Abstract
An intelligent approach, which uses adaptive network based fuzzy inference systems (ANFIS) based on experimental designs, is used to characterise the tribological behaviour of undoped and Zr doped diamond-like carbon (DLC) films that are deposited using magnetron sputtering. An orthogonal array experiment is used and the effect of the deposition parameters on the films is determined. The films are analysed using X-ray photoelectron spectroscopy (XPS) and scanning electron microscopy (SEM). Friction and wear tests are performed, using a pin-on-disk tribometer. This study identifies a group of highly developed hillock-like textures and lower wear volume loss is evident in the undoped and Zr doped films. The C1s core level XPS spectra show that the undoped and Zr doped films formed have a relative content of sp3 and sp2 hybrids. It is found that a value that is close to the estimated value, 0·5±0·05, of sp3/sp2 ratio results in better tribological properties in the undoped and Zr doped DLC films. These predicted values and the experimental results, for which an ANFIS predicts the tribological behaviour of the DLC films, are similar. The experimental results demonstrate that the tribological properties of DLC multilayer films are accurately predicted by an ANFIS. The results obtained for the ANFIS model are also compared to those for an ANN model and a Fuzzy system and it is shown that an ANFIS gives more reliable modelling of these sputtering processes and is more accurate and flexible than ANN and Fuzzy system models, which verifies the reliability and feasibility of this approach.
Keywords
Introduction
Hard coatings are widely used as protective surface coatings, to provide the requisite resistance for specific environments. They are essential for wear protection in tools that are used in conventional machining processes, such as cutting tools, cold forming tools, plastic injection moulds and tools for powder compaction.1–5 Several studies of wear failure have shown that the progressive deterioration of a metallic surface in a major industrial plant, due to wear, ultimately leads to a reduction in plant efficiency and, at worst, a shut-down. Wear damage to materials directly and indirectly costs the Unite States hundreds of billions of dollars annually.10–13 There are different types of tribosystems, such as coatings and surface treatments, and the wear that occurs in these tribosystems is reduced by the application of a coating. It is of the utmost importance that the effect of surface wear on materials is minimised and the most economical mean of wear protection is used, to offset the high costs associated with component replacement. The effects of wear can be prevented or repaired by the use of specific coatings. These are frequently used to reduce wear and friction in materials. The use of advanced deposition technology allows a great diversity of coatings, but no method allows the tribological behaviour of a coating to be tailored to a particular application.
Over the last 50 years, many processes have been developed to improve surface properties that use various advanced techniques, such as vacuum arc deposition, magnetron sputtering, pulse laser deposition and ion beam assisted deposition,14–19 all of which have been used to grow hard coatings during the last decade. Each of these techniques is designed for specific materials and for specific applications. Physical vapour deposition is widely used in industry as a protective coating for various tools. Magnetron sputtering is the most commonly used process for physical vapour deposition and has been extensively studied, because of its high deposition efficiency and low waste of materials. This has been demonstrated by many successful applications in industry over the last fifteen years. Magnetron sputtering provides a more efficient solution to many surface engineering problems and is used to improve the service life of various tools, moulds and machine parts. Interest in magnetron sputtering has been growing rapidly, because studies have shown that it produces remarkable mechanical, electrical, magnetic and optical properties.20–23
During the last decade, the number of new coating materials, structures, combinations and applications has increased exponentially. Various multilayer coatings have been produced for industrial production processes. Hard coatings use hard material that contains nitrides, carbides and borides of transition metals [such as TiN, CrN and (TiAl)N], carbon based coatings (such as diamond and diamond-like carbon), alumina and cubic boron nitride,4,5 all of which have been extensively studied and which are used to protect against wear, as reported in the literature.24–29 Diamond-like carbon and related materials can be used in a wide variety of applications that require wear protection and/or low friction. Many of these films/coatings have been found to display unique mechanical, physical and electronic properties, so they are potentially applicable in tribology, as well as in protective and decorative coatings. Some studies have reported that diamond-like carbon (DLC) coatings provide good protection