Difference Image-Based Training Sets for Automatic Defect Classification at Outgoing Inspection

Outgoing quality control/assurance (OQA) and defect inspection in the semiconductor industry is of utmost importance in the realm of high-volume manufacturing. Finished wafers exiting an integrated circuit (IC) fabrication facility frequently demand a high rate of inspection to ensure outgoing quality. The intent of outgoing quality inspection is to ensure that reliable and functional chips are manufactured. When establishing an effective outgoing quality system, both to prevent financial loss from return material authorization and to enable effective and timely investigation into defect issues, the accurate classification of defectivity is paramount. This paper proposes to construct an enhanced (augmented) training set based on defect-reference difference images, rather than the direct use of the defect image, in a machine learning model for classifying defects at outgoing inspection. The model using the enhanced dataset proves to have more accurate classification for critical defectivity. The proposed methodology of augmenting the training data resulted in a model with an overall recall of 88.8%, more than a 4% improvement from using only the defect image to train the model; further, the weighted precision shows an almost 5% increase using the proposed methodology. It is presumed that the enhanced dataset has a stronger capability of predicting defect characteristics by reducing background noise.