基于数据驱动的大气压射频放电等离子体数值模拟研究

DATA-DRIVEN PLASMA SIMULATION ON ATMOSPHERIC RADIO FREQUENCY DISCHARGE PLASMAS

随着人工智能技术的进步,结合低温等离子体的物理特点,数据驱动技术由于其独特的优势在低温等离子体的研究中正逐渐兴起.本研究以深度神经网络(DNN)模型在大气压射频放电中的计算研究为例,讨论了数据驱动方法在低温等离子体模拟研究中的应用.对于低温等离子体的研究而言,数据驱动研究所需要的数据可以来自于实验诊断和数值计算,根据等离子体物理特性的不同也可以选择不同的数据驱动模型.粒子模型与流体模型是低温等离子体研究中常用的两类计算模型,基于这两者的模拟数据组成的训练集,DNN可以实现对大气压射频放电的动理学特性等各种特性的实时预测.首先通过将流体模型与粒子模型计算结果与DNN模型的预测结果相比较,验证了DNN模型在给定精度下的有效性.然后基于流体模拟数据,利用DNN探究了α和γ模式下输入电流密度和放电间隙对大气压射频放电特性的影响,最后借助于粒子模拟数据构建的训练集,讨论了大气压射频微放电的频率效应,特别是电子能量分布函数(EEDF)的演化.预测结果表明,经过大约1 h的训练后,DNN只需要耗时0.01 s左右就能以极高的计算精度(与数值模拟之间的相对误差小于0.5%)获得电子密度、电场强度和EEDF等大气压射频放电的特定物理信息,而基于流体模拟或者粒子模拟中获得一组稳定的模拟结果分别需要大约半个小时和几十个小时.可以说,在相同的计算精度下,经过训练后DNN的预测效率较传统数值模拟效率提高了约10^(5)~10^(7)倍,可以近乎实时地给出计算结果.另一方面,基于有限的训练集,DNN可以快速产生大量的特定预测数据,这将极大丰富和强化原有的数值模拟效果,更好地体现射频等离子体的演化规律.本研究以DNN模型在大气压射频放电数值模拟中的应用为例,表明数据驱动技术的引入将有力地推动低温等离子体研究的发展.

With the advancement of artificial intelligence,data-driven techniques are emerging in the field of low temperature plasma due to their unique advantages.In this study,the application of deep neural network(DNN)in atmospheric radio frequency(RF)discharge is taken as an example to investigate the numerical simulation of lowtemperature plasma based on data-driven methods.For data-driven low-temperature plasma studies,training data could be selected from experimental diagnostics and numerical simulations,and different data-driven models can also be selected according to various plasma properties.Particle model and fluid model are commonly used in low-temperature plasma simulations and based on a training dataset consisting of particle simulation and fluid simulation data,DNN enables real-time prediction of various properties of atmospheric RF discharges,including kinetic properties.The effectiveness of the DNN is verified by comparing the DNN prediction results with the numerical simulation results.Subsequently,based on the fluid simulation data,the DNN is employed to investigate the effects of input current density and electrode spacing on the atmospheric RF discharge operating inαandγmodes,and finally the frequency effects of atmospheric RF micro-discharge,especially the evolution of the electron energy distribution function(EEDF),are discussed based on the training dataset consisting of particle simulation data.The prediction results show that after about one hour of training,DNN only takes about 0.01 second to obtain the specific discharge characteristics(such as electron density,electric field,and EEDF)with very high accuracy(relative error less than 0.5%with respect to the simulation results).In contrast,it takes about half an hour and tens of hours to obtain stable simulation results in fluid simulation and particle simulation,respectively.It can be said that the prediction efficiency of trained DNN is about 10^(5)~10^(7) times higher than the computational efficiency of traditional numerical simulations,and the prediction results can be provided in near real-time.In addition,DNN can rapidly generate infinite prediction data based on limited training data,which could greatly enrich and strengthen the original numerical simulation and better describe the evolutionary behavior of the RF plasma.In this study,the application of DNN in atmospheric RF discharges is used as an example to show that datadriven techniques could strongly promote the development of low-temperature plasma.