from wear, because of their high wear resistance, extreme hardness, low friction coefficient and the chemical inertness of the coating. Many studies have examined the properties and tribological behaviour of the films and the wear behaviour of DLC films has been extensively studied. Recently, metal containing diamond-like carbon (Me-DLC) coatings have attracted increasing interest, because they provide high levels of protection against surface damage and because it is possible to synthesise materials that have unique physical–chemical properties. Because of this widely recognised potential, magnetron sputtering Me-DLC techniques and the resulting unique properties of the films that they produce have been studied extensively. It is noted that most of these studies concern DLC films, but very little study the prediction of the tribological properties of Me-DLC films. A sputtering process generally demonstrates nonlinear behaviour, because of the numerous variables and the stochastic nature of the process.6–9 Therefore, the use of traditional modelling methods to model mechanical properties does not ensure robust quality. Intelligent systems and optimization methods are used to increase quality. Of the intelligent systems, fuzzy logic, neural networks and genetic algorithms that are used to characterise various sputtering processes have attracted growing interest, because of their ability to produce adaptive predictions. Fuzzy logic systems allow uncertainty in modelling applications and artificial neural networks are particularly useful models for complex problems. This paper combines fuzzy logic and neural networks to control sputtering processes. An alternative method, a fuzzy inference system based on neural networks (ANFIS), is used to produce the desired reulst more efficiently for sputtered Me-DLC coatings. 6 ANFIS, as an effective alternative solution to this complicated problem and gives comparatively better results for the modelling of manufacturing processes than a traditional fuzzy inference system, which uses time consuming trial and error procedures to solve the basic problem using fuzzy ‘if-then’ rules, using artificial neural networks. This study determines the morphology, structure, bonding energy states, friction coefficient and wear behaviour of undoped and Zr doped DLC multilayer films. An ANFIS method for sputtering processes that uses learning procedures is developed, using a Taguchi experiment, which produces an accurate and reliable model for the prediction of tribological properties in undoped and Zr doped DLC films. This experiment gives an understanding of the effects of variables on undoped and Zr doped DLC multilayer films and simulates the tribological behaviour of the films, to produce better wear qualities.
Experimental details
A closed field unbalanced magnetron system was used. A sputtering magnetic field that supplies four rectangular targets (300×100×10 mm) was operated in an unbalanced mode. Three 99·5% pure targets: zirconium, chromium and titanium, were used for deposition. Multilayer films, which had interlayers of Zr\ZrC elements on the films, were used. The SKD11 substrate matrices used in the experiment had dimensions of 4×4×0·7 cm. The deposited surface morphology and the wear scars of the undoped and Zr doped DLC coatings were examined using a field emission scanning electron microscope (JEOL JSM-6700F) and their structures were characterised using X-ray photoelectron spectroscopy (XPS). The wear profile produced by the wear tester was recorded as numerical two-dimensional coordinates, using a TalyScan 150. The coatings on the SKD11 substrates were caused to slide against a WC ceramic ball. The friction and wear properties of the Zr doped DLC coatings were measured using a friction tester. The pin specimen was a tungsten carbide ball, with a diameter of 12 mm, the normal load was 2 N, with a sliding diameter of 20 mm, the rotational speed of the disk was 286 rev min−1, the sliding speed was 0·3 m s−1 and the length of the test was 1000 m. The friction coefficient and the sliding time were recorded automatically, during the test. The code and the levels for the control parameters for unbalanced magnetron sputtering are shown in Table 1. Most of the factors have three levels, but Factor A (Deposition Material) has two. The Taguchi method, which uses orthogonal arrays, was used to reduce the time and cost for the experiments. Orthogonal arrays for L18 tests were employed. There are three categories for the quality of the characteristics in the analysis of the S/N ratios: the lower the better, the higher the better and the nominal the better. These were used to measure the deviation in the quality of the characteristics from the desired values. An analysis of variance was also performed, to determine the process parameters that are statistically significant. These significant parameters were then used for an ANFIS model.
Control factors and their levels for Zr doped DLC films
Model analysis
The architecture for an Adaptive Neural-Fuzzy Inference System (ANFIS) was first proposed by Jang, in 1993, and has proven useful for modelling non-linear functions.
