DATA-DRIVEN PLASMA SIMULATION ON ATMOSPHERIC RADIO FREQUENCY DISCHARGE PLASMAS

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 low-temperature 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 in α and γ modes, 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 10⁵ ~ 10⁷ 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 data-driven techniques could strongly promote the development of low-temperature plasma.