MODELLING OF MICROHARDNESS VALUES BY MEANS OF ARTIFICIAL NEURAL NETWORKS OF AL/SiCP METAL MATRIX COMPOSITE MATERIAL COUPLES PROCESSED WITH DIFFUSION METHOD

Abstract-In this study, modelling of microhardness values by means of artificial neural networks of Al/SiCp metal matrix composite material couples with diffusion method and manufactured by powder metallurgy process, were obtained using a backpropagation neural network that uses gradient descent learning algorithm. After diffusion bonding and relevant test, to prepare the training and test (checking) set of the network, results were recorded in a file on a computer. Then, the neural network was trained using the prepared training set. At the end of the training process, the test data were used to check the system accuracy. As a result the neural network was found successful in the prediction of modelling of microhardness values of Al/SiCp metal matrix composite material couples processed with diffusion method and behavior.


INTRODUCTION
Aluminum-based, particulate-reinforced metal matrix composites (MMCs) are of concerns for structural carrying outs where weight saving is of primary concern. There are several production techniques to getting in manufacturing the MMC materials. An artificial neural network is a parallel-dispensed information proceeding system. It stores the specimens with dispensed coding, thus forming a trainable nonlinear system. The main idea of neural network draw near to resembles the human brain functioning. Given the inputs and longing outputs, it is also self-adaptive to the habitat so as to respond different inputs rationally. The neural network theory deals with learning from the preceding obtained data, which is named as training or learning set, and then to check the system accomplishment using test data [1][2]. Artificial Neural networks (ANNs) have been used to model the human vision system. They are biologically inspired and contain a large number of simple processing elements that perform in a manner analogous to the most elementary functions of neurons. Artificial neural networks learn by experience, generalize from previous experiences to new ones, and can make decisions. Neural elements of a human brain have a computing speed of a few milliseconds, whereas the computing speed of electronic circuits is on the order of microseconds. The ANNs are parallel process elements which has characteristic in below.
-ANN is a mathematical model of a biological neuron.
-ANN has very process elements which are related another.
-ANN keeps knowledge with connection weights.
Neural network models provide an alternative approach to implementing enhancement techniques. A simple process element of the NNs is given in Fig.1. Output of i th process element at this simple model is given at Equation 1.
In there, a is activation function, θ i is threshold value of i th process element. Knowledge processes of process element compose from two parts: input and output. Output of i th process element is calculated with Equation 2 [3][4]. Neural networks were procured a basically different draw near to material modeling and material processing control techniques than statistical or numerical procedures. This method is feasible in many areas of engineering and has produced promising to prepare results in the areas of material modeling and proceduring. One of the main advantages of this approach is that there is no need to make a priori assumptions about material behavior although in more were sophisticated neural network modeling projects one may take advantage of the information of the procedure in network design. Even though multi-layered neural network models cannot make sure a global minimum solution for any given problem, it is a sensible assumpttions that if the network is trained on a extensive database with a suitable representation project, the resulting model will approximate all of the laws of mechanics that the actual material or process obeys [5][6].
Neural networks are essentially connectionist system, in which different nodes called neurons are interconnected. A typical neuron accepts one or more input signals and procures an output signal trusting in the proceduring function of the neuron. This output is conveyed to connected neurons in varying intensities, the signal intensity being decided by the weights. Feed forward networks are jointly used. A feed forward network has a consecutive of layers consisting of a number of neurons in each layer. The output of neurons of one layer come to exists input to neurons of the achieving layer. The first layer, called an input layer, accepts data from the outside world. The last layer is the output layer, which sends knowledge out to users. Layers that lie between the input and output layers are called hidden layers and have no direct touch with the environment. Their presence is needed in order to procure complexity to network architecture for modeling non-linear functional kinship. After choosing the network architecture, the network is tested by using back propagation algorithm, where back propagation algorithm is the productive optimization method used for underrating the error through weight arrangement The trained neural network has to be experimented by supplying testing data [7].
The basic fundamentals to build the system model on the basis of NN consist of: (a) connecting the artificial neurons into a network with respect to certain rules and a topology; (b) regulating the weights between neurons in term of an proofreading criterion; (c) establishing the topology and free parameters of the NN by learning specimen data (input patterns) repeatedly; (d) determining the system model by taking advantage of the strong learning ability of ANN (Fig. 2). The information included in the illustration data was acquired via the improved back propagation (BP) learning algorithm. The parameters of the BP network were defined as follows: The input vectors [X = x0, x1, . . . , xn1] T The output vectors [Y = y0, y1, . . . , ym1] T where the symbols n, h and m represented the number of neurons in the input layer, the hidden layer and the output layer, sequentially.
Joining of the powder metallurgy products (P/M) by diffusion bonding process is important both to protect the microstructural properties of parent materials and bonding behavior of joining materials [9]. Diffusion bonding is a solid state coalescence of contacting surfaces occur at a temperature below the melting point (T m ) of the materials to be joined with the loads and the period, below those that would cause macro deformation and a significant properties change at the parent materials [10]. The process is depended on a number of parameters in particular, bonding temperature, atmosphere, time, pressure and surface roughness. Process pressure is selected high enough to dislocate the surface oxides. Bonding period should be selected long enough for the completion of the diffusion mechanism at the interface. Diffusion bonding is an advanced bonding process in which two materials, similar or dissimilar, can be bonded in solid state. Process temperature is selected as 0.5-0.7 (T m ). Super plasticity may occur at low temperatures [11-12].

