In the High Velocity Oxygen Fuel (HVOF) technology, the coating properties are sensitive to the behaviors of in-flight particles, which are mainly influenced by the processing parameters. However, due to the complex chemical and thermodynamic reactions, the real-time optimization of the coating properties during the HVOF process is still a challenging issue. This study focused on establishing an Artificial Neural Networks (ANN) model to analyze the influence of the processing parameters on the characteristics of in-flight particles. Hydroxyapatite (HA) powders were selected to deposit onto the stainless steel substrates via an improved HVOF spraying system. Combined with an Accuraspray-g3 system applied to acquire the temperature and velocity of inflight HA particles, the artificial neural network algorithm was well trained to predict the velocity and temperature of in-flight particles. The relationship between the variations of the operating parameters (gas flow rates and fuel-to-oxygen ratio) and the behaviors of in-flight HA particles was investigated, which therefor contributes to analyzing and optimizing the mechanical performance and crystallinity of the HA coatings.

Traditionally, large moulds for manufacturing of CFRP (carbon fiber reinforced plastics) parts are machined from a solid metal block making this way of manufacturing very energy and time consuming. Using wire arc spraying thin-walled moulds can be produced by spraying onto an original mould and separating the coating. In order to create a reliable and high quality product the manufacturing process needs to be highly reproducible. Thus the spraying process requires monitoring and control, which can be done using artificial neural networks (ANN). In our approach, for monitoring the process the diagnostic system PFI (Particle Flux Imaging) is used to characterize the spray particle stream, which is essentially achieved by fitting an ellipse to an image of the particle stream. Comparing deviations from a reference ellipse recorded for an “optimal” coating process provides data that can subsequently be used for process control. Investigations performed by the method of design of experiments (DOE) show a very strong correlation of the parameters pressure, current, and voltage with certain parts of the PFI data: for example the semi-minor axis of the ellipse depends linearly on voltage and current but quadratic on pressure. These results can further on be used to control the coating process by ANN. This paper discusses the application of this method and its feasibility for industrial use.

On-line process control appears today as a potential way to improve thermal spray coating quality and reliability. In order to reach that goal, several requirements are needed. Firstly, reliable sensors have to be developed to accurately diagnose in-flight particle characteristics but also some other parameters, such as coating surface temperature during spraying. Secondly, a robust command, linked to a model, is required to insure the stability of the control system. Finally, an extensive and representative database is required as reference. Several methodologies can be implemented to define such a structure. Artificial Neural Networks (ANNs) and neuromimetic models, based for example on fuzzy logic, appear as an interesting way with a large potential of improvement to control non-linear systems, such as the thermal spray processes.

One approach for controlling the twin wire arc spray (TWAS) process is to use optical properties of the particle beam like length or brightness of the beam as input parameters for a process control. The idea is that changes in the process like eroded contact nozzles or variations of current, voltage and/or atomizing gas pressure can be detected through observation of optical properties of the particle beam. It can be assumed that if these properties deviate significantly from those obtained from a beam recorded for an optimal coating process the spray particle and so the coating properties change significantly. Thus, the goal is to detect these optical deviations and compensate occurring errors by adjusting appropriate process parameters for the wire arc spray unit. One cost effective method for monitoring optical properties of the particle beam is to apply the process diagnostic system PFI (Particle Flux Imaging): PFI fits an ellipse to an image of a particle beam thereby defining easy to analyze characteristic parameters by relating optical beam properties to ellipse parameters. Using artificial neural networks (ANN) mathematical relations between ellipse and process parameters can be defined. Thus in the case of a process disturbance through the use of an ANN-based control new process parameters can be computed to compensate particle beam deviations. In this paper, it will be shown that different process parameters can lead to particle beams with the same PFI parameters.

