Improving Virtual Metrology Predictions via Transfer Learning and Active Learning

Abstract

Unlike traditional metrology in semiconductor manufacturing, it uses physical methods to measure wafers that are both resource-intensive and time-consuming, increasing possibilities of causing defects to production of wafers. Virtual metrology (WM) predicts wafer measurement using sensor data, enabling real-time, non-intrusive monitoring of process performance. Our study introduces a smarter approach to virtual metrology by combining regression modeling with transfer learning to enhance model generalization under varying manufacturing conditions. The framework consists of two stages: first we fine-tuning a base model using a limited amount of labeled data from the target domain, then followed with iterative refinement via active learning in which the most uncertain predictions are identified and incorporated into the training set. This method improves prediction in the target domain, especially in cases where standalone models do not perform well.  Experimental results demonstrate proposed framework significantly outperforms models trained solely on target domain data. There is significant improvement in the refined model, achieving 77.80% in Mean Absolute Error (MAE) and 58.095% in Root Mean Squared Error (RMSE) compared to the original model. In addition, improvements in recall and reductions in false positive rates were observed, indicating the method is more effective at identifying abnormal wafers. Active learning helps to select the most appropriate sample for labelling to reduce the need for extensive datasets. The purposed method is advantageous in high-mix, low-volume (HMLV) manufacturing industry settings, where some stage or products are produced in small amounts. This innovative approach to virtual metrology aims to streamline semiconductor manufacturing, minimize defects, and optimize resource utilization by delivering strong, adaptable predictive capabilities.