结合机器学习的大气压介质阻挡放电数值模拟研究

Numerical study of discharge characteristics of atmospheric dielectric barrier discharges by integrating machine learning

大气压介质阻挡放电是应用中常用的放电形式,通常使用等离子体流体模型进行理论描述.本文针对大气压均匀介质阻挡放电每半个电压周期出现一次或多次电流脉冲的特性,基于机器学习方法构造一个全连接多层神经网络,采用误差反向传播算法,并设计了一个通用的隐藏层结构,将计算数据或实验数据作为训练集,借助于人工神经网络程序研究大气压介质阻挡放电的电流密度、电子密度、离子密度和电场强度等宏观与微观放电特性.通过分析计算结果可知,在给定合适训练集的条件下,构造的机器学习程序与流体模型能以近乎相同的计算精度(误差小于2%)来描述大气压介质阻挡放电等离子体性质,同时计算效率远高于求解流体模型,并能极大地拓展放电参数的遍历范围.本文的算例表明,将机器学习程序与现有的流体模型或动理学模型结合起来,将极大地提高大气压放电等离子体的模拟效率与效果,深化对放电等离子体的认识.

In recent years,with the development of gas discharge technology at atmospheric pressure,the application of low temperature plasma has received widespread attention in pollution prevention,disinfection,sterilization,energy conversion and other fields.Atmospheric dielectric barrier discharge is widely used to produce low temperature plasma in various applications,which is usually numerically investigated by using fluid models.The unique advantages of machine learning in various branches of physics have been discovered with the advancement of big data processing technology.Recent studies have shown that artificial neural networks with multiple hidden layers have a pivotal role in the simulation of complex datasets.In this work,a fully connected multilayer BP(back propagation)network together with a universal hidden layer structure is developed to explore the characteristics of one or more current pulses per half voltage cycle of atmospheric dielectric barrier discharge.The calculated data are used as training sets,and the discharge characteristics such as current density,electron density,ion density,and electric field of atmospheric dielectric barrier discharge can be quickly predicted by using artificial neural network program.The computational results show that for a given training set,the constructed machine learning program can describe the properties of atmospheric dielectric barrier discharge with almost the same accuracy as the fluid model.Also,the computational efficiency of the machine learning is much higher than that of the fluid model.In addition,the use of machine learning programs can also greatly extend the calculation range of parameters.Limiting discharge parameter range is considered as a major challenge for numerical calculation.By substituting a relatively limited set of training data obtained from the fluid model into the machine learning,the discharge characteristics can be accurately predicted within a given range of discharge parameters,leading an almost infinite set of data to be generated,which is of great significance for studying the influence of discharge parameters on discharge evolution.The examples in this paper show that the combination of machine learning and fluid models can greatly improve the computational efficiency,which can enhance the understanding of discharge plasmas.