A Neural Network Approach to Analyze FDC Data in Semiconductor Manufacture

Abstract

Fault Detection and Classification (FDC) Data reflects machine real-time status that provide information for diagnosing processing quality. In practice, most semiconductor engineers analyze FDC data by experience. Usually, they only use parameters which they thought are important and resolve data in an inherent subjective method. However, with the development of sensor and communication technology, there are increasing numbers of variables collected in milliseconds. Therefore, analyzing by human experience becomes more complicated, inconsistent, and unreliable. To solve this problem, we proposed a system to transform FDC data into summary statistics that can assist engineer to trace root causes and provide forecasted inline value for corresponding process. The system consists of two parts: a Convolutional Neural Network (CNN) model that pre-process FDC Data and a neural network model based on transformer framework to predict inline value. Numerical results show that the mixed model outperform simple transformer or CNN model and nearly a hundred times faster than pure Long Short-Term Memory (LSTM) model.