Machine Learning Regression Approach to the Nanophotonic Waveguide Analyses
Machine Learning Regression Approach to the Nanophotonic Waveguide Analyses
复制标题
DOI：
10.1109/jlt.2019.2946572
复制
发表时间：
2019-12-15
影响因子：
4.7
通讯作者：
Rahman, B. M. A.
中科院分区：
文献类型：
--
作者：
Chugh, Sunny;Ghosh, Souvik;Rahman, B. M. A.
机器学习是人工智能的一种应用，它专注于计算机算法的开发，通过从提供的数据中提取模式来自动学习。机器学习技术可以有效地用于有大量参数需要优化的问题，也可以用于无法开发执行任务的特定指令算法的问题。本文将有限元模拟与机器学习技术相结合，对不同集成光子器件的模式有效指数、功率约束和耦合长度进行了预测。首先，我们使用COMSOL Multiphysics准备了一个数据集，然后将该数据用于训练，同时优化机器学习模型的各种参数。通过改变波导宽度、高度、工作波长和其他器件尺寸来记录不同的模态解参数。对一个槽波导结构进行了详细的研究，以评估不同的机器学习模型参数，包括层数、节点数、激活函数的选择等。经过训练后，该模型用于预测新输入设备规格的输出。该方法比直接数值模拟技术更快地预测不同器件参数下的输出。对于槽波导、条形波导和定向波导耦合器设计，预测输出的绝对百分比误差小于5%。这项研究为将基于机器学习的优化技术用于集成硅光子器件铺平了道路。
Machine learning is an application of artificial intelligence that focuses on the development of computer algorithms which learn automatically by extracting patterns from the data provided. Machine learning techniques can be efficiently used for a problem with a large number of parameters to be optimized and also where it is infeasible to develop an algorithm of specific instructions for performing the task. Here, we combine the finite element simulations and machine learning techniques for the prediction of mode effective indices, power confinement, and coupling length of different integrated photonics devices. Initially, we prepare a dataset using COMSOL Multiphysics and then this data is used for training while optimizing various parameters of the machine learning model. Waveguide width, height, operating wavelength, and other device dimensions are varied to record different modal solution parameters. A detailed study has been carried out for a slot waveguide structure to evaluate different machine learning model parameters including number of layers, number of nodes, choice of activation functions, and others. After training, this model is used to predict the outputs for new input device specifications. This method predicts the output for different device parameters faster than direct numerical simulation techniques. Absolute percentage error of less than 5% in predicting an output has been obtained for slot, strip, and directional waveguide coupler designs. This study paves the step towards using machine learning based optimization techniques for integrated silicon photonics devices.