Semi-Supervised Deep Kernel Active Learning for Material Removal Rate Prediction in Chemical Mechanical...

Area of Science:

• Materials Science
• Chemical Engineering
• Machine Learning

Background:

• Material removal rate (MRR) is crucial in chemical-mechanical planarization (CMP) but challenging to measure directly.
• Existing data-driven virtual metrology (VM) methods often discard unlabeled data, leading to information loss.
• Accurate MRR prediction is vital for process control and optimization in semiconductor manufacturing.

Purpose of the Study:

• To propose a novel semi-supervised deep kernel active learning (SSDKAL) model for accurate MRR prediction in CMP.
• To leverage both labeled and unlabeled data effectively to improve prediction accuracy.
• To address the limitations of existing VM methods in handling large unlabeled datasets.

Main Methods:

• Feature extraction using clustering-based phase partition and phase-matching algorithms.
• Integration of a deep network to replace the kernel in Gaussian process regression for enhanced feature extraction.
• Application of semi-supervised regression and active learning strategies to maximize the utility of unlabeled samples.

Main Results:

• The proposed SSDKAL model demonstrated superior performance on a CMP dataset compared to supervised and co-training based semi-supervised methods.
• Achieved a lower mean square error across various proportions of labeled samples.
• Outperformed physics-based VM, Gaussian-process-based regression, and stacking models in prediction accuracy.

Conclusions:

• The SSDKAL model effectively utilizes unlabeled data for improved MRR prediction in CMP.
• This approach offers a more efficient and accurate virtual metrology solution compared to existing methods.
• The findings highlight the potential of deep kernel active learning in optimizing complex manufacturing processes.