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

Area of Science:

• Optics and Photonics
• Computational Physics
• Machine Learning Applications

Background:

• Designing optical devices requires extensive simulations to optimize parameters.
• Current simulation methods for frequency-domain Maxwell equations can be computationally intensive.
• Fast and accurate simulations are crucial for efficient optical device design.

Purpose of the Study:

• To develop a deep learning model for accelerating optical simulations.
• To improve the speed and accuracy of predicting transmittance in 2D slit arrays.
• To introduce novel loss functions and evaluation methods for regression tasks in optical modeling.

Main Methods:

• A deep learning approach was developed to accelerate a frequency-domain Maxwell equation solver.
• A dataset was generated using an open-source optical simulator.
• A novel loss function combining root-mean-squared error and binary cross-entropy was proposed.
• A four-layer convolutional neural network (CNN) was trained and evaluated.

Main Results:

• The deep learning model achieved 160,000 times faster simulation results compared to the traditional simulator.
• The model demonstrated high accuracy in predicting transmittance per wavelength (R² score: 0.86).
• The proposed loss function improved the prediction of variations in multiple regression outputs.

Conclusions:

• Deep learning significantly accelerates optical simulations for device design.
• The developed model offers a highly accurate and efficient alternative to conventional methods.
• The approach is expected to be valuable for the simulation and design of various optical devices.