Microstrip Patch Antenna Design Using a Four-Layer Feed Forward Artificial Neural Network Trained by Levenberg-Marquardt Algorithm

This study proposes an efficient Artificial Neural Network (ANN) model for predicting the dimensions and feed point of microstrip patch antennas with three types of geometrical shape. The geometrical shapes investigated are square, triangular, and trapezoidal. Initially, the dimensions of these patches are varied using Advanced Design System (ADS) electromagnetic field simulator. The simulated performance parameters named resonance frequency, gain, and reflection coefficient are recorded and exported to the MATLAB workspace. Subsequently, the ANN is designed and trained using these performance parameters as input variables, while the dimensions and feed point serve as output variables. The Levenberg-Marquardt (LM) algorithm is employed for training due to its efficiency and faster convergence compared to other algorithms. The proposed ANN predicts not only the dimensions but also the feed point of the patch antennas. Therefore, patch antennas with three fundamental geometrical shapes can be designed using the same ANN removing computational complexity for the designers. The ANN contains a multi-layered network architecture that learns and generalizes complex patterns through the LM algorithm and weight optimization based on the datasets without any feature extraction like Deep Neural Network (DNN). Finally, the proposed ANN’s prediction results are represented in terms of mean-squared-errors (MSE) and regression plots. The ANN’s best obtained MSE and regression values are <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$10^{-6}$ </tex-math></inline-formula> and 0.99999, respectively, which show improvement over some recent models indicating robust accuracy and precision in predictions. To validate the performance of the proposed ANN, a trapezoidal patch antenna is simulated and fabricated based on one of the output values predicted by the ANN. It is found that the simulated and fabricated results approximately match, thereby validating the performance of the proposed ANN.