Parametric drifts and fluctuations occur during plasma spraying. These drifts and fluctuations originate especially from the electrode wear and intrinsic plasma jet instabilities: the plasma net power varies in this case modifying significantly its thermodynamic properties and transport coefficients hence modifying the momentum and heat transfers to the particles. It is possible to control the in-flight particle characteristics by adjusting continuously the operating parameters, in particular the power parameters. Due to the large amplitudes of these drifts and fluctuations, the strategy to adopt will depend on the required corrections to apply to the particle characteristics. Developing a reliable controller requires: (i) implementation of reliable sensors to accurately diagnose in-flight particle characteristics but also some other parameters, such as surface temperature during spraying; (ii) development of a robust command, to insure the stability of the control system. Fuzzy logic permits to define parametric correction rules and the command can be based on these algorithms; (iii) linking of the robust command to a predictive model. Artificial neural networks, among other artificial intelligence protocols such as genetic algorithms, proved to be able to predict in-flight particle and coating characteristics; (iv) validation of the corrections with an extensive database used as a reference. Coupling neural protocols to fuzzy logic should permit the development of such an independent controller which could adjust in real time the operating parameters as a function of the measured in-flight particles to manufacture a coating with identical conditions among its entire thickness. This paper aims at presenting the methodology and the controller and at simulating the spray operation.

Plasma spraying is a key industrial coating process that exhibits intricate nonlinear interactions among process parameters. This complexity makes accurate predictions of particle properties, which greatly affect process behavior, very challenging. Specifically, particle velocities and temperatures profoundly impact coating quality and process efficiency. Conventional methods often require empirical correlations and extensive parameter tuning due to their limited ability to capture the underlying physics within this intricate system. This study introduces Physics-Informed Neural Networks (PINNs) as a solution. By seamlessly integrating known physical laws and constraints directly into the model architecture, PINNs offer the potential to learn the underlying physics of the system. For comparison, Artificial Neural Networks (ANNs) are also developed. Computational Fluid Dynamics (CFD) simulations of a plasma generator and plasma jet model provide data to train both ANN and PINN models. The study reveals an improvement in particle velocity prediction through the proposed PINN model, demonstrating its capability to handle complex relationships. However, challenges arise in predicting particle temperature, warranting further investigation. The developed models can aid in optimizing the plasma spraying process by predicting essential particle properties and guiding necessary process adjustments to enhance coating quality.

In this work, an artificial neural network (ANN) model was developed to investigate the application of Cr 3 C 2 -25NiCr coatings by HVOF spraying and predict the resulting properties based on flow rates, stand-off distance, and other parameters. HVOF coatings were sprayed and tests were conducted to generate data for training, validating, and testing the model. The model was trained with an R-value of 0.99965 to predict the relationship between spray parameters and coating properties including hardness, porosity, and wear rate. The reliability and accuracy of the model was subsequently verified using independent test sets.

During plasma spray process, many intrinsic operating parameters allow tailoring the in-flight particle characteristics (temperature and velocity), thus affecting the final coating characteristics. Among them, plasma enthalpy, thermal conductivity, momentum, density, etc. result from the selection of extrinsic operating parameters such as the plasma torch nozzle geometry, composition and flow rate of plasma forming gases, the arc current intensity, etc. The complex relationships among those operating parameters make it difficult to fully predict their effects. Moreover, temporal fluctuations (anode wear for example) require "real time" corrections to maintain particle characteristics to targeted values. In addition, substrate temperature has to be maintained to targeted values depending upon the feedstock to be sprayed, the geometry of the part to be coated, its thermal capacity, etc. An expert system was built to optimize and control some of the main extrinsic operating parameters. This expert system includes two parts: 1) an artificial neural network (ANN), which predicts an extrinsic operating window and 2) a fuzzy logic controller (FLC) to control it. The paper details the general architecture of the system, discusses its limits and typical characteristics. An example is finally presented.

In this study, the magnetic properties of iron-based coatings obtained by HVOF thermal spraying were investigated. These properties were correlated to alloy type, heat treating temperature and coating thickness using artificial neural network method. Among coating characteristics, porosity is regarded as an influential parameter on the magnetic properties. Therefore, the role of the microstructure was particularly emphasized in this study. Magnetic properties, especially coercivity, showed a weak decrease with the addition of alloying elements. However, these properties remained nearly unchanged at high temperatures and for large coating thicknesses. Porosities are regarded as defects anchoring Bloch walls and consequently promoting an increase of coercivity.

Off-line control of thermal spraying generally implements the quantification of the energetic, environmental, injection and geometric conditions. The interdependence of these parameters with regard to the deposit properties were studied with different statistical methodologies. The aim was to point out the large possibilities of the Artificial Neural Network methodology in the process optimization. In that way, experiments were designed considering atmospheric plasma spraying and the following operating parameters: arc current intensity, powder feed rate, carrier gas flow rate, total plasma gas flow rate, primary hydrogen content and scanning step. These parameters were related to the coating porosity level and hardness (HK). For each case, simple correlations were studied with none linear regressions, design of experiments and a Multilayer Perceptrons (MLP, with two hidden layers). Non linear correlations were compared with MLP based on the Average Sum of Absolute Error (ASAE) and other correlation factors (RSS, R 2 , Ra 2 , STD, RS and AR).