6
In ANFIS, an artificial Neural Network (ANN) is capable of self-learning and a fuzzy logic inference system deals with fuzzy language information and simulates judgment and decision making in the human brain. A training algorithm is used for artificial neural networks, to tune rule based fuzzy systems that approximate the methods by which humans process information. These can be used to solve many complicated engineering problems in various fields. This system not only learns adaptively, but also describes and processes fuzzy information and demonstrates judgment and decision making. An ANFIS is an example of a readily available system that uses an ANN for fuzzification, fuzzy inference and defuzzification, in a fuzzy system. An ANFIS uses Gaussian functions for several fuzzy sets, linear functions for the rule outputs and Sugeno's inference mechanism. Depending on the structure of the If-then rules, as in the Mamdani and the Takagi–Sugeno model, a set of fuzzy rules and a first order function are used. The rule base contains k fuzzy Takagi and Sugeno type if-then rules, as follows:
is the linguistic label of the fuzzy set (
),…, (
), and
represents the model outputs of the Sugeno fuzzy inference system. Using this hybrid algorithm, a gradient descent is used to tune the antecedent membership functional parameters (
) and the least squares method is used to identify the consequent linear parameters (
). As shown in Fig. 1, a fuzzy layer, a rule based layer, a normalisation layer, an inference layer and a total output layer. Each layer has several nodes, each of which is described by the node function. The procedures for the ANFIS algorithms are:
Frameworks of ANFIS with five layers including inputs, rules, normalisation, inference and outputs
Layer 1: every node j in this layer is a square node with a node function. It is a fuzzy layer, in which
are the inputs of nodes.
are linguistic labels used in fuzzy theory to divid the membership function
is the linguistic label (small, medium, or large) that is associated with this node function and
is the membership function of
. In each membership function, the parameters in a layer are defined as premise parameters. This study uses a Gaussian shaped function as the input for the nodes. A Gaussian shaped function is specified by two parameters (μ, σ) as follows
Layer 2: circled with by π labelled nodes, this layer multiplies the incoming signals and outputs the product. This represents the firing strength of a rule
Layer 3: every node in this layer, labelled N, calculates the average ratio of the ith rule's firing strength
Layer 4: every node I that has a square node with a node function in this layer is utilised
is the output of layer 3 and the parameters,
, are referred to as consequent parameters.
Layer 5: the node in this layer, marked by a circle and labelled Σ node, computes the overall output as the sum of all of the incoming signals
The error function E which is defined as the sum of the squares of the differences between the desired value, dj and the adaptive fuzzy system outputs
, is given by
During the training process, an ANFIS dynamically adjusts these parameters using back propagation, whereby the antecedent parameter set (μ, σ) of the fuzzy model is tuned according to the error function. This gives the fuzzy system a neural learning capability, which makes it more adaptive. As a result, the network can accurately describe the mapping between the input and the output data. An ANFIS that models the anti-wear properties of undoped and Zr doped DLC films is developed that uses a hybrid learning approach to optimise the parameters of the adaptive network. This gives a good a measure of the accuracy with which the fuzzy inference system models the input/output data for a given set of parameters that are used for DLC coatings.
