Sudden Concept Drift Detection and Adaptation in Virtual Metrology for Semiconductor Manufacturing

Virtual metrology (VM) has emerged as a promising technology for semiconductor manufacturing; however, abrupt changes in production conditions can significantly degrade prediction accuracy. Therefore, detecting sudden concept drift becomes crucial. Micro Cluster Nearest Neighbor (MC-NN), a method that offers a potential solution for sudden concept drift detection, was originally designed for supervised classification tasks. As a result, it faces challenges when applied directly to high-dimensional, semi-supervised regression problems. To address these issues, this work proposes a novel Model Rebuilding Scheme (MRS) to not only detect but also adapt to sudden concept drift in high-dimensional, semi-supervised scenarios. Experimental results show that Advanced Automatic Virtual Metrology with Convolutional Neural Networks (AVM<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CNN</sub>), enhanced by MRS, outperforms MC-NN and the VM architecture proposed by Lang <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">et al</i>. [6]. Furthermore, Advanced AVM<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CNN</sub> with MRS can achieve a 24.4% average reduction in mean absolute error compared to Advanced AVM<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CNN</sub> without MRS. The proposed framework is promising to provide a robust method for detecting and adapting to sudden concept drift in next-generation VM applications, ensuring sustained prediction accuracy.