An Evaluation of Continual Learning for Advanced Node Semiconductor Defect Inspection

Amit Prasad,Bappaditya Dey,Victor Blanco,Sandip Halder

2024-07-18

Abstract:Deep learning-based semiconductor defect inspection has gained traction in recent years, offering a powerful and versatile approach that provides high accuracy, adaptability, and efficiency in detecting and classifying nano-scale defects. However, semiconductor manufacturing processes are continually evolving, leading to the emergence of new types of defects over time. This presents a significant challenge for conventional supervised defect detectors, as they may suffer from catastrophic forgetting when trained on new defect datasets, potentially compromising performance on previously learned tasks. An alternative approach involves the constant storage of previously trained datasets alongside pre-trained model versions, which can be utilized for (re-)training from scratch or fine-tuning whenever encountering a new defect dataset. However, adhering to such a storage template is impractical in terms of size, particularly when considering High-Volume Manufacturing (HVM). Additionally, semiconductor defect datasets, especially those encompassing stochastic defects, are often limited and expensive to obtain, thus lacking sufficient representation of the entire universal set of defectivity. This work introduces a task-agnostic, meta-learning approach aimed at addressing this challenge, which enables the incremental addition of new defect classes and scales to create a more robust and generalized model for semiconductor defect inspection. We have benchmarked our approach using real resist-wafer SEM (Scanning Electron Microscopy) datasets for two process steps, ADI and AEI, demonstrating its superior performance compared to conventional supervised training methods.

Computer Vision and Pattern Recognition,Artificial Intelligence,Machine Learning

What problem does this paper attempt to address?

The paper primarily addresses the detection and classification of new types of defects that continuously emerge in the semiconductor manufacturing process. Traditional supervised learning methods may encounter the problem of "catastrophic forgetting" when dealing with new types of defect datasets. This means that the model forgets previously learned knowledge when learning new tasks, leading to a decline in performance on old tasks. Additionally, storing all historical data for retraining or fine-tuning becomes impractical in real production environments (especially in high-volume manufacturing scenarios) due to the massive amount of data.

To solve the above issues, the authors propose a task-agnostic continual learning method based on meta-learning. This method allows the model to incrementally add new defect categories and can be extended to create a more robust and generalized model for handling semiconductor defect detection. In this way, the model can retain old knowledge while continuously learning new knowledge.

Specifically, the study conducted three case studies:

1. **Case Study 1**: Evaluated the model's ability to minimize forgetting of previously trained defect categories when incrementally learning new defect categories.
2. **Case Study 2**: Further tested the model's performance in incrementally learning new defect categories at different manufacturing stages (AEI images) and assessed its forgetting of defect categories previously trained on the entire ADI dataset.
3. **Case Study 3**: Compared three different training strategies: (i) training a model with all defect categories at once; (ii) training a portion of defect categories first, then fine-tuning for new categories; (iii) using the proposed incremental supervised training strategy, training a portion of defect categories first, then gradually adding new categories for training.

Through these case studies, the paper demonstrates the effectiveness of the proposed method, especially in reducing "catastrophic forgetting." Ultimately, the experimental results show that the proposed continual learning strategy can effectively address these issues, thereby improving the accuracy and efficiency of semiconductor defect detection and classification.