New neural network-based method for Stribeck curve construction during chemical mechanical planarization

Abstract

Chemical mechanical planarization (CMP) has been widely used in integrated circuit (IC) processing industry to achieve both local and global surface planarity through combined chemical and mechanical actions. In CMP, Stribeck curve helps determine the lubrication mechanism of CMP processes since it provides direct evidence of the extent of contact among wafer, pad asperities and slurry particles. Traditionally, the procedure for constructing the Stribeck curve is as follows: (1) polish wafers at various pressures and sliding velocities to obtain the COF (coefficient of friction) value for a particular Sommerfeld number; (2) put the experimental data in the plot of COF vs. Sommerfeld number; (3) construct the Stribeck curve via curve fitting. In this study, a new neural network-based method, in contrast to curve fitting, is presented to obtain the Stribeck curve. A multilayer radical basis function (RBF) neural network is constructed where the ratio of the sliding velocity to the pressure is the input parameter while the COF value is the output. The neural network is trained using experimental data. After training, the neural network can successfully capture the non-linear relationship between the input and output. Therefore, COF for a particular combination of pressure and velocity can be accurately predicted.