Characterizing Volume Density of Subwavelength Particles at 220-325 GHz Using Deep Neural Network and Nonfeatured Scattering Matrix

The feasibility of using machine-learning algorithm on classification and numerical prediction method for characterizing volume density is explored. The deep neural network (DNN) is exploited to describe the relationship of input and output data when the analytical modeling or simulation is unavailable. In this letter, this approach is exemplified for the extraction of relative volume density of subwavelength particles at 220-2013;325-00A0;GHz. The training based on the phase of transmission coefficients ascertains classification accuracies of 99.9-0025; and prediction mean squared error of 0.0186. In addition, the training based on the real and imaginary parts of the scattering matrix can also achieve high classification accuracy (> 94.6-0025;). It concludes that the DNN can autonomously retrieve correlation of electromagnetic properties from the nonfeatured real and imaginary parts of the scattering matrix.