Results and discussion
Wear performance analysis
Table 2 shows the overall experimental layout and the results. The wear volume values for each experimental test are calculated using S/N ratios, with an average and a deviation. Most of the tests on the Zr doped DLC\CZr\Zr films, which give an average value of 3·025×10−2 mm3, have higher S/N ratios than those for the undoped DLC\CZr\Zr films, which have an average of 3·256×10−2 mm3. The complete wear volume values range from 1·355×10−2 to 4·902×10−2 mm3. The results show that the Zr doped DLC coatings demonstrate an obvious improvement, in terms of the friction coefficient and wear volume values, over undoped DLC coatings. A higher average coefficient of friction of 0·328±0·119 is also observed for undoped DLC coatings, but the average coefficient of friction for Zr doped DLC coatings is lower, at 0·290±0·126, for a normal load of 2N. In the study by Lubwama et al.,32 the average coefficient of friction for undoped DLC coating is cited as 0·6, for a normal load of 0·1 N. The decrease in the wear volume values for DLC films, for larger S/N ratios, seen in Table 2, is also seen in Fig. 2. It is worthy of note that the L2 trial for an undoped DLC\ZrC\Zr film and the L14 trial for a Zr doped DLC\ZrC\Zr film show larger S/N ratios of 37·35db and 36·29db respectively, which shows good anti-wear properties in all cases. The higher the S/N ratio, the lower is the wear volume value and the smaller is the deviation. In Fig. 6, a high resolution C1s spectra with sp3 and sp2 hybrids is seen. A comparison of the sp3/sp2 ratio with the wear volume values shows that a sp3/sp2 ratio of nearly 0·55±0·05 results in good anti-wear performance for the undoped and Zr doped DLC films, because the desirable parts of the adherent sp2-rich phases dominate the slippery transfer formations in the undoped and Zr doped DLC films. This figure clearly shows that the sp3/sp2 ratio shifts to a higher or lower value, which results in poor anti-wear performance. This is reasonable evidence that undoped and Zr doped DLC films with a sp3 rich or a sp2 rich phase form a precursor to diamond or graphite, which results in poorer anti-wear properties after sputtering.
Comparison of between SNR, wear volume and SP3/SP2 ratios for undoped and Zr doped DLC films in the orthogonal arrays
Response plot for S/N ratios of wear volume
control factors and their levels for Zr doped DLC films in orthogonal arrays
XPS analysis
Further details of the chemical binding and the elemental composition of undoped and Zr-doped DLC films are shown in Fig. 4. The film composition in L2, L5, L14 and L14 tests was determined by XPS spectra, including the Zr 3d, Ar 2p3/2, N 1s, C 1s and O 1s peaks in the undoped and Zr-doped DLC films. The existing data4,5 show that the Zr 3d5/2 peaks at 180 eV and the Ar 2p3/2 peak at 239 eV are associated with DLC bonded coatings, but the peaks are not apparent, while the C1s peaks at 286 eV are seen and a sharp peak in the C1s region is noted. The O1s peaks at 534 eV are attributed to film contamination, following exposure to air, and the N1s peaks at 400 eV are clearly present for the DLC coatings. As shown in Fig. 4, the N1s peak becomes smaller in the L5 bonds, but the intensity of the C1s increases in the L2 bonds. According to the existing data, the C 1s peaks at 285 eV are detected and all of the peaks are strong in the C1s spectra, which significantly affects the performance of undoped DLC films. As shown in Figure 5, the C1s core XPS spectra for undoped and Zr-doped DLC films is divided into two phases, with patterns of sp3 and sp2 peaks. Figure 5a shows that the broad C1s spectra in the L2 films are composed of the SP3 bonding peak area of 874 at 287·359 eV carbons and the SP2 bonding peak area of 1552 at 286·083 eV carbons, which correspond to the sp2 and sp3 bonded carbons, respectively. The estimated value of the sp3/sp2 ratio is 0·563, which results in good anti-wear behaviour in the overall tests. Figure 5b shows that the estimated sp3/sp2 ratio, the integral area under its curve, is 0·415 in the L5 test, which results in the worst tribological properties. The SP3 bonding peak area of 607 at 287·166 eV carbons and the SP2 bonding peak area of 1462 at 286·263 eV carbons are shown. In addition, Figure 5c shows that the broad C1s spectra in the L14 films are composed of the SP3 bonding peak area of 881 at 287·363 eV carbons and the SP2 bonding peak area of 1554 at 286·084 eV carbons, which correspond to the sp2 and sp3 bonded carbons, respectively. The estimated value of the sp3/sp2 ratio is 0·567, which is the same as that for Fig. 5a, and results in relatively good anti-wear behaviour. However, Fig. 5d shows that the estimated sp3/sp2 ratio, which is the integral area under its curve, is 0·810 in the L15 test, which results in the worst tribological properties, similarly in Fig. 5a. The SP3 bonding peak area of 936 at 286·910 eV carbons and the SP2 bonding peak area of 1163 at 285·420 eV carbons are evident.
