Sputter process variables prediction via data mining

Media manufacturing process plays a vital role in the hard disk drive (HDD) industry. Sputtering is an important process within the media manufacturing process that provides magnetic properties, which in turn affects the performance of the media in a HDD assembly. In this study, an attempt has been made to predict the critical magnetic characteristics and yield of the media, which in turn affects the overall product performance, by resorting to data mining technology. Feature selection was carried out using methods such as principal components analysis (PCA), and weight information of multilayer feed-forward neural networks (MLFF) and general regression neural networks (GRNN). In addition, classification and regression trees (CART) was used to generate rules to understand the dataset, as neural networks are considered as black boxes without rule extraction mechanism. A new hybrid architecture was also developed combining the top inputs identified by MLFF, GRNN and CART. It is concluded that models constructed using feature selection carried out by MLFF or GRNN performed very well. These models are preferred since they use the minimum number of input variables. Hence they are practically and economically more viable than other models with similar performance. The predictions have been validated using the 10-fold cross validation method, to ensure that the results are not due to any anomaly in the dataset.