EXPERIMENTAL STUDIES
SiC particulate Al alloy MMCs specimens to be produced by hot pressing method were fabricated by powder metallurgy process. 1 % Mg, 3 % Si powders were mixed with 99 % Al. SiC particles with a 42 µm mean diameter were added to the matrix at 5, 10, 20 (wt) % fractions. Powders were properly mixed with mechanic mixers for homogeneity of the formation. The mixture was cold compacted at 400 MPa in the φ12x60 mm steel dies. This is followed by sintering at 600 ºC in argon atmosphere for 30 minutes. Finally the specimens were hot compacted and extruded at 200 MPa pressure. Mixture rations and density of specimens were presented in Table 1. Work pieces were prepared for diffusion bonding and surfaces to be joined were protected against corrosion and oxidation. Al alloy MMC specimens with 5-5, 5-10, 5-20, 10-10, 10-20 ve 20-20 % SiC (wt) fractions were coupled and bonded at diffusion bonding apparatus. Schematic illustration of diffusion bonding apparatus is given in  After the bonding process, specimens were tested for microhardness. Specimens were cut perpendicular to the bonding interface to facilitate longitudinal microstructure cross section examinations. Micro hardness values were measured at interface and it's both side by Hv scale under a load of 100 gr. (Fig. 4). Grinding of the surface were followed by etching with Keller etchant. Metallographic evaluations and investigations were made by the aid of optical microscopy and SEM.

RESULTS AND DISCUSSION
Deformation of surface asperities by plastic flow and creep, grain boundary diffusion of atoms to the voids and grain boundary migration, volume diffusion of atoms to voids can be listed as a sequence of metallurgical stages of the diffusion bonding. Especially with aluminum alloys diffusion bonding can be achieved with adherent surface oxides. In general, the oxide is not removed, but is simply dispersed over a greater surface area in an enclosed environment, in which oxidation cannot recur.
At elevated temperatures diffusion mechanism were accelerated and diffusion period were decreased to achieve the same coalescence. Relatively poor coalescence were achieved at 550-575 ºC process temperatures. Voids were traced at the bond interface. SEM micrographs of weld specimens bonded at 550-575 ºC process temperatures presented at Fig.6 and 7.  temperatures. SEM micrographs of weld specimens bonded at 600-625 ºC process temperatures were presented at Fig.8 and 9.  In this study, prediction of microhardness of diffusion bonded MMC couples were performed by using a back-propagation neural network that uses gradient descent learning algorithm. a-Bonding process temperature, bonding period and SiC particulate (wt) fractions were used as the model inputs while the microhardness was the output of the model. These datas were obtained from experimental works. b-Comparison of experimental microhardness test results with predicted values inline with bonding parameters were presented in Table 3. Experimental microhardness of specimens have shown a consistency with predicted results differing 0.01-3. This trained values can lead maximum 5 % error in microhardness calculations.  c-The Sum-squared error (SSE) graphic trained for 623500 Epochs was presented in Fig.11. Figure 11. Sum-Squared Error Curve Versus İteration Number

CONCLUSIONS
The overall performance of the model was quite satisfactory. The low error fractions indicate that ANNs could be a useful tool for modeling and predicting microhardness of bonded interfaces of SiC p reinforced Al alloy MMCs. Under given conditions, and with prescribed materials predicted microhardness can be utilized by designers and process engineers as and where necessary. Given and predicted values of the ANN system can also be employed at feasibility programs at no cost. This can be handled as a cost saving item at advanced production planning.