In thermal spray process, the in-flight particle characteristics such as particle size, velocity and temperature influence significantly their flight duration as well as their melting degree. Consequently, they influence the splat formation and ultimately the coating properties. Thus, the knowledge of the interactions between the process parameters and the in-flight particle characteristics is very important for optimizing the coating properties. Artificial Neural Network (ANN) concept was used to predict in-flight particle velocity and temperature considering the case of alumina (Al 2 O 3 -TiO 2 ) coatings. Databases of in-flight particle characteristics (diameter, velocity and temperature) versus spray process parameters (arc current intensity, hydrogen rate and plasma gas composition) were collected. ANN was trained with the database to establish the relationships linking the particle diameter and spray process parameters to particle velocity and temperature. Then, the established ANN relationships permitted to determine the inflight particle velocity and temperature versus their diameter for given spray process parameters. These velocity and temperature data were then used to determine the time for complete particle melting and the particle dwell-time before impact by an analytical model for given operating conditions.

After having established an offline process control based on the optical process diagnostic system PFI (Particle Flux Imaging) the process robustness has been increased by experimental identification of the impact of the noise factors „electrode wear“ and „injector wear“ in the APS process. Neural networks were used for implementing a process controller into a GTV APS process control center. In combination with the PFI system a tool was installed in the process center that is able to predict coating quality by analyzing some characteristics of the plasma and the particle plume. The neural network can be trained for all applications and all feedstock materials. An offline controller is instructing the operator through a desktop in order to train the network. The training, that means the monitoring of the process through different parameter setups and its reactions, is generated and executed automatically. Due to the fact that controlling the process parameters cannot influence every aspect of the coating quality, noise factors have to be regarded. For the APS process the electrode wear and the injector wear were identified as the most influencing noise factors. Both were analyzed by means of Design of Experiment (DOE) and long-term monitoring (200 hrs). The samples were characterized by light microscopy and different coating test methods that were chosen with respect to the coatings functions (e.g. wear resistance). The result of this work is a set of parameters that are as robust against both noise factors as possible and that are adapted to certain changes by a neural network process control.

The use of numerical modeling in the development and optimization of thermal spray processes has been hampered by the complexities of the process, with up to 50 interdependent variables involved, and the need to identify and explicitly define each relationship. This work demonstrates an alternative modeling approach that employs neural network learning algorithms. Plasma spraying was selected as the test case because it presents the greatest challenge in terms of processing parameters and the magnitude of their effect on layer quality. This paper provides an introduction to neural networks and its use in thermal spraying. A companion paper in these same proceedings presents an example in which the method is used to link processing parameters with in-flight particle characteristics in a dc plasma jet. Paper includes a German-language abstract.

In current aircraft engines challenging coatings or coating systems with different functions are used. Many of these coatings are applied to different components with thermal spray processes. Thermal spraying is a very sensitive and very complex process, which is influenced by numerous controllable variables like the powder feed rate, the gas flow rates, etc. as well as not controllable variables like the torch wear, varying powder properties, etc. With conventional process control based on linear algorithms it is not possible to enduringly create constant coating properties, because they cannot describe the complexity of all influencing variables. In this work, the possibility of a closed loop was investigated exemplarily for an atmospheric plasma spray process (APS). During serial production a data base was collected, consisting information about torch and plume conditions as well as powder and coating properties. This data base was used to train different neural networks (NN). With regard to the automation of the APS, the NN obtained target values of relevant coating properties and should calculate the needed control variables. The result of this work shows the difficulties in the quantification of relevant influencing variables and the feasibility of the plasma spray process control with neural network.

Parametric drifts and fluctuations occur during plasma spraying. These drifts and fluctuations originate primarily from electrode wear and intrinsic plasma jet instabilities. One challenge is to control the manufacturing process by identifying the parameter interdependencies, correlations and individual effects on the in-flight particle characteristics. Such control is needed through methods that (i) consider the interdependencies that influence process variability and that also (ii) quantify the processing parameter-process response relationships. Artificial intelligence is proposed for thermal spray applications. The specific case of predicting plasma power parameters to manufacture grey alumina (Al 2 O 3 -TiO 2 , 13% by wt.) coatings was considered and the influence of the plasma spray process on the in-flight particle characteristics was investigated.

