A Machine Learning Approach for Efficient Mach-Zehnder Modulator Design

Abstract

The advancement of silicon photonics modulators is vital for achieving high-speed optical communications. However, designing optical modulators is a complex and resource-intensive task due to the large number of design parameters. Traditionally, the design of silicon photonics modulator depends on electrical and optical simulations. This work introduces employing machine learning models as a powerful design approach for MOS-like Mach-Zehnder Modulators to overcome the traditional design complexity. The proposed design and methodology build upon prior successful efforts in developing electro-optic modulators at the device level. The RandomForestRegressor model is developed to predict the effective refractive index (n_e) and a HistGradientBoostingRegressor for the absorption coefficient (k). These models show high prediction accuracy, with a mean absolute percentage error (MAPE) of 0.02% for the effective refractive index (n_e) and 0.80% for the absorption coefficient (k) for the test dataset. The developed models can predict the performance of MOS-like Mach-Zehnder Modulators within a few milliseconds, exhibiting a minimal margin of error. These results highlight the potential of integrating machine learning in photonic device design to simplify optimization, reduce the high computational and enable efficient exploration of new design spaces.