New Deep Learning Paradigm For Pulsed Discharge Plasma Catalysis Modeling

Plasma catalysis exploits the mutual synergistic effect between plasma and catalyst and has promising applications in energy conversion and atmospheric pollutant control.The results of numerical simulations can effectively guide experiments and supplement physical information that cannot be easily measured experimentally,which can significantly save research costs and improve research efficiency.Establishing numerical models of plasma catalysis,combining numerical simulation results with experimental diagnostic data,decoupling the mutual synergistic effects between plasma and catalyst,and studying the mechanism and characteristics of energy conversion and utilization in plasma catalysis are academic hotspots and frontier topics in the intersection of gas discharge,multiphase catalysis and plasma physics.In recent years,with the continuous advancement of artificial intelligence and machine learning technology,deep neural networks have provided a new paradigm for the numerical simulation of pulsed discharge plasma catalysis.However,both tightly coupled differential equations and complex sets of plasma-catalytic reactions suffer from poor convergence and high computational resource requirements for numerical simulations on multiple time scales;also,incomplete data sets of plasma and surface reaction rate coefficients are a major obstacle to the development and application of numerical simulations of pulsed discharge plasma catalysis.In response to the above problems,this paper systematically builds up a plasma-catalyzed reaction data set,including reaction equations and corresponding reaction rate coefficients between various gases used for energy conversion and air pollution control.On this basis,the calculation of different types of surface reaction rate coefficients between the plasma and catalyst surface components are given respectively.In addition,multilayer feed-forward deep neural networks are introduced into the numerical simulation of the kinetics of pulsed discharge plasma and pulsed discharge plasma catalysis.The Matlab script file is called in the Python language under the Py Charm compiled environment to construct a deep neural network algorithm based on the results of the zero-dimensional kinetic numerical simulation,and the initial input parameters required for the kinetic simulation are extracted from the specific experimental data The reduced electric field（E/N）is used to link the pulsed discharge plasma experiments with the zero-dimensional kinetic simulations.This was validated in the CH₄/Ar pulsed discharge plasma model and the N₂/H₂pulsed discharge plasma catalytic model.The work performed has resulted in the following studies:（1）The deep neural network was extended with rich input and output data based on the kinetic simulation results,and the computational results of the deep neural network extensions were in good agreement with the kinetic simulation results.The relative errors of the CH₄/Ar pulsed discharge plasma model and the N₂/H₂pulsed discharge plasma catalytic model were1.15×10^-3and 4.19×10^-4,respectively,at which time the neural network architectures were 7hidden layers with 80 neurons each and 8 hidden layers with 70 neurons each,respectively.（2）For the CH₄/Ar pulsed discharge plasma model,the E/N amplitude predicted by the neural network was used as the input value for the kinetic numerical simulation,and the average relative error between the numerical simulation results and the experimental results was 0.28%.Argon plays a non-negligible role in the methane-to-hydrogen process,52.9%of the hydrogen is produced by the reaction of excited state argon with the background gas methane,and the dilution of argon changes the energy distribution function and energy loss fraction of the methane plasma.（3）For the N₂/H₂pulsed discharge plasma catalytic model,the E/N amplitude predicted by the neural network was used as the input value for the kinetic numerical simulations,and the average relative error between the numerical simulation results and the experimental results was0.61%.the surface reaction involving the Ru/Mg O catalyst contributed about 87%of the ammonia in the synthesis of ammonia gas,of which 83.0%was produced by the Eley-Ridea reaction NH（s）+H₂→NH₃+Surf reaction.Deep learning provides a new paradigm for the numerical simulation of pulsed discharge plasmas,offering a new approach to parameters that are difficult to determine in numerical simulations of plasma dynamics.More importantly,in addition to the initial demonstration of the capabilities of deep learning in these tasks,it provides a new and promising numerical tool for the numerical simulation of plasmas in the plasma domain.Meshless deep learning can solve to some extent the convergence problem in numerical simulations by the finite element method or the finite volume method,and offers the possibility to optimize the research process and integrate the results,assimilating experimental and simulation data.