Physics-Informed Graph Neural Networks for the Inverse Design of GHz Reconfigurable Antenna

Cindy Pan, Naveen Verma, James SturmIEEE European Conference on Antenna and Propagation 2025 · OpenReview

TL;DR

This paper introduces a novel physics-informed Graph Neural Network (GNN) that can predict the precise switch configurations needed to achieve a desired radiation pattern for reconfigurable antennas. By incorporating fundamental physics like Kirchhoff's Current Law into the GNN's architecture, the model learns the complex relationship between antenna geometry, switch states, and resulting electromagnetic behavior, paving the way for more efficient and dynamic antenna design.

What Problem Does This Paper Address?

Designing reconfigurable antennas, which can dynamically alter their radiation patterns, is crucial for advanced wireless communication systems. The challenge lies in the "inverse design" problem: given a desired radiation pattern, how do we determine the specific settings (in this case, the states of radio-frequency switches) of the antenna to achieve it? Traditional methods rely on computationally expensive iterative simulations and manual optimization. While machine learning offers a promising alternative, the intricate spatial relationships and physical interactions within an antenna array make it difficult for standard neural networks to effectively learn these complex mappings. This paper tackles this gap by proposing a GNN that can directly predict switch configurations from desired surface current distributions, which are directly linked to radiation patterns.

Methods

The core of this work is a physics-informed Graph Neural Network (GNN). The antenna array is represented as a graph where antenna patches are nodes and the RF switches connecting them are edges.

  • Node Features: The surface current density within each antenna patch is averaged to create a single complex value, serving as the node feature. This captures the electromagnetic state of each patch.
  • Edge Features: The state of each RF switch (open or closed) is encoded as an edge weight.
  • Physics Integration: The GNN is designed to emulate Kirchhoff's Current Law (KCL). This is achieved through specialized message-passing layers that aggregate information from neighboring nodes, mimicking how current flows and distributes within the antenna network. This physics-informed approach allows the GNN to learn the underlying electromagnetic principles rather than just statistical correlations.
  • Architecture: The GNN consists of an "interaction network" with message-passing layers for relational reasoning, followed by an edge classification MLP. This architecture learns to infer messages (analogous to current flow) between patches and then classifies each edge (switch) as ON or OFF.
  • Training: The model is trained using a weighted binary cross-entropy loss function. This loss is weighted to penalize incorrect predictions on "non-trivial" switches (those with significant current flow) more heavily than on "trivial" switches, preventing overfitting and focusing on critical components.

Datasets

The authors generated their dataset using CST Microwave Studio, a full-wave electromagnetic simulation software.

  • Data Generation: They randomly generated 3000 switch configurations. For each configuration, they performed forward simulations to obtain the corresponding surface current distributions and radiation patterns. This dataset serves as the ground truth for training and evaluating the GNN.
  • Validation: The model was trained using 9-fold cross-validation on this simulated dataset.

Key Insights and Ideas

  • GNNs for Antenna Topology: Representing the antenna as a graph is a natural fit, allowing the GNN to effectively capture the spatial connectivity and relationships between antenna elements and switches.
  • Physics-Informed Learning: Integrating KCL into the GNN architecture is a key innovation. This allows the model to learn more robustly and generalize better by understanding the fundamental physics governing antenna behavior, rather than relying solely on data patterns.
  • Inverse Design from Current Distribution: The ability to predict switch configurations directly from surface current distributions is a significant step towards practical inverse design. This bypasses the need to directly infer radiation patterns from switch states, which is often more complex.
  • Weighted Loss for Critical Components: The use of a weighted loss function that prioritizes non-trivial switches is crucial for achieving high accuracy in predicting the configurations that significantly impact the antenna's performance.

Key Results

  • High Test Accuracy: The model achieved a test accuracy of 93.2% in predicting switch configurations.
  • Performance Metrics: For non-trivial switches, the model demonstrated a True Positive rate of 87.97% and a True Negative rate of 98.02%.
  • Accurate Pattern Replication: Forward simulations using the GNN-predicted switch configurations successfully replicated complex far-field radiation patterns with a low Mean-Square Error (MSE) of approximately 0.001-0.003 in the tested examples.
  • EVM and SNR: On average across 10 randomly sampled validation configurations, the model achieved an Error Vector Magnitude (EVM) of 19.85 A/m and a Signal-to-Noise Ratio (SNR) of 35.32.

Limitations and Future Work

  • Simulation-Based Validation: The current validation is entirely based on simulations. The authors acknowledge the need for real-world experimental validation to confirm the model's performance in practice.
  • Scalability: While the architecture is optimized for the current antenna dimensions (7x7 array), future work will explore its scalability to antennas with varying dimensions.
  • Generalizability: The model is trained on a specific antenna design. Investigating its performance on different reconfigurable antenna architectures would be a valuable next step.

Summary updated

Abstract

Reconfigurable antennas, as a subclass of meta- surfaces, offer innovative and dynamic capabilities for wireless communication systems. Specifically, enabling radiation pat- tern reconfigurability allows for flexible beam steering through reverse-engineering of antenna parameters such as surface cur- rent distributions. In this work, we present a physics-informed machine learning model, leveraging fundamental physics such as Kirchhoff’s current Law, to predict the switch configurations of 2-dimensional antenna arrays. We utilize a graph neural network (GNN) to effectively capture the spatial relationships between radio-frequency (RF) switches and antenna patches, closely emulating the antenna topology. Simulation results demonstrate that our approach successfully predicts switch configurations needed to generate complex far-field radiation patterns.
BibTeX
@inproceedings{pan2025physics,
  author = {Cindy Pan and Naveen Verma and James Sturm},
  title = {{Physics-Informed Graph Neural Networks for the Inverse Design of GHz Reconfigurable Antenna}},
  booktitle = {IEEE European Conference on Antenna and Propagation},
  year = {2025},
  url = {https://openreview.net/forum?id=7kh9q0U4PZ},
}
Figure