Prediction of reflection responses for randomized free-form antenna designs using a convolutional neural network

Abstract

This study presents a convolutional neural network optimized to predict the reflection responses from pseudo-random, multi-dimensional free-form antenna designs using 50 vertices. The proposed framework does not rely on expert knowledge and electromagnetic simulations during inference of the responses. The main contributions include: (1) a new dataset comprising 108,710 pseudo-random antenna topologies and their corresponding reflection characteristics; (2) a convolutional neural network capable of predicting the reflection response over a wide frequency range directly from complex pseudo-random antenna geometries; (3) a comprehensive evaluation, including training procedures, performance assessment, and comparison with alternative approaches. The input of the model is a 2×51 tensor that represents vertex coordinates of the antenna, while the output is a reflection characteristic (in decibels) in the form of a 251-point vector spanned over a frequency range from 3 GHz to 8 GHz (with 20 MHz step). The dataset is publicly available to ensure reproducibility. The hyperparameters for the network training were selected via Bayesian optimization. Furthermore, the model was trained and validated using 10-fold cross-validation, whereas its performance was evaluated using a separate test set. The model achieved a mean squared error of 3.788, a mean absolute error of 0.911, a coefficient of determination of 0.438, and a root mean squared error of 1.946 for the entire frequency range. The average prediction time of 1.9 milliseconds per sample enables a rapid evaluation of thousands of candidate designs, alleviating the computational bottleneck associated with extensive cost of electromagnetic simulations in the exploratory phase of antenna development.