Modelling and next-value prediction of beam propagation from grating structures using a simplified transformer model

Area of Science:

• Photonics and optical engineering
• Computational physics
• Machine learning applications

Background:

• Silicon photonics gratings are crucial for coupling light between integrated circuits and free space.
• Traditional simulation methods like Finite-Difference Time-Domain (FDTD) are computationally intensive and time-consuming.
• Accurate prediction of electric field (E-field) distribution is essential for device design and performance optimization.

Purpose of the Study:

• To develop and evaluate a simplified transformer model for predicting E-field distribution from silicon photonics gratings.
• To assess the accuracy and efficiency of the transformer model compared to conventional FDTD simulations.
• To demonstrate the potential of machine learning for accelerating photonic simulations.

Main Methods:

• Utilized a simplified transformer model for next-value prediction.
• Performed Finite-Difference Time-Domain (FDTD) simulations to generate E-field data for gratings with varying pitches (0.6 to 1.6 µm).
• Trained the transformer model using E-field data from a 0.6 µm grating and predicted E-fields for other grating pitches.

Main Results:

• Achieved prediction accuracy up to 92.5% for E-field distribution.
• Model training completed in 1908.4 seconds, a significant speedup over multi-hour FDTD simulations.
• Demonstrated that transformer models can predict E-field with minimal training data.

Conclusions:

• A simplified transformer model offers a highly efficient and accurate method for predicting E-field in silicon photonics.
• The developed model can significantly expedite FDTD simulations when integrated into existing software.
• This approach highlights the potential of AI in advancing computational photonics research and development.