Unraveling the resonant frequency of H-shaped microstrip antennas using a deep learning approach

Abstract

This paper introduces a novel physics-informed learning approach that combines principles from physics with deep learning techniques to optimize the simulation process of microstrip antennas. These deep learning-based approaches are preferable because traditional full-wave models used in antenna design are computationally intensive and require significant memory due to their reliance on iterative algorithms, leading to exponential increases in resource demands as input parameters grow. In contrast, the proposed deep learning method requires significant computational resources only during training, with a constant time complexity of O(1) during deployment. This results in much faster modeling, allowing a broader range of antenna configurations to be processed more quickly, thereby improving the efficiency of the design workflow. Unlike conventional deep learning methods that rely solely on data, our approach leverages the underlying physical laws governing antenna behavior, particularly beneficial when labeled data is scarce or difficult to obtain. We propose a bias observational physics-informed learning technique by integrating physical laws into the loss function, which includes two components: Neuron Loss, the standard MSE measuring prediction accuracy against actual data, and Physics Loss, which penalizes deviations from physical laws as represented by a cavity model. The total loss combines these two, with higher physics loss indicating poorer alignment with physical principles and lower physics loss suggesting better adherence to them. This approach refines predictions by balancing data fidelity with physical constraint, wherein the dataset is sourced from simulations and real-world measurements. This strategy ensures model uncertainty and broad generalization capabilities. Computational efficiency is a key consideration, with our approach implemented on low-specification hardware, emphasizing optimal resource and power consumption. The H-shaped microstrip antennas (HMAs), known for its wide and dual-band properties, serves as the target antenna for our study. We employ three sequential models’ recurrent neural networks (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU)—integrated with a cavity model-driven resonance frequency representation to maintain the resonance mode TM¹⁰ at prediction. Comparative analysis of these models encompasses execution time, prediction convergence, loss reduction, prediction score (R²), as well as memory and CPU usage. This research contributes four main sections elucidating the methodology, experimental setup, and results analysis, underscoring the efficacy of our deep learning approach in antenna optimization.