Joint modeling of classification and regression for improving faulty wafer detection in semiconductor manufacturing

In the semiconductor manufacturing process, it is important to identify wafers on which faults have occurred or will occur to avoid unnecessary and costly further processing and physical inspections. This issue can be addressed by formulating the faulty wafer detection problem as a predictive modeling task, in which the process parameters/measurements and subsequent inspection results concerning the faults comprise the input and output variables at the wafer level, respectively. To achieve improved predictive performance, this paper presents a joint modeling method that incorporates classification and regression tasks into a single prediction model. Given the output variables in both binary and continuous forms, the prediction model simultaneously considers both the classification and regression tasks to complement each other, where each task predicts the binary and continuous output variables, respectively. The outputs from these two tasks are combined to predict whether a wafer is faulty. The entire model is implemented as a neural network, and is trained by optimizing a single objective function. The effectiveness of the model is demonstrated with a case study using real-world data from a semiconductor manufacturer.