Self-Supervised Virtual Metrology for CMP Material Removal Rate Prediction
Research collaboration between SUTD, A*STAR, and NUS
Overview
As semiconductor manufacturing advances toward the 2nm node and beyond in 2026, chemical mechanical planarization (CMP) remains a critical process for achieving wafer surface uniformity. Accurate prediction of material removal rate (MRR) is essential for process control and yield optimization. This research presents a self-supervised virtual metrology approach developed jointly by the Singapore University of Technology and Design (SUTD), A*STAR, and the National University of Singapore (NUS).
The Challenge
Traditional metrology requires offline measurement of test wafers, introducing latency and cost. Virtual metrology (VM) models predict quality metrics from in-situ sensor data, enabling real-time process control. However, conventional supervised VM models require large labeled datasets, which are expensive and time-consuming to obtain.
Self-Supervised Solution
The proposed method leverages self-supervised learning (SSL) to pre-train representations on unlabeled sensor data, reducing dependence on labeled examples. Key advantages include:
- Reduced labeling effort: Models learn useful features from abundant unlabeled process data.
- Improved generalization: Robust performance across varying process conditions.
- Real-time prediction: Enables inline MRR estimation for advanced process control (APC).
Relevance in 2026
With the industry's push toward digital twins and AI-driven fab operations, self-supervised virtual metrology aligns with trends in smart manufacturing. As fabs adopt more sensor data and edge computing, SSL-based VM offers a scalable path to predictive process control without proportional increases in metrology overhead.
Conclusion
This collaborative research demonstrates the potential of self-supervised learning to enhance virtual metrology for CMP, contributing to more efficient and reliable semiconductor manufacturing.
Note: This article is based on the provided title and affiliations; technical details are inferred from the context of self-supervised learning and virtual metrology.
