Mixture and pore volume fraction estimation in AI_2O_3/SiC ceramic cake using artificial neural networks

摘要

In this study, the percentage of alumina in produced Al_2O_3/SiC ceramic cakes and the pore volume fraction in the ceramic cake are obtained by designing a back propagation neural network that uses a gradient descent learning algorithm. Artificial neural network (ANN) is an intelligent technique that can solve non-linear problems by learning from the samples. Therefore, some experimental samples have been firstly prepared to train the ANN to get it to give (to estimate) pore volume fraction (percent) and A1_2O_3 wt percent in Al_2O_3/SiC ceramic cake for any given SiC (g) amount. The most important point in this paper is that after ANN is trained using some experimental samples, it has given approximately correct outputs for some of experimental inputs that have not been used in the training. Firstly, to prepare a training set some results are experimentally obtained. In these experiments, we have obtained Al2O3/SiC ceramic cakes; SiC is obtained commercially, particles dimensions are obtained by a chemical process method. In order to prepare ceramic preforms, a chemical process was used rather than one using a mixing of ceramic powders to obtain porous Al_2O_3/SiC ceramic foams. These products were heated in a ceramic crucible in a furnace. It foamed and an Al_2O_3/SiC cake was obtained. Resulting A1_2O_3 grains had a 3D honeycomb structure and SiC particles were in the alumina grains. Consequently, a homogeneous powder mix and porosity were obtained within the cake. The morphology of the powder connections was a networking with flaky particles. These flaky alumina particles provided a large amount of porosity, which was desired for ceramic preforms to allow liquid metal flow during infiltration. A resulting high porous ceramic cake (preform) was obtained. In the preparation of ANN training module, the amount of SiC (g) is used as the input and the percentage of alumina in produced Al_2O_3/SiC ceramic cake and high porous ceramic cake (preform) are used as outputs. Then, the ANN is trained using the samples obtained in the experimental processes. In this paper, the alumina and pore volume fraction in the produced cake have been estimated for different amounts of SiC using neural network efficiently instead of time consuming experimental processes.