Comparison of selecting special five tests between L2, L5, L14, L15 and optimal trial
X-ray photoelectron spectroscopy for wide scan energy spectrum detailed between L2, L5, L14 and L15 undoped and Zr doped DLC films
X-ray photoelectron spectroscopy with sp2/sp3 level of C1s spectrum detailed on undoped and Zr doped DLC films
Tribological properties
Table 2 shows the L18 experimental results, in which the anti-wear performance of each experimental test of undoped and Zr doped DLC films is evaluated by computing their S/N ratios. The experimental results show considerably different anti-wear properties and a corresponding lower wear volume loss, during the wear test, compared with the substrate,4–6 which indicates good wear resistance. The results for the frictional behaviour after tribo-tests for sliding are shown in Fig. 7. The typical variation in the wear track is similar to that for other tests. The undoped and Zr doped DLC films are worn to a slight and a severe degree respectively. The wear track for the slight wear tests is not wrinkled when viewed with a looking glass, but the severe wear tests are rough, with obvious groove microfracture features, despite the microcracks and plastic deformation. Figures 7a–c shows that the width of the wear track is relatively small and the transfer film with a large area of zirconium oxides comes together. This transferred film appears to be a lubricating layer and seems to be more effective in improving the anti-wear performance of undoped and Zr doped DLC films. The wear volume of the two samples is similar to that for glassy carbon. It is relatively low in the undoped and Zr doped DLC coatings and the wear scars, with widths of 0·85 and 0·91 mm, have extremely small values of 1·355±0·063×10−2 and 1·534±0·148×10−2 mm3, which corresponds to S/N ratios of 37·35db and 36·26db respectively, during sliding. As shown in Fig. 7b–d, a few slight dimples and long narrow furrows form on the wear profile of the wear tracks on the films. The wear scars, with widths of 1·6 and 1·8 mm, have a relatively large value of 4·902±0·162×10−2 and 4·837±0·532×10−2 mm3, which corresponds to lower S/N ratios of 26·19db and 26·27db respectively, during sliding. This wear profile appears to be mainly abrasive wear from plowing and grooving, as seen in the OM images. Further details of the wear profile of the scar tracks are shown in Fig. 7e, where the wear profile of the scar tracks is computed using an α-step profiler. The degree or amount by which the worn profiles differ is found for both wear tests with undoped and Zr doped DLC films. The worn tracked curves in the slight wear tests that are obtained by this study have ear scar features that irregularly fluctuate, but the worn tracked curves in the severe wear tests are more undulated and have a comparatively rough texture. In addition, the difference in the friction coefficient of undoped and Zr doped DLC films, compared to that for tungsten carbide, are shown in Fig. 7f. In this waterfall figure, the differences in the friction coefficient for the L2, L5, L14 and L15 tests are compared. Few parts of the wear scar show an extreme condition in the slight wear test for the Zr doped DLC films, which results in an understandable reduction in friction. The value of the friction coefficient is much lower than that of the undoped DLC films. A comparison of the S/N ratios shows that an increase in the wear volume is clear for the undoped and Zr doped DLC films with a larger frictional coefficient. As previously mentioned, the slight wear tests for undoped and Zr doped DLC films demonstrate better wear behaviour and a lower friction coefficient, but the severe wear tests demonstrate higher wear volume for the undoped and Zr doped DLC films, which have wavelike variations in the frictional coefficient curves, within specified limits, which indicates poor tribological behaviour.
