accuracy improvement over the single-SIM baseline across multiple datasets.

• The paper introduces a distributed EMNN architecture using UAV-mounted SIMs for analog semantic encoding and robust task-oriented communications.
• It employs over-the-air computation to fuse spatially diverse semantic features, outperforming centralized baselines by 8–10% in accuracy.
• A novel Temperature-Adaptive Gradient Optimization method ensures end-to-end training of meta-atom coefficients under dynamic channel conditions.

Distributed Electromagnetic Neural Networks for Task-Oriented Semantic Communications: An In-Depth Analysis

Introduction and Motivation

This work introduces a task-oriented semantic communications (SemCom) architecture leveraging distributed electromagnetic neural networks (EMNNs) realized through unmanned aerial vehicles (UAVs) equipped with stacked intelligent metasurfaces (SIMs). Traditional SemCom systems predominantly utilize digital neural networks for semantic encoding and decoding, which imposes high computational loads, elevated energy consumption, and introduces latency—factors that are critical constraints for UAV and edge computing scenarios. The approach in this study addresses these limitations by shifting computation into the analog electromagnetic (EM) domain, employing programmable wavefront manipulation on the SIMs for semantic feature extraction and encoding. The system's distribution across multiple UAVs introduces spatial diversity, supporting robustness against channel fading and enabling task-relevant, efficient communications.

System Architecture and Methodology

The proposed system comprises multiple UAV-mounted SIMs that act as source and semantic encoders, with a ground receiving station (GRS) handling semantic decoding. Each SIM is structured as a multilayer uniform planar array (UPA) of meta-atoms, where each meta-atom’s amplitude and phase can be controlled independently. This physical architecture allows the EMNN to carry out wave-based analog computations, exploiting parallel, low-latency processing inherent to the physics of wave propagation. The UAV-SIM configurations support collaborative extraction and fusion of semantic features through spatially distributed over-the-air computation (AirComp).

Key architectural features:

• Image-to-Wave Mapping: The first metasurface layer in each SIM encodes image data onto the amplitude-phase profiles of meta-atoms (for grayscale, the phase represents pixel value; for color, the luminance and chrominance components map to real and imaginary parts).
• Wave-Domain Semantic Processing: Subsequent metasurface layers perform semantic feature refinement via programmable modulation, emulating deep neural network layers in the analog domain.
• Spatial Diversity and Fusion: Multiple SIMs (each on a UAV) process the same semantic input from different vantage points and transmit signals to the GRS. AirComp at the receiver aggregates these spatially distributed semantic features, mitigating localized deep fades and enhancing resilience.

The wireless channel between each SIM and the GRS is modeled as Rician fading. The GRS, with an antenna array, decodes the semantic class using either pre-detection (waveform-level fusion) or post-detection (power-level fusion) combining strategies, with each antenna corresponding to a semantic category.

Training Paradigm: Temperature-Adaptive Gradient Optimization

Analog EMNNs, operating in the wave domain, lack digital normalization, and channels often introduce significant attenuation, compressing the effective dynamic range and impeding straightforward gradient propagation. To address this, the paper introduces a Temperature-Adaptive Gradient Optimization (TAGO) algorithm:

• Trainable Temperature Parameter ($T$): Incorporated into the softmax activation to adjust amplification of class scores, compensating for non-ideal normalization and improving gradient flow.
• End-to-End Differentiable Training: All SIM meta-atom coefficients (amplitude $\alpha$ and phase $\theta$) and $T$ are learned via backpropagation with Adam optimization and step-based learning rate decay.
• Decentralized Gradient Updates: During training, the GRS computes gradients and distributes gradient signals to respective UAVs for local EMNN parameter updates. This reflects practical constraints where communication resources between devices are limited.

Experimental Results and Key Numerical Findings

Simulations employed the MNIST and CIFAR-10 datasets, with system configurations tuned to emulate realistic THz-band UAV communications (carrier at 0.3 THz, $\lambda=1$ mm, UAV-GRS distances of 100 m, 4 UAVs/SIMs, and 4-layer SIMs with 121 meta-atoms each).

Strong empirical findings:

• Distributed EMNN outperforms centralized baselines: Averaged over multiple datasets, the distributed architecture achieves at least 8–10% greater classification accuracy compared to a single centralized SIM-based EMNN with an equivalent parameter count.
• Superior spatial diversity gains: Increasing the number of UAV-mounted SIMs consistently improves robustness and classification accuracy due to enhanced spatial diversity and AirComp-enabled fusion.
• Optimal temperature regulation: There exists a non-trivial optimum for the temperature parameter $T$; too small $T$ yields brittle convergence and sharp, unstable softmax activations, while large $T$ dilutes feature discrimination. The learned $T$ via TAGO delivers consistently superior results compared to fixed-temperature baselines.
• Across all datasets, distributed EMNNs attain highest accuracy: For MNIST, classification accuracy peaked at 93.06%, while CIFAR-10 (four-class subset) reached 71.77%; the gap to standard DNNs and centralized EMNNs is maintained even under parameter-matched conditions, demonstrating that the performance advantage is due to distributed processing and not just parameter scaling.

Practical and Theoretical Implications

Practical Implications

The architecture’s reduction of digital computation in favor of analog, wave-domain inference drastically cuts latency and energy consumption—central priorities for UAVs and edge devices. The spatially distributed design increases system reliability under channel impairments and node failures, and is highly suitable for deployment in dynamic or disaster-prone environments where connectivity is intermittent or infrastructure is compromised.

Moreover, by realizing semantic encoding and feature extraction at the physical layer, the system avoids excessive front-haul bandwidth usage, facilitating direct task-oriented communication without transmitting superfluous bit representations.

Theoretical Implications and Future Directions

This paradigm demonstrates the feasibility of large-scale, distributed, analog-computation-based neural inference for semantic communications—an approach fundamentally distinct from standard digital SemCom. The physics-driven operation of SIM/EMNN relaxes many bottlenecks of digital hardware, such as analog-to-digital conversion, quantization noise, and clocking constraints.

Key directions and challenges for further research include:

• Scalable Energy Management: As larger SIM arrays increase payload and control energy requirements, joint optimization of energy harvesting, UAV flight dynamics, and metasurface control is crucial [ameur2025EnergyManagement].
• Mobility and Phase Adaptation: Real-time adaptation to mobility-induced Doppler and phase shifts, ensuring robust AirComp under non-static node distributions.
• Advanced Channel Architectures: Integrating more sophisticated channel models (e.g., non-line-of-sight, severe scattering) and joint learning of channel and semantic processing.
• Wave-Domain Machine Learning Theory: Extending information-theoretic analyses for analog, EMNN-based inference and quantifying the fundamental limits of semantic capacity in such hybrid analog/digital networks.

Conclusion

This paper presents a UAV-enabled, distributed electromagnetic neural network framework for task-oriented semantic communications, operating entirely in the analog wave domain using programmable metasurfaces. The combined effect of spatial diversity, AirComp fusion, and temperature-adapted training delivers marked improvements in task accuracy and robustness over centralized and digital baselines. This architecture opens new perspectives for scalable, low-latency, and energy-efficient semantic communications, especially pertinent in wireless edge and autonomous networks. Future work should explore integration with energy harvesting, robust mobility handling, and the theoretical underpinnings of analog semantic processing architectures.