Replication of Particle Trajectories in the Plasma Jet with Two Consecutive Residual Neural Networks

Abstract

AbstractDue to the complexity of the multi-arc plasma spraying process in combination with the harsh ambient conditions, i.e., extremely high temperatures and velocities, the use of numerical analysis, such as modern methods from computational fluid dynamics (CFD), is unavoidable to gain a better understanding of the coating process. However, the tradeoff between the accuracy of the increasingly sophisticated CFD models and their computation time has always been a concern. This study presents a novel machine learning approach capable of predicting the temperatures, velocities, and coordinates of the in-flight particles in a plasma jet. To this end, two individual residual neural networks are trained consecutively with CFD simulation data sets, in a way that the deviations between the targets and predictions of the first network are used as additional inputs for the second network. The results for test data not used during the training of the networks reveal that the simulated particle trajectories in the plasma jet can be fully replicated by the developed machine learning approach. This indicates the potential of the approach to replace the CFD simulations of the plasma jet, which reduces the computation time from several hours to a few seconds.