*3.3. Manufacturing*

ML approaches were also employed in the area of manufacturing technology, for example, for process monitoring or in quality control/image recognition [6]. There are also some studies related to tribology, particularly regarding friction stir welding [93–95], but

also for forming or machining [96]. Sathiya et al. [97] modeled the relationship between friction welding process parameters as heating pressure, heating time, upsetting pressure as well as upsetting time and output parameters (tensile strength and metal loss) when joining similar stainless steel by means of a back propagation ANN with 9 neurons in the single-hidden layers. The database (14 data points) was generated from corresponding experiments. Subsequently, different optimization strategizes based upon the ANN's prediction were compared: Genetic algorithm, simulated annealing algorithm, and particle swarm optimization. Among them, the genetic algorithm was reported to be most suitable and good agreement was found between the prediction of tensile strength and metal loss for optimized process parameters with respective validation experiments. Similarly, Tansel et al. [98] and Atharifar [99] applied and confirmed the suitability of ANNs for optimizing friction stir welding processes. However, the latter further introduced an optimization of the back propagation ANNs using a genetic algorithm (genetically optimized neural network systems) to maximize the prediction quality. Anand et al. [100] compared the performance of an ANN (4:9:2) with a response surface methodology approach (quadratic polynomial models) when optimizing friction welding with respect to tensile strength and burn-off length. The data (30 data sets) were generated with experiments within a five-level, four variable centrale composite DoE (CCD). It was observed that the ANN featured higher accuracy by a factor of two compared to the response surface. In turn, Dewan et al. [101] compared back propagation neural networks with adaptive neuro-fuzzy interference systems (ANFIS) [102] when predicting tensile properties in dependency of spindle speed, plunge force and welding speed from a rather small database (73 data points). Here, 1200 different ANFIS models were developed with varying number and type of membership functions as well as input combinations. It was reported the optimized ANFIS provided lower prediction errors than the ANN.

In addition to process optimization, ML approaches have also been used for monitoring friction stir welding. Baraka et al. [103] made use of process signals (traverse and downward tool force) to predict the weld quality. This was based upon frequency analysis by FFT, and an interval type 2 radial base function (RBF) neural network trained by an adaptive error propagation algorithm that effectively provided continuous feedback to the operator with an accuracy above 80%. Das et al. [104] also used real-time process signals (torque) for internal defect identification in friction welding. The experimentally obtained signals were analyzed by discrete wavelet transformation, statistical features (dispersion, asymmetry, excess) as well as general regression models and ML methods, namely SVM and back propagation ANN (3:5:1, log-sigmoid transfer functions) trained by the gradient descent method to predict tensile strength. It was reported the prediction performance of the SVM (0.5% error) was superior to regression (13.6%) and the ANN (3.1%).

Regarding other manufacturing processes, Fereshteh-Saniee et al. [105] trained a feedforward back propagation ANN with 21 neurons in the single hidden-layer (tansigmoid transfer function) from over 700 FE simulations to determine material flow and friction factors in one-step ring forming. Thereby, obtained load curves showed good agreement with experimental validation tests, featuring an accuracy of 99% and 97% for grease lubricated and dry conditions, respectively. The difference was traced back to higher variations of friction for unlubricated forming. Furthermore, Bustillo et al. [106] attempted to predict surface roughness and mass loss during turning, grinding, or electric discharge machining based upon surface isotropy levels and different ML approaches: Artificial regression trees, multilayer perceptions (MLP), RBF networks, and random forest. The most accurate approach for predicting the loss of mass was found to be RBF, while the MLP most precisely predicted the arithmetic mean roughness. However, the model parameters of both approaches had to be tuned very carefully and even small changes led to a substantial increase of errors. In contrast, satisfactory accuracy without any tuning stage could be obtained using the random forest ensembles. It was also reported that the prediction quality was comparatively sound even outside the training record as well as for smaller data sets.

The works in the area of manufacturing technology are summarized in Table 3 according to the subject, the database, the inputs and outputs, and the ML approach.

**Table 3.** Overview of ML approaches successfully applied in the area of manufacturing technology.


#### *3.4. Surface Engineering*

Approaches to enhance the tribological behavior of components by modifying their surfaces can be subsumed under the term surface engineering [107]. This involves adjusting the surface topography with and without compositional changes through as well as the application of coatings. Examples include, among others, tailoring the roughness and/or statistically distributed or discrete micro-textures, carburizing, nitriding, anodizing, electroplating, weld hardfacing, thermal spraying, chemical, or physical vapor deposition (CVD, PVD) [107]. Some studies have also applied ML approaches to better understand or design the surface modifications.

