Data-Efficient Machine Learning Algorithms for the Design of Surface Bragg Gratings

Abstract

Deep learning models, with a prerequisite of large databases, are common approaches in applying machine learning for inverse design in photonics. For these models, less expensive, approximate methods are usually used to generate large databases, which limit their applications. In this study, we compare the performance of data-efficient machine learning (ML) models for predicting the characteristics of surface Bragg gratings in semiconductor ridge waveguides. We employ the 3D finite-difference time-domain method which is very accurate but time-consuming to generate a database. We analyze the performance of different ML models including support vector regression and extreme gradient boosting (XGBoost) on our limited data. We show that the XGBoost significantly outperforms other models on a smaller database. Our results pave the way for the data-efficient design of integrated photonic components with accurate but time-demanding simulations.

Optimizing data acquisition: a Bayesian approach for efficient machine learning model training

“…The high-dimensional data can make it difficult for SVR to find an effective hyperplane in such a high-dimensional space [32]. On the other hand, XGBoost represents a complicated algorithm which has proven to be flexible in learning intricate relationships [16,34]. As we will see in the following not only the complexity of the database, but also the complexity of the ML algorithm plays a crucial role in the overall performance.…”

“…The SVR model used in this study employs a kernel to map the data to a higher-dimensional space for easier problem solving. We compared in our previous study [16] linear, RBF, and polynomial kernels, determining that the RBF kernel yields the highest precision. The adjustable parameter ϵ in SVR specifies the width of the so called 'tube' around the function being evaluated.…”

“…We provide an example by building a database that includes the characteristics of Bragg gratings and their respective reflectance spectra. The database is generated using finite-difference time-domain (FDTD) simulations [15][16][17]. A ML model trained on this small database should significantly outperform the same model trained on the same number of data points collected by traditional means (uniform or randomly distributed data points).…”

“…As an example, designing Bragg gratings to achieve a precise optical response is crucial for fabricating diode lasers with specific applications, ranging from telecommunications to sensors, optical atomic clocks and quantum technologies [18][19][20][21][22][23]. Inverse design of these Bragg gratings using ML models [16,24] necessitates collecting a substantial amount of data in multi-dimensional parameter spaces to achieve high predictive accuracy. This, in turn, can be time-consuming and resource-intensive.…”