Reducing Complexity and data-set-Size Through Physics Inspired Tandem Neural Network

Owing to ample potential and versatile capabilities of artificial intelligence to solve intricate scientific problems, two regression-based artificial neural network (ANN) models are proposed to design and optimize nano-structured meta-atoms. The proposed forward predicting ANN depicts that considering the complete structural and material information of the cylindrical nano-pillar meta-atoms could predict the corresponding electromagnetic (EM) response (amplitude and phase of transmission) with a mean squared error (MSE) as low as
$\mathbf{2.1}\times \mathbf{10}^{-\mathbf{3}}$
. Thus, it replaces the conventional EM simulations performed using high-end commercial software's, while significantly saving time and computational resources. Inverse design deep-learning model is also presented, which is connected with the pre-trained forward model and trained in a tandem architecture to provide an optimum set of dimensions and material, given the target response as its input. Furthermore, a comparative study regarding the number of hidden layers of the ANN and the amount of training dataset size is performed for the proposed forward and tandem inverse models to analyze the effect of considering extra underlying physics related information, i.e., wavelength regime and the EM spectral information. This study reveals that considering the extra information can lead to a significant reduction in the obtained MSE. Specifically, the proposed model could achieve a decent MSE even with a smaller amount of training dataset. Hence, the use of artificial intelligence models significantly reduces the training time and computational complexity of the proposed solution.