Simulation acceleration for transmittance of electromagnetic waves in 2D slit arrays using deep learning

Abstract

When designing new optical devices, many simulations must be conducted to determine the optimal design parameters. Therefore, fast and accurate simulations are essential for designing optical devices. In this work, we introduce a deep learning approach that accelerates a simulator solving frequency-domain Maxwell equations. Our model achieves high accuracy while predicting transmittance per wavelength in 2D slit arrays under certain conditions to achieve 160,000 times faster results than the simulator. We generated a dataset using an open-source simulator and compared its performance with those of other machine learning models. Additionally, we propose a new loss function and performance evaluation method for creating better performance models with multiple regression outputs from one input source. We observed that using a loss function that adds binary cross-entropy loss, which predicts whether the differential of the transmittance is positive or negative at wavelengths adjacent to the root mean-squared error of the transmittance value, is more effective for predicting variations in multiple regression outputs. The simulation results show that a four-layer convolutional neural network model demonstrates the best accuracy (R² score: 0.86). The overall approach presented here is expected to be useful for simulating and designing optical devices.