Prediction of Particle Properties in Plasma Spraying Based on Machine Learning

Abstract

Thermal spraying processes include complex nonlinear interdependencies among process parameters, in-flight particle properties and coating structure. Therefore, employing computer-aided methods is essential to quantify these complex relationships and subsequently enhance the process reproducibility. Typically, classic modeling approaches are pursued to understand these interactions. While these approaches are able to capture very complex systems, the increasingly sophisticated models have the drawback of requiring considerable calculation time. In this study, two different Machine Learning (ML) methods, Residual Neural Network (ResNet) and Support Vector Machine (SVM), were used to estimate the in-flight particle properties in plasma spraying in a much faster manner. To this end, data sets comprising the process parameters such as electrical current and gas flow as well as the in-flight particle velocities, temperatures and positions have been extracted from a CFD simulation of the plasma jet. Furthermore, two Design of Experiments (DOE) methods, Central Composite Design (CCD) and Latin Hypercube Sampling (LHS), have been employed to cover a set of representative process parameters for training the ML models. The results show that the developed ML models are able to estimate the trends of particle properties precisely and dramatically faster than the computation-intensive CFD simulations.

Science-based, data-driven developments in plasma processing for material synthesis and device-integration technologies

“…They have demonstrated that the trends of particle properties are estimated to be extremely faster than the averaged plasma jet simulation (from 3 h to less than 5 s). 242) By technological outcomes, this field is becoming a hot topic for controlling using the ML and AI techniques without physical understanding of the significant parameters. We emphasize the importance of understanding the physics of the system in order to achieve comprehensive control of critical parameters.…”

Machine learning tools in production engineering

“…The goal is to establish a digital twin of the simulation process which significantly reduces the computation time while still sufficiently predicting the general behavior of the particles. An approach for training a ResNet similar to the one presented in the following and a comparison of ResNet results with results of a support vector machine can be found in [25]. Further ResNet results are also available in [37].…”

“…First, we study the prediction of specific cutting forces for different materials and compare the results with measurement data from, e.g., [22]. Second, the particle behavior in a plasma spraying coating process [23][24][25] is predicted based on large simulation data sets. These use cases stem from cooperations within the DFG project EXC-2023 "Internet of Production".…”

Continuous limits of residual neural networks in case of large input data

“…Note that even the choice d = 1 is possible and, in addition, that networks of this type have been already proved to satisfy different formulations of the universal approximation theorem [31][32][33]. Further, they have been also applied to several (realworld) training problems [34,35].…”