Simulation Neural Network Trained by FDTD Dataset for Determining Freq. Responses of Metagratings
4Abstract: In recent years, Convolutional Neural Network (CNN) has been used for various machine learning tasks such as Image Recognition, Computer Vision, and Natural Language Processing. Recent developments on its application in designing electromagnetic devices such as metamaterials lead to the fast and precise computation of seemingly slow and difficult to obtain frequency responses through numerical methods such as Finite Difference Time Domain (FDTD) and Finite Element Analysis (FEM). This study aims to develop a CNN, to be called a Simulation Neural Network (SNN), that can produce frequency responses of metagratings of different geometric configurations faster than FDTD can do. The CNN architecture is composed of 10 convolutional blocks, including the input block, and 4 fully dense layers, plus another dense layer for the sigmoid function, producing the discretized frequency responses. The SNN is written in Tensorflow 2 programming framework and trained using the frequency response data from the FDTD simulation of 2450 samples of metagratings, with loss set to mean squared error, metric set to accuracy, and optimizer set to Adam optimization algorithm. For the transmission dataset, the SNN training reported a loss of 1.3751e-4 and an accuracy of 82.63%, while the in the validation set, a mean square error of 2.496e-04 at 83.20% accuracy. Also, the SNN can produce predictions within seconds for bulk testing, in contrast to FDTD, giving off results within minutes. Plotting the FDTD results together with the predicted plots from SNN shows close predictions from the numerical method dataset. This makes the SNN a powerful complement for numerical methods in analyzing electromagnetic structures.
Key Words: Neural Network; Metamaterials; Finite Difference Time Domain