Defect Sampling via Dynamic Weighting for Semiconductor Wafer Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating defect samples in semiconductor manufacturing are inefficient in capturing low-count defect types, prone to fluctuations, and require tuning classifiers, leading to systematic errors and biased results.
Innovation Solution
A computer-implemented method that identifies defects with the most diversity in attribute values, generates tiles based on proximity, separates defects into dynamic sample bins, and randomly selects defects from these bins to create a representative sample, eliminating the need for tuned classifiers and ensuring statistically unbiased estimates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If random sampling is used to select defects from the detected population, then the sampling scheme is easy to setup and re-normalizable, but it is not efficient in capturing lower count defect types and introduces spatial bias
Solution Approach 1:
The patent implements dynamic sampling weights that are recalculated for each wafer based on the detected defect population characteristics. Instead of using fixed random sampling, the system dynamically adjusts the probability of selecting each defect based on its type and frequency, ensuring that rare defect types are captured while maintaining ease of implementation through automated weight calculation.
Solution Approach 2:
The patent changes the sampling parameters by introducing defect-type-specific sampling weights rather than uniform random selection. The system modifies the sampling distribution based on the observed defect population, adjusting selection probabilities to optimize capture of low-count defect types while preserving the re-normalizable property of the sampling scheme.
2Productivity
If class code based sampling is used to select defects randomly from bins, then the sampling is more stable and efficient for well performing binners, but it requires trained classifiers and periodic performance monitoring
Solution Approach 1:
The patent extracts the classification function from the sampling process. Instead of requiring trained classifiers to bin defects before sampling, the system directly samples from the raw detected defects using dynamically calculated weights based on defect attributes. This removes the need for separate classifier training and monitoring infrastructure while maintaining sampling efficiency.
Solution Approach 2:
The sampling system performs its own defect characterization and weight calculation without requiring external classifiers. The system automatically analyzes defect attributes and computes sampling weights based on the observed population, making the process self-sufficient and eliminating the need for periodic classifier performance monitoring.
3Reliability
If rule-based sampling is used to select defects based on rules, then the sampling is re-normalizable, but it has similar characteristics to class code based sampling requiring defect classification
Solution Approach 1:
The patent inverts the traditional approach by not classifying defects into bins before sampling. Instead, it samples directly from the unclassified defect population using weights derived from basic defect attributes. This reversal eliminates the need for rule-based classification while preserving re-normalizability through the dynamic weighting scheme.
4Adaptability or versatility
If spatial diversification is enforced in random sampling to avoid excessive sampling of dies and clusters, then the sampling covers more areas, but it introduces subtle bias into the renormalization process that results in systematic errors
Solution Approach 1:
The patent applies local quality by allowing different sampling strategies for different defect types and locations. Instead of enforcing uniform spatial diversification rules, the system calculates location-specific sampling weights based on the actual defect population distribution, ensuring that each region is sampled proportionally to its defect significance without introducing artificial bias.
Data Source
AI summary
Methods and systems for generating defect samples are provided. One method includes identifying a set of defects detected on a wafer having the most diversity in values of at least one defect attribute and generating different tiles for different defects in the set. The tiles define a portion of all values for the at least one attribute of all defects detected on the wafer that are closer to the values for the at least one attribute of their corresponding defects than the values for the at least one attribute of other defects. In addition, the method includes separating the defects on the wafer into sample bins corresponding to the different tiles based on their values of the at least one attribute, randomly selecting defect(s) from each of two or more of the sample bins, and creating a defect sample for the wafer that includes the randomly selected defects.