Enhancing wafer defect detection via ensemble learning

Abstract

Wafer inspection is crucial for semiconductor manufacturing, as it identifies defects in wafers before manufacturing. Wafer defect detection avoids wasting time and production capacity, boosts productivity, and assures production quality. In this paper, we propose an ensemble learning-based method for wafer defect detection that fuses the classification results of four models, namely, ResNet, ResNeSt, ResNeSt + CBAM, and ResNeSt + Self-attention. During the integration phase, we employ a hybrid strategy that combines weighted averaging and voting to determine weight coefficients. Our analysis shows that the model’s performance surpasses that of the arithmetic mean model within an interval of 0.8–1, according to our mathematical derivations. Furthermore, results demonstrate and substantiate that optimal performance is attained by setting the weighting value to 1. We experimentally validated the effectiveness of the proposed method on the WM-811k industrial dataset. In the experiments, the ensemble learning based method achieves an accuracy of 99.70%, which outperforms the individual model. Our approach outperforms the traditional arithmetic mean model by combining the strengths of all prediction models to improve prediction accuracy. Experimental results demonstrate that the proposed method has the potential to be an ideal option for wafer defect detection.