#### 3.4.1. Coatings

Cetinel [108] used a single-hidden layer feed forward ANN to predict the COF and wear loss of thermally sprayed aluminum titanium oxide (Al2O3-TiO2) coatings. The database was created by reciprocal pin-on-block tribometer tests under dry as well as acid conditions different loads. In the ANN, the test conditions were the inputs and—after

trial-and-error testing of different configurations—the hidden layer consisted of 80 neurons. Furthermore, the ANN provided 63 outputs in the form of the COF and linear wear progress at different times of the experiments. Thus, the tribological behavior over the test period could be mapped very well. Sahraoui et al. [109] analyzed the friction and wear behavior of high-velocity oxy-fuel (HVOF) sprayed Cr-C-Ni-Cr and WC-Co coatings as well as electroplated hard chromium by means of a feed forward ANN. The database consisted of 180 training and 180 test data sets from dry-running pin-on-disk tribometer tests of the coated test specimens against brass disks at various normal loads and sliding speeds. An ANN with sigmoid transfer functions as well as two hidden layers (6 and 4 neurons) was found to be suitable for predicting the COF within variabilities between 5.8% and 10.8%. The main advantage of the model in this study was that the friction coefficients could be predicted comparatively well for a range of parameters up to 7 times larger than those contained in the training data. Upadhyay and Kumaraswamidhas [110,111] applied a back propagation ANN to optimize multilayer nitride coatings on tool steel deposited by unbalanced reactive PVD magnetron sputtering. The input parameters comprised bias voltage and gas flow rate as well as time, velocity and load within pin-on-disk sliding tests. The data was split into 70% training, 15% validation, and 15% test data. Training was based on the LM function and the most favorable ANN consisted of 20 neurons in the hidden layer. Thus, the wear rate as well as the COF could be predicted within errors of less than 10%.

#### 3.4.2. Surface Texturing

Otero et al. [112] attempted to optimize surface micro-textures fabricated by photolithography and chemical etching processes in order to reduce the COF of EHL contacts by means of an ANN. The data was obtained from tests on a mini-traction machine (steel ball-on-micro-textured copper disk) at various loads, total speeds and slip conditions. The ANN consisted of 7 inputs (average velocity, SRR, load, minor and major axis dimensions, depth and texturing density), 20 neurons in the hidden layer, and the COF as output. Thus, load case-dependent ranges for beneficial texture parameters could be derived. Additionally, referring to tests on samples with pores or micro-textures on a lubricated mini traction machine at different test conditions, Boidi et al. [113] applied an RBF to predict the wear behavior of sintered components. The database included 1704 experimental sets with different sum velocities and slip, as well as geometric or statistical characteristics of the dimples, grooves and pores, respectively. A Hardy multiquadric RBF was found to provide an excellent fit with an overall correlation of 0.93, especially with regard to the standard deviations of the tribological experiments. Mo et al. [114] utilized statistical methods as well as a back propagation ANN with 60 neurons in the single-hidden layer to investigate the role of micro-texture shape deviations and dimensional uncertainties on the tribological performance. The database was founded on physical modeling approaches in the form of simulations of parallel, hydrodynamically lubricated (HL) contacts and randomly split into 70% training and 30% validation data. The trained ANN was able to predict the relationships between geometric micro-texture parameters (e.g., dimple diameter, depth, area density etc.) and the frictional force as well as the load carrying capacity with an accuracy of 99.7% and 97.5%, respectively. Thus, the influences of statistical deviations (e.g., roundness errors, standard deviations of the dimensional parameters, etc.) could be estimated and optimal, robust optima could be retrieved by means of a genetic algorithm. Similarly, Marian et al. [115,116] utilized a MOP [74] to model the influence of micro-textures in EHL contacts as well as an EA to optimize the micro-texture geometry and distribution. Based upon a LHS (70 data sets) and contact simulations, the contact pressure, lubricant film height, and frictional force were predicted with CoPs larger than 82%, allowing subsequent optimization with an EA. Zambrano et al. [117] used reduced order modeling (ROM) to predict and optimize the frictional behavior of surface textures in dynamic rubber applications under different operating conditions. It is noteworthy that this was based on a limited number of experimental measurements and the ROM was

fed with microscope-based texture measurements. In this sense, besides nominal texture parameters, the real geometries as well as their deviations and uncertainties have been evaluated with good accuracy.

The works in the area of surface engineering are summarized in Table 4 according to the subject, the database, the inputs and outputs, and the ML approach.