Variations of a–d worn scar surfaced SEM images, e cross-section curves of worn scar and f friction coefficient during sliding
Optimisation
In order to estimate the performance of the films used in the experiments, each test was replicated three times, in different areas of the wear tracks. The results and the S/N ratios, using a formula that fits smaller-the-better properties, are shown in Table 2. The response plots for the S/N rations for the orthogonal array experiments are shown in Fig. 3. Clearly, the larger S/N ratio indicates better performance for the Zr doped DLC film. The larger the S/N ratio, the more significant is the factor, so the optimal parameters are determined in accordance with the effect of the S/N ratio, so the undoped and Zr doped DLC films can be clearly distinguished. The optimal parameters for the sputtering process are: a substrate bias of −40 V, a C/Zr target current of 4 A, a N2/Ar mixture of 0·14 sccm, a CH4 flowrate of 2 sccm, a sputtering distance of 10 cm, a sputtering time of 30 min and a pulse frequency of 100 kHz. The confirmation experiments for the optimal settings for all of the orthogonal array experiments are compared. Figure 4 shows that four tests: L2, L5, L14 and L15, were performed under specific conditions, such as smaller or larger S/N ratios, for both the undoped and Zr doped films, in order to compare the results with those for the optimal tests. In Fig. 4, The graph of a Gaussian is used and the thinnest bold solid curve indicates that the optimal test produces the least wear volume loss, with only a little deviation. It is clear that the optimal settings for the control factors are obviously resistant to noise, which indicates good reproducibility. An analysis of variance (ANOVA) was also performed using the S/N ratios, to determine the effects of the factors, and this was used to determine the factors that influence the quality of the films. The ANVOA separated the total variability of the wear volume values, which is measured as the sum of the squared deviations from the total mean of the wear volume values, into the contributions of each of the factors and the error. Table 3 lists the percentage contribution of each of the factors to the total sum of the squared deviations and this is used to evaluate the effect of the factor on the wear response. The results of the ANOVA indicate that at least four parameters: the C and Zr target current, the N2/Ar mixture, the sputtering distance and the pulse frequency, have the most significant effect on the wear performance of the undoped and Zr doped DLC films. These significant factors explain nearly 80·87% of the experimental variation, so they were selected for further modelling with the ANFIS.
Variance analysis of wear volume values in orthogonal array experiments
Simulation results
In order to understand the performance of the ANFIS, the significant parameters from the analysis of variance for the undoped and Zr doped DLC films, as listed in Table 3, were assigned to the adaptive network based fuzzy systems that were modelled. Figure 8 shows the results for an ANFIS, using a subtractive clustering method. Clearly, subtractive fuzzy clustering significantly reduces the number of rules. As shown in Table 4, the data from the L19 to L28 tests was also tested to identify the predictive accuracy of the ANFIS model. Using the fuzzy training system previously mentioned, the parameters associated with the membership functions are tuned in the learning process and the fuzzy logic system is optimized. For an ANFIS, the optimal curves for the RMS errors for training data and test data for the 180 learning epochs are 6·644×10−6 and 1·031×10−5 respectively. The model performance is validated using root mean square error (RMS). A low RMS value indicates a better model. The ANFIS predictor also fits the percentage errors for the actual data. The ANN, the Fuzzy system and the ANFIS predictors use the same factorial settings. For less tests, the errors are closely grouped around 6%, with a maximum error of 6·07%. However, the distribution of the predictive error fluctuates unsteadily for both the fuzzy system and the ANN predictor, where many errors are more than 10%. In other words, the ANFIS predictor produces an average error of 2·72, but both the ANN and the Fuzzy system predictors generate average errors of 10·19% and 11·04 respectively. Test predictions and comparisons for the experimental data are shown in Fig. 9, which compares the ANFIS, the ANN and the Fuzzy system predictors. An ANFIS predictor produces an average error of 6·02, but both the ANN and the Fuzzy system predictors generate average errors of 17·32% and 19·52 respectively. It is clear that the ANFIS predictor produces predictions that are more accurate than those of the other two methods: the ANN and the Fuzzy system. If the experimental results are compared with the predicted values, the RMSE values for the ANFIS model are lower, which indicates a better performance than that of either the ANN or the Fuzzy system models. As a result, the ANFIS model developed