Next-Value Prediction of Beam Waists From Mixed Pitch Grating Using Simplified Transformer Model

Abstract

In this study, a simplified transformer model is used to predict the beam waist of 1,092 nm light coupled out from SiN-based mixed pitch gratings at various heights. The beam waists data at various heights above the grating is first compiled. Then, we used a sequence of the current beam waist values, z-positions, and the computed mathematical indicators (features) to predict the next beam waist value (labels). Optimized transformer model yields average percentage error (APE) of 6.6% between the predicted and actual beam waists, which corresponds to 93.4% prediction accuracy. This study provides a pioneering approach to using natural language processing model to perform predictive modelling on photonics data, and possible extrapolation of photonics data using transformer model.

Keywords

• Attention
• gratings
• multi-head attention
• photonics integrated circuits
• quantum computing
• self-attention