Wafer Map Defect Pattern Recognition Based on Self-Organizing Incremental Neural Network

Wafer map defect pattern (WMDP) recognition is critical to evaluate the semiconductor manufacturing process. Traditional WMDP recognition algorithms, such as deep learning-based methods, only focus on improving the accuracies of the seen patterns in the training set but ignore the recognition and incremental generalization of new emerging patterns. The model can only be retrained by using the overall dataset composed of new and old patterns to gain the ability to recognize the new patterns. To address the above problems, this paper proposes a WMDP recognition method based on self-organizing incremental neural network. The proposed method detects abnormal patterns according to the relationship between input response and adaptive threshold of winner neurons, updates the model only with new pattern samples by dynamically adjusting the number or connection of neurons. Experiments show that the proposed method has the ability to classify known patterns and discover unknown abnormal patterns. Its unique competitive learning mechanism makes it possible to update the model only with new pattern samples without catastrophic forgetting.