Grid-wise simulation acceleration of the electromagnetic fields of 2D optical devices using super-resolution

Area of Science:

• Computational physics
• Optical engineering
• Machine learning for simulation

Background:

• High-resolution simulations are crucial for device design but computationally expensive.
• Existing simulation methods face temporal limitations due to increasing computational demands with resolution.
• The cost of physical testing necessitates accurate and efficient simulation techniques.

Purpose of the Study:

• To develop a computationally efficient model for predicting high-resolution simulation results.
• To reduce the temporal limitations associated with high-resolution electromagnetic field simulations.
• To achieve high simulation accuracy with significantly lower computational cost.

Main Methods:

• Introduction of the fast residual learning super-resolution (FRSR) convolutional neural network model.
• Application of super-resolution techniques to predict electromagnetic fields.
• Utilizing residual learning and post-upsampling for enhanced accuracy and reduced computation.

Main Results:

• The FRSR model achieved high accuracy (R2: 0.9941) in predicting high-resolution outcomes.
• Demonstrated an approximately 18-fold increase in execution speed compared to traditional simulators.
• Achieved the shortest training time (7000 s) among super-resolution models evaluated.

Conclusions:

• The FRSR model offers a viable solution for accurate, low-cost, and fast electromagnetic field simulations.
• This approach effectively addresses the trade-off between simulation accuracy and computational resources.
• The developed model enhances the feasibility of high-resolution simulations in practical device design.