Effects of hyperparameters and machine learning approaches in forecasting absorption behavior of GHz disk-shape metamaterials

Abstract

The last decade has witnessed an increasing interest in metamaterial absorbers (MMAs) because of their huge potential in a wide range of applications including energy harvesting, photodetectors, sensors, light modulators, infrared camouflage and wireless communication. Recently, machine learning (ML) has become one of the modern and powerful tools that can examine the design data in order to forecast the absorption behavior with much less effort and cost-effectiveness than conventional experimental and computation approaches. In this work, we utilize two ML algorithms, Polynomial Regression (PR) and Random Forest Regression (RFR), to predict the absorption strength and frequency of a symmetric disk-shape metamaterial structure operating within 10 and 16[Formula: see text]GHz. The proposed models are trained on hundreds of simulation-generated samples. We show that fine-tuning some hyperparameters results in higher forecasting performance. The dependence of predicted results on input parameters demonstrates that PR has better performance in predicting absorption strength, while both algorithms share similar accuracy in predicting the absorption frequency.