Bayesian GLM for Predicting No-Defect Physical Failure Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The semiconductor industry faces prolonged IC development process timelines due to delays in fault analysis, particularly when No Defect Found (NDF) outcomes are common during Physical Failure Analysis (PFA), leading to increased costs, delayed product releases, and inefficient use of resources.
Innovation Solution
A Bayesian Generalized Linear Model (BGLM) is used to predict PFA outcomes, incorporating historical data and prior knowledge to estimate the probability of NDF or Defect-Found results, thereby optimizing the PFA process and reducing unnecessary costly procedures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Physical Failure Analysis (PFA) is performed on all failing IC devices, then defect detection capability is improved, but development time and costs increase significantly
Solution Approach 1:
The patent applies preliminary action by performing statistical analysis and predictive modeling before conducting PFA. The system analyzes test data, identifies patterns, and predicts which devices are likely to yield defects during PFA. This preliminary screening allows the most promising candidates to be selected for PFA, avoiding unnecessary PFA on devices unlikely to provide useful information, thereby reducing overall development time while maintaining defect detection capability.
Solution Approach 2:
The patent implements self-service by enabling the system to automatically analyze test data, generate predictions, and identify candidate devices for PFA without requiring manual review of each failing device. The automated statistical analysis and predictive algorithms serve the selection process independently, reducing human intervention time and accelerating the overall workflow.
2Measurement precision
If PFA is performed on all failing IC devices, then defect identification accuracy is improved, but resource consumption and costs increase
Solution Approach 1:
The system performs preliminary statistical analysis and predictive modeling to identify high-probability candidates for PFA before the actual PFA process. By predicting which devices are most likely to yield defects based on test data patterns, the system ensures that PFA resources are concentrated on devices where defect identification is most probable, maximizing resource efficiency while maintaining high defect identification accuracy.
Solution Approach 2:
The patent changes the parameter of device selection from uniform (all failing devices) to selective (predicted high-yield devices). By using statistical parameters and prediction scores to filter and rank candidate devices, the system transforms the resource allocation strategy, directing PFA resources to devices with the highest expected value while reducing overall resource consumption.
3Productivity
If PFA processes are reduced to save time and resources, then development efficiency is improved, but defect detection capability deteriorates
Solution Approach 1:
The system uses automated statistical analysis and predictive algorithms to independently identify the most promising PFA candidates from test data. This self-service capability ensures that even with reduced PFA volume, the selected candidates have high probability of yielding defects, thereby maintaining defect detection capability while improving development efficiency through selective resource allocation.
Solution Approach 2:
The patent transforms the approach by changing from quantity-based PFA (number of devices) to quality-based PFA (predicted yield probability). By using prediction scores and statistical parameters to select candidates, the system ensures that each PFA process has high expected value, maintaining overall defect detection capability while reducing total PFA volume and improving efficiency.
Data Source
AI summary
A Physical Fault Analysis (PFA) outcome prediction tool utilizes previously-generated evaluation data and associated PFA outcome data to generate a Bayesian Generalized Linear Model (BGLM), and then utilizes the BGLM to generate a PFA outcome prediction for newly-submitted evaluation data that operably characterizes measured operating characteristics of an IC chip that is being developed. The BGLM generation methodology by utilizing a Generalized Linear Model (GLM) in a Bayesian framework to form a hierarchical model representing the evaluation data and associated PFA outcome data as a linear combination. The PFA outcome prediction includes a credible interval of a posterior distribution that effectively represents a cross-sectional portion of the BGLM corresponding to the newly-submitted evaluation data. The previously-generated evaluation data and associated PFA outcome data are stored in a training data library, which is updated to include newly-submitted evaluation data and associated PFA outcome after each PFA is performed.