In this study, a neural network is used to model the complex relationships associated with dc plasma spraying. The paper describes how training and test data were experimentally obtained for alumina-titania particles processed under different conditions in order to assess various learning approaches and the predictive ability of the model. As reported, 80% of the test database was successfully recognized and the misclassifications are the result of not having enough data sets to adequately cover the wide range of values obtained during the experiments. This work considers in-flight particle characteristics as a first step of a more global approach that includes coating microstructure and mechanical properties as well. Paper includes a German-language abstract.

The very singular pore network architecture of thermal spray coating results from the combination of several phenomena occurring during the coating buildup process. Close pores, usually of large dimensions, result from stacking defaults when impinging particles spread onto previously deposited layers. Intra-lamellar cracks, perpendicular to the substrate surface, develop mostly in ceramic lamellae during solidification. Inter-lamellar cracks, parallel to the substrate surface, depend significantly from the surface tension characteristics during the particle spreading stage. One way or another, these characteristics are related to the processing parameters implemented to manufacture the coating and they significantly modify the coating characteristics, their cohesion, their compliance and their impermeability, among the most significant. Al 2 O 3 -TiO 2 (13% wt.) coatings were atmospheric plasma sprayed implementing several sets of processing parameters, among which power parameters (i.e., arc current intensity, plasmas gas flow rates, etc.), feedstock injection parameters (i.e., carrier gas flow rate, injector internal diameter, etc.) and environment parameters (i.e., spray angle, etc.) were varied. Pore contents were analyzed implementing image analysis. Pore network connectivity was analyzed implementing an electrochemical test: the higher the passivation potential of the substrate, the higher the coating pore network connectivity.

In an atmospheric plasma spray (APS) process, in-flight powder particle characteristics, such as the particle velocity and temperature, have significant influence on the coating formation. The nonlinear relationship between the input process parameters and in-flight particle characteristics is thus of paramount importance for coating properties design and quality control. It is also known that the ageing of torch electrodes affects this relationship. In recent years, machine learning algorithms have proven to be able to take into account such complex nonlinear interactions. This work illustrates the application of ensemble methods based on decision tree algorithms to evaluate and to predict in-flight particle temperature and velocity during an APS process considering torch electrodes ageing. Experiments were performed to record simultaneously the input process parameters, the in-flight powder particle characteristics and the electrodes usage time. Various spray durations were considered to emulate industrial coating spray production settings. Random forest and gradient boosting algorithms were used to rank and select the features for the APS process data recorded as the electrodes aged and the corresponding predictive models were compared. The time series aspect of the data will be examined.

Thermal spraying is a complex physical process consisting of three main sub-systems: flame/plume generation, powder/flame/plume interaction and coating build-up. While mathematical and CFD models provide valuable insight about the individual modules of a thermal spray process, it is very difficult to gain overall insight of the whole process and dependencies between different inputs and outputs using mathematical and CFD analysis, due to very complex and interconnected nature of the thermal spray process. In this work, a sophisticated experiment has been conducted to collect enough data for the sake of developing data-driven model of a plasma spray process. Metco 204 powder feedstock material and F4 gun have been used. An optimized number of data samples has been chosen by applying common industrial input parameters in the experiment. The developed neural network model is able to predict the coating quality parameters with acceptable average accuracy of above 90% on test data by considering all relevant measurement error deviations of the process analysis methods. A sophisticated user-interface has been developed to enable the use of the model for coating parameter development as well as the designing recipe for target coating characteristics. The developed model can be used for different purposes: parameter development, off-line coating quality control, and eventually adaptive coating control.

The reproducibility of the coating properties is the result of an efficient control of the thermal spray process. Actually, it is not possible for the sprayer to guarantee the temporal reproducibility of the coating properties without online control. However, adjusting the system fluctuations with regards to the temperature, velocity and size of particle could alter some properties that depend or not on the in-flight particle characteristics. This study aims at discussing the existing correlations between the processing parameters and the coating properties established with a Neural Network methodology. It demonstrates the segmentation of the correlations by comparing the result of the merging procedure of two sub-network structures and the direct correlation from the processing parameters to the coating properties. The sub-structures are built considering respectively the in-flight particle characteristics relationship with the operating conditions and the in-flight particle characteristics relationship with the coating properties.