Yield Root Cause Analysis for Process Tool Interaction Effects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In semiconductor and TFT-LCD manufacturing, identifying the root causes of yield loss is challenging due to the complexity of interactions between process tools and parameters, leading to difficulties in enhancing production yield, especially during the research-and-development and ramp-up phases where data is limited and high-dimensional variable selection problems arise.
Innovation Solution
A two-phase method and system that utilize algorithms like the triple phase orthogonal greedy algorithm (TPOGA), least absolute shrinkage and selection operator (LASSO), and regression tree to identify key process devices, parameters, and inter-tool interactions causing yield losses, and determine threshold values affecting production yield, thereby enabling continuous improvements in the production line.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional big data analysis is performed on all production-related data to find root causes of yield loss, then comprehensive analysis coverage is achieved, but the complexity and difficulty of searching for root causes increases significantly due to enormous and complicated data
Solution Approach 1:
The patent segments the enormous production data into structured datasets with specific schemas for different process stages. It divides the analysis into multiple phases (data collection, data processing, model training, validation) and separates individual process tool analyses from interaction effect analyses, making the complex data analysis task manageable and systematic
Solution Approach 2:
The patent introduces an intermediary structured data schema that transforms raw production data into organized datasets with defined relationships between process tools, parameters, and outcomes. This intermediary structure acts as a bridge between raw data and analytical models, reducing the complexity of direct analysis on unstructured enormous data
2Measurement precision
If yield management system analyzes all production tools to find root causes of defects, then comprehensive defect detection is achieved, but the system becomes ineffective during R&D and ramp-up phases when the number of workpieces is small
Solution Approach 1:
The patent performs preliminary actions by collecting and structuring data from multiple process stages before conducting the actual root cause analysis. It pre-processes data into organized schemas, pre-identifies potential interaction effects between process tools, and prepares training datasets in advance, enabling effective analysis even with limited workpiece samples during R&D phases
Solution Approach 2:
The patent shifts from analyzing only individual process tool outputs to analyzing interaction effects between multiple process tools as an additional dimension. By examining how combinations of process tools affect yield, the system can detect root causes with fewer workpieces, effectively adding a new analytical dimension that compensates for limited sample sizes
3Measurement precision
If interaction effects between process tools are identified and analyzed, then more accurate root cause identification is achieved, but the computational complexity and data processing requirements increase
Solution Approach 1:
The patent segments the interaction effect analysis into distinct phases: first identifying individual process tool impacts, then analyzing pairwise interactions, and finally examining higher-order interactions. This segmented approach allows computational resources to be allocated efficiently at each stage rather than attempting to analyze all interactions simultaneously, reducing overall computational complexity
Data Source
AI summary
Embodiments of the present disclosure provide a two-phase process for searching root causes of a yield loss in a production line. In a first phase, an interaction between two process tools, that between two parameters, or that between one process tool and one parameter that is likely to cause the yield loss is identified. In a second phase, a threshold of the parameter that is likely to cause the yield loss and is obtained from the first phase is identified. In each phase, two different algorithms can be used to generate a reliance index (RII) for gauging the reliance levels of their search results.