AI-Powered Fault Detection In Semiconductor Fabrication: A Data-Centric Perspective

Abstract

This paper presents an AI algorithm to detect faults in semiconductor fabrication. The algorithm works in multiple stages. In the first stage, a signal classification approach is proposed. The segmentation of long time series signals is performed, and a one-class classification algorithm is proposed. The class is then created using the normal samples to identify the anomalies later in the diffusion process. In the second stage, a neural network-based tool is proposed to explain and filter the results from the first model.

Semiconductor fabrication is a highly complex and automated technological process that consists of hundreds of steps. As equipment becomes more advanced, the quality of the equipment and control of the process improve, making the detection of faults more difficult and therefore more significant. In semiconductor manufacturing, chip-level faults should be detected using equipment-level signals throughout the fabrication process.

Semiconductor fabrication is a combination of a series of unit processes with different themes that create 2D or 3D patterns on a silicon wafer. Each time a signal is collected, it is a time series consisting of hundreds of thousands of high frequency data points. In addition, these signals are affected by other variables such as the temperature of the manufacturing environment, recipe parameters, and noise. To understand such complex and high-frequency data for a specific unit process, the first approach is to characterize each signal using time intervals as a result of statistical information such as mean, median, maximum, minimum, kurtosis, and more. A machine learning approach is then used to detect anomalies, which impacts the normal operation of equipment and productivity [1].