Wafer Defect Root Cause Analysis Using 2D Image Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for identifying the root cause of defects in semiconductor structures during fabrication are inaccurate and time-consuming, often requiring expert knowledge and manual classification, which can be error-prone and inefficient.
Innovation Solution
A method using machine-learning algorithms trained on prior knowledge of the fabrication process to analyze 2-D image datasets acquired during the wafer run, determining process deviations as the root cause of defects, thereby facilitating faster and more accurate defect identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual classification and expert knowledge are used to identify defect root causes, then measurement precision may be improved, but loss of time increases significantly
Solution Approach 1:
The patent replaces manual expert classification (mechanical human analysis) with automated machine learning algorithms that process 2-D image datasets. The system uses trained models to automatically identify defect root causes, eliminating the time-consuming manual review process while maintaining or improving accuracy through consistent algorithmic application.
Solution Approach 2:
The patent creates a virtual model of the defect analysis process through machine learning algorithms that replicate expert knowledge. The system trains models on labeled 2-D image datasets to capture expert classification patterns, then uses these learned models to automatically analyze new defects, copying expert capabilities into an automated system.
2Measurement precision
If 3-D imaging is used to analyze defects, then measurement precision improves, but productivity decreases due to invasive and time-consuming procedures
Solution Approach 1:
The patent extracts the essential defect information from 2-D image datasets without requiring 3-D imaging. The machine learning algorithms process only 2-D images to identify defect characteristics and root causes, eliminating the need for time-consuming 3-D imaging procedures while retaining sufficient diagnostic capability.
Solution Approach 2:
The patent uses readily available 2-D image datasets that can be quickly captured and processed, replacing the need for expensive and time-consuming 3-D imaging. The system processes inexpensive 2-D images through automated algorithms to achieve defect analysis that previously required elaborate 3-D imaging procedures.
3Productivity
If automated machine learning analysis is used, then productivity increases, but measurement precision may deteriorate without proper training
Solution Approach 1:
The patent performs preliminary training of machine learning algorithms on labeled 2-D image datasets before deployment. The system pre-processes training data, trains models on known defect patterns, and validates performance beforehand. This preliminary action ensures the automated system achieves the necessary measurement precision before being used for actual defect analysis at high speed.
Solution Approach 2:
The patent implements feedback mechanisms where the machine learning system's predictions are validated against ground truth labels during training and can be refined through iterative improvement. The system learns from training data feedback to continuously improve its root cause identification accuracy while maintaining high processing speed.
Data Source
AI summary
A method includes obtaining at least one 2-D image dataset of semiconductor structures formed on a wafer including one or more defects during a wafer run of a wafer using a predefined fabrication process. The method also includes determining, based on at least one machine-learning algorithm trained on prior knowledge of the fabrication process and based on the at least one 2-D image dataset, one or more process deviations of the wafer run from the predefined fabrication process as a root cause of the one or more defects. A 3-D image dataset may be determined as a hidden variable.