Wafer Quality State Prediction via Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining the quality state of a wafer discard entire wafers as rejects if a single wafer level test value is incorrect, failing to provide precise quality assessment for individual chip areas.
Innovation Solution
A computer-implemented method using machine learning algorithms to predict wafer level test values based on process control monitoring metrics collected at different coordinates, allowing for precise quality state determination of chips and minimizing rejects by approximating test values without additional measurements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single wafer level test value is determined for the entire wafer, then the quality decision is made for the entire wafer, but the precision of quality assessment for individual chip areas deteriorates and unnecessary rejects increase
Solution Approach 1:
The patent divides the wafer into multiple spatial regions (e.g., center, edge, intermediate areas) and determines separate wafer level test values for each region. This segmentation allows precise quality assessment for individual chip areas while reducing unnecessary rejects, as only defective regions are identified rather than discarding the entire wafer.
Solution Approach 2:
The patent applies different quality criteria and test value thresholds to different spatial regions of the wafer. By determining region-specific wafer level test values and comparing them against region-specific thresholds, the method enables localized quality assessment that accounts for spatial variations in manufacturing processes, thereby improving measurement precision without increasing overall reject rates.
2Measurement precision
If wafer level test values are determined for multiple areas of the wafer, then the quality assessment precision improves, but the measurement time and complexity increase
Solution Approach 1:
The patent performs preliminary clustering of process control monitoring metrics during the manufacturing process to predict spatial distributions of wafer level test values before actual measurement. This preliminary action identifies which regions are likely to contain defects, allowing focused measurement efforts on critical areas and reducing overall measurement time while maintaining high quality assessment precision.
Solution Approach 2:
The patent uses process control monitoring metrics (which are easier and faster to obtain) as proxies or copies to predict wafer level test values. By establishing correlations between process metrics and final test values, the method enables indirect assessment of quality in multiple regions without performing time-consuming direct measurements in all areas.
3Measurement precision
If process control monitoring metrics are collected at multiple coordinates, then the prediction accuracy of wafer level test values improves, but the device complexity and measurement setup increase
Solution Approach 1:
The patent employs a unified machine learning model that can process multiple types of process control monitoring metrics (e.g., Vth, Vbe, other technology-specific parameters) collected at different coordinates and predict multiple wafer level test values simultaneously. This multi-functional approach consolidates what would otherwise require separate measurement systems for each metric and prediction task, reducing overall device complexity while maintaining high prediction accuracy.
Solution Approach 2:
The patent transforms process control monitoring metrics into predicted wafer level test values through parameter changes enabled by machine learning algorithms. By learning the functional relationships between process parameters and test values from training data, the system can accurately predict test values without requiring complex direct measurement setups, thereby improving prediction accuracy while minimizing device complexity.
Data Source
AI summary
A computer-implemented method determines a quality state of a wafer. The method includes providing at least three process control monitoring metrics of the wafer. Each process control monitoring metric is collected on the wafer at a different process control monitoring coordinate. The method further includes inputting the at least three process control monitoring metrics and the different process control monitoring coordinates of the process control monitoring metrics into at least one machine learning algorithm, and outputting at least three approximated wafer level test values by the at least one machine learning algorithm. The method also includes determining the quality state of the wafer based on the at least three approximated wafer level test values.