A study on the improvement of machine learning (ML)-based defect prediction model reflecting process variations

This paper presents the research findings and application results of the ML-Statistics Risk Pattern Predictor (ML-SRPP), a novel approach that combines pattern segmentation, Greedy algorithm-based sampling, and unbiased statistical estimation to predict and mitigate process-related defects in advanced semiconductor technology nodes.

The ML-SRPP framework first extracts the pattern type and usage frequency of the product design using a pattern segmentation technique, then applies the Greedy algorithm to select the most representative patterns within the measurement constraints. An unbiased estimation method is used to ensure 99% reliable process variation data, which is then incorporated into the ML model.

The enhanced ML-SRPP model can predict the statistical risk of critical patterns, identifying the minimum and maximum critical dimension (CD) values that patterns can exhibit. This capability enables early detection and mitigation of potential defects related to CONTACT, VIA, and metal layers, contributing to significant yield improvements in the latest 3nm products.

The paper demonstrates the effectiveness of the ML-SRPP approach through several case studies, including CONTACT_A open risk, VIA_A not open/short risks, CONTACT_B short risks, and VIA_B to METAL_A short risk prediction and prevention. The methodology has been applied to 2 nm and 1.4 nm technology nodes, showcasing its scalability and importance for advanced semiconductor development.

This paper was presented at the 2025 SPIE Advanced Lithography + Patterning symposium.

What you'll learn:

• How to leverage pattern segmentation, Greedy algorithm-based sampling, and unbiased statistical estimation to comprehensively characterize process variations in advanced semiconductor manufacturing.
• The process of developing an enhanced ML-based risk pattern predictor (ML-SRPP) that can accurately forecast the statistical risk of critical patterns, going beyond simple CD prediction.
• Practical applications of the ML-SRPP approach in identifying and mitigating potential defects related to CONTACT, VIA, and metal layers, leading to substantial yield improvements in 3 nm, 2 nm, and 1.4 nm technology nodes.

Who should read this:

• Semiconductor process engineers, R&D teams, and yield enhancement experts involved in the development and optimization of advanced technology nodes.
• Engineering managers and decision-makers in semiconductor fabs and design houses who are responsible for accelerating the introduction of new technologies and improving manufacturing yields.
• Researchers and data scientists working on the application of machine learning and statistical techniques to solve complex challenges in semiconductor manufacturing, particularly in the areas of defect prediction and process control.