Wafer Map Defect Pattern Detection via Supervised Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In semiconductor manufacturing, supervised machine learning techniques for defect pattern detection are limited by the lack of sufficient high-quality training samples, particularly for hybrid patterns and small changes, which reduces their effectiveness in identifying root causes and improving yield.
Innovation Solution
The method generates defect pattern variants from wafer maps with systematic defects and superimposes them on maps without systematic defects to create positive training data, deriving a trained machine-learning model for recognizing known defect patterns, which can be used to analyze newly fabricated wafers and adjust manufacturing processes or equipment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised machine learning techniques are used for defect pattern detection, then defect pattern recognition capability is improved, but the lack of sufficient high-quality training samples reduces effectiveness
Solution Approach 1:
The patent creates synthetic training samples by copying and transforming existing defect patterns. Specifically, it extracts defect patterns from wafer maps, generates transformed versions through rotation, flipping, and scaling operations, and combines these with random noise to create realistic training data. This copying approach enables the model to learn from multiple variations of defect patterns without requiring additional physical wafer samples.
Solution Approach 2:
The patent performs preliminary actions by pre-processing and transforming defect patterns before they are used for training. It extracts defect patterns from training wafer maps, applies geometric transformations (rotation, flipping, scaling), and combines them with random noise to create pre-transformed defect patterns. These preliminary transformations enrich the training dataset with diverse examples before the machine learning model receives the data.
2Quantity of substance
If more training samples are generated through transformation, then training data quality and quantity are improved, but computational complexity increases
Solution Approach 1:
The patent performs all necessary transformations and data generation operations in advance during the training phase. By pre-extracting defect patterns, applying geometric transformations (rotation, flipping, scaling), and combining with random noise before model training begins, the system prepares enriched training data upfront. This preliminary action eliminates the need for complex real-time processing during actual defect detection operations.
Data Source
AI summary
Various aspects of the disclosed technology relate to training and applying a machine learning model for defect pattern detection. Defect pattern variants of one or more defect patterns are generated. The one or more defect patterns are extracted from wafer maps of wafers having at least systematic defects. Each of the generated defect pattern variants is superimposed on wafer maps of wafers having no systematic defects to generate positive training data of wafer maps, which are included in a training dataset. Based on the training dataset, a trained machine-learning model for recognizing known defect patterns on wafer maps is derived.