Accurate and efficient prediction of photonic crystal waveguide bandstructures using neural networks

Abstract

We demonstrate the use of neural networks to predict the optical properties of photonic crystal waveguides (PhCWs) with high accuracy and significantly faster computation times compared to traditional simulation methods. Using 100,000 PhCW designs and their simulated bandstructures, we trained a neural network to achieve a test set relative error of 0.103% in predicting gap guided bands. We use pre-training to improve neural network performance, and numerical differentiation to accurately predict group index curves. Our approach allows for rapid, application-specific tailoring of PhCWs with a runtime of sub-milliseconds per design, a significant improvement over conventional simulation techniques.