Self-Supervised Virtual Metrology For CMP MRR Prediction (SUTD, A*STAR, NUS)
Original reporting by Semiconductor Engineering

Chemical Mechanical Planarization (CMP) Material Removal Rate (MRR) prediction is a vital process in semiconductor manufacturing, directly influencing chip quality and production efficiency. Achieving precise control requires accurate, real-time estimations of MRR, a task often hampered by the need for extensive labeled data and the inherent challenges of manufacturing environments. Addressing this critical challenge, researchers from the Singapore University of Technology and Design (SUTD), A*STAR, and National University of Singapore (NUS) have introduced Global Regression Pretraining (GREP), a novel self-supervised, attention-based framework for CMP MRR prediction.
Unlike traditional methods that demand vast quantities of labeled sensor data, GREP employs a sophisticated two-stage approach. It first leverages abundant *unlabeled* sensor signals to learn robust temporal representations through global proxy-target regression. This innovative self-supervision allows the model to grasp underlying process dynamics without explicit human annotation, overcoming a major bottleneck in data availability.
Proving its value
Following this comprehensive pretraining, the framework undergoes supervised fine-tuning to precisely estimate MRR. This powerful combination significantly enhances GREP's practical utility, as demonstrated in rigorous evaluations against a feature-engineered supervised baseline. GREP exhibited superior performance and resilience, particularly excelling when faced with limited labeled data, substantial sensor noise, and common temporal process drift. This breakthrough offers a robust solution for a critical aspect of semiconductor fabrication, promising improved chip quality and efficiency.
The Global Regression Pretraining (GREP) framework developed by researchers at SUTD, A*STAR, and NUS represents a significant advancement in virtual metrology for chemical mechanical planarization (CMP) processes. By leveraging a self-supervised, attention-based architecture, GREP demonstrates superior accuracy and robustness in predicting material removal rates (MRR), even when confronted with limited labeled data, sensor noise, and temporal process drift. This capability addresses critical challenges in semiconductor manufacturing, offering a more reliable and efficient approach to real-time process monitoring and control compared to traditional feature-engineered baselines.
Broader Industrial Impact
The implications of GREP extend far beyond the precise domain of semiconductor manufacturing. This research highlights the increasing maturity and practical applicability of self-supervised learning for complex industrial time-series data. The ability to extract meaningful representations from vast quantities of unlabeled sensor signals, followed by fine-tuning for specific tasks, is a paradigm shift. This approach can unlock significant efficiencies and quality improvements across various heavy industries—from aerospace and automotive to energy and pharmaceuticals—where large datasets are common but expert-labeled data is scarce and expensive. By providing a robust method for predictive maintenance, anomaly detection, and process optimization, GREP and similar frameworks promise to accelerate the adoption of advanced AI in mission-critical industrial applications, driving towards more autonomous and intelligent manufacturing ecosystems worldwide.
Frequently asked questions
- What is CMP MRR prediction and why is it crucial for semiconductor manufacturing processes?
- Chemical Mechanical Planarization (CMP) is a critical step in semiconductor fabrication, ensuring a smooth, uniform wafer surface. Material Removal Rate (MRR) prediction involves estimating how much material is removed during this process. Accurate MRR prediction is crucial for maintaining wafer quality, minimizing defects, and optimizing manufacturing efficiency, directly impacting chip performance and production costs. Virtual metrology offers a non-invasive way to achieve this.
- What is self-supervised learning, and how does it benefit industrial predictions like CMP MRR?
- Self-supervised learning is a machine learning approach where a model learns representations from data without explicit human-provided labels. Instead, it generates labels from the data itself, often by predicting missing parts or transformations. This is highly beneficial in industrial settings where labeled data is scarce or expensive. For CMP MRR, it allows models to leverage abundant unlabeled sensor data to learn robust temporal patterns, improving prediction accuracy even with limited supervised data.
- What is the GREP framework and how does it predict CMP material removal rates?
- The Global Regression Pretraining (GREP) framework is an attention-based, self-supervised method designed for Chemical Mechanical Planarization (CMP) Material Removal Rate (MRR) prediction. It first learns temporal representations from vast amounts of unlabeled sensor signals through global proxy-target regression. Subsequently, these learned representations are fine-tuned with a smaller set of labeled data for precise MRR estimation, enhancing prediction accuracy and robustness, especially under challenging conditions like noise or limited data.