A Novel Multi-Scale Feature Learning-Based Residual Network Improves Semiconductor Wafer Defect Classification Accuracy

Wafer inspection is a crucial step in the semiconductor manufacturing process. Accurate inspection of wafer defects is of great significance in ensuring chip quality, improving production efficiency, and reducing cost losses. However, most wafer defects are irregular and tiny which brings great challenges to defect inspection. At present, most wafer defect inspection is also inefficient and has low accuracy. Addressing this, a novel wafer defect classification algorithm based on the residual network was proposed in this study. The algorithm proposes a multi-scale feature learning block that can obtain finer-grained features and distinguish different types of defects as the network depth increases. The performance of this algorithm was evaluated using the MixedWM38 dataset of semiconductor wafers and a classification accuracy of 99.13% was achieved, demonstrating competitive results when compared to other methods.