is an effective model for the analysis of the tribological properties of undoped and Zr doped DLC coatings by decision makers. The effect of the sputtering parameters on the anti-wear properties of the films was also studied with respect to the spraying distance, the pulse frequency, the C/Zr target current and the flowrate of the N2/Ar mixture, using the significant variables for each response. Equation (1) is plotted in Fig. 10 as the contour plots for the ANFIS model, for the four significant variables that affect wear volume losses if the other variables are optimal, using the wear volume losses as the response. Figure 10a shows that the wear volume loss increases rapidly as the spraying distance increases to 11 cm, when the pulse frequency is lower (<100 kHz). The wear volume loss at a lower pulse frequency is also larger when there are changes made to the carrier gas, as shown in Fig. 10b. An increase in the sputtering distance results in a decrease in the pulse frequency, because the gas is weakly ionised when it crosses the systems. This result demonstrates that the lowest wear volume losses occur when there is a short sputtering distance and a high pulse frequency. At this spraying distance the system requires less time to deposit, which produces coarse coatings. Several similar counter plots are also seen in Fig. 10c–e. The corresponding contours show a considerable curvature, which demonstrates that the larger wear volume losses are exactly located inside the design boundary. The wear volume loss increases from 2·0×10−2 to 5·2×10−2 mm3. By comparison, a change in the C/Zr target current does not result in any change in the wear volume losses when the flow rate of N2/Ar mixture is medium (4A)(Fig. 10c), or when the sputtering distance is close to 10 cm (Fig. 10f). Figure 10d–f shows that the middle area of the wear volume does not vary if the pulse frequency is decreased, when the accelerating voltage is high (>70 V) and when the flowrate for the carrier gas is low (5 L min−1). Figure 10d shows that the wear volume losses are reduced if the pulse frequency is high and the flowrate of the N2/Ar mixture is high. Multiple phases are seen in the films, which lead to the generation of denser and more homogeneous coatings. When the C/Zr target current and the sputtering distance have their maximum values, Fig. 10f shows that a higher C/Zr target current leads to a thicker film, which can move a larger part of the massive particles because the adatoms have a higher impact mobility. This increases the wear volume losses. As a result, coarser coatings with thick surface films are produced. The figures clearly show that it is easy to predict the relationship between the responses and the design variables and to determine the nature of the stationary area and to gain other information regarding the nature of a sputtering system.
RMS errors of training and testing algorithms with learning epochs of 180 for ANFIS model
Comparisons between experimental data and predicted values in ANN, Fuzzy system and ANFIS models for wear volume of undoped and Zr doped DLC films
Response surface pictures of tribological behaviours of DLC films in ANFIS model
Errors between testing algorithms for tribological behaviours in ANN, Fuzzy system and ANFIS models
Conclusion
The tribological behaviour of undoped and Zr doped DLC films is studied, using magnetron sputtering processes. During the wear test, the Zr doped DLC films exhibit relatively lower friction and wear volume values and demonstrate excellent wear resistance. A slightly worn surface with a glassy carbon phase is apparent in the undoped and Zr doped DLC films with a higher S/N ratio, but severely worn films with slightly narrow furrows are clearly visible in the undoped and Zr doped DLC films with a lower S/N ratio. Futhermore, a comparison of the results obtained by analysis of the XPS spectra for undoped and Zr doped DLC coatings shows that the estimated value of 0·55 for the sp3/sp2 ratio of carbon bonds results in better wear resistance properties. In addition, the most significant factors for the tribological behaviour of undoped and Zr doped DLC coatings are identified using an analysis of variance. The C and Zr target current, the N2/Ar mixture, the sputtering distance and the pulse frequency account for nearly 80·87% of the experimental variation. In the ANFIS, the predicted and the observed values are clearly close, which indicates that the model is a good predictor, within the experimental limits. Also, the proposed ANFIS model is also compared with an ANN and a Fuzzy system. The results show that the ANFIS model gives a more accurate and reliable prediction of the tribological properties of undoped and Zr doped DLC coatings than the ANN or the Fuzzy system. Accordingly, the ANFIS is proven to be a reliable and feasible method for forecasting the tribological properties of undoped and Zr doped DLC films.
Footnotes
Acknowledgements
This work is based upon work that was supported by grant no. NSC: 102-2622-E-132-001-CC3 and by MEA: 101-2511-S-132-002, Taiwan.