Nuisance Mining for Novel Defect Discovery in Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for configuring nuisance filters in optical inspections of semiconductor wafers, particularly those based on machine-learning algorithms, risk missing defects not present during the training process, leading to incomplete defect detection and reduced throughput.
Innovation Solution
A method that involves creating a nuisance bin, partitioning the defect population, segmenting it, selecting and replicating defects to generate new training data, and training a binary classifier to segregate novel defects from nuisance defects, enabling the discovery of unknown defects and improving filter maturity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If ML-based nuisance filters are used for defect classification, then productivity and automation are improved, but the risk of missing novel defects increases due to training set limitations
Solution Approach 1:
The system performs preliminary actions by generating synthetic defect data before actual inspection. The nuisance filter is trained on augmented datasets that include artificially created defect patterns, preparing the model to recognize novel defect types that may appear during production inspections.
Solution Approach 2:
The system creates copies of existing defect patterns through synthetic data generation. By replicating and transforming known defect characteristics, the system produces training examples of novel defect types without requiring physical samples, enabling the ML model to learn from these synthesized copies.
2Reliability
If training set accumulation is used to improve filter maturity, then defect detection reliability is improved, but the process becomes more complex and time-consuming
Solution Approach 1:
Instead of accumulating physical training samples over time, the system copies and synthesizes defect data computationally. Synthetic defect patterns are generated by transforming existing defect characteristics, creating diverse training examples without requiring additional physical samples or extended training periods.
Solution Approach 2:
The system changes parameters of existing defect patterns to generate synthetic novel defects. By modifying characteristics such as size, shape, intensity, and position parameters of known defects, the system creates diverse training examples that expand the model's capability without requiring physical sample accumulation.
3Reliability
If manually constructed nuisance filters are used, then domain knowledge can be leveraged to reserve signal space for novel defects, but the process is tedious and subjective
Solution Approach 1:
The system automatically copies and transforms defect patterns through computational synthesis, eliminating manual filter construction. Synthetic defect data is generated by algorithmic transformation of existing patterns, replacing the tedious manual process of reserving signal space while maintaining the benefit of domain knowledge embedded in the synthesis algorithms.
Solution Approach 2:
The system replaces manual mechanical processes of filter construction with automated computational methods. The subjective and tedious manual adjustment of nuisance filters is substituted with objective algorithmic synthesis of training data, maintaining reliability while dramatically improving ease of manufacture.
4Productivity
If defects are removed from inspection to reduce nuisance rates, then productivity is improved, but the ability to discover new defects is reduced
Solution Approach 1:
The system creates synthetic copies of potential novel defects for training purposes, preserving the ability to discover new defect types even when removing actual defects from production inspection. The synthetic data maintains the information about novel defect patterns without requiring them to be present in the production stream.
Solution Approach 2:
The system performs preliminary analysis and synthetic data generation before production inspection. By preparing synthetic training data in advance that encompasses potential novel defect types, the system preserves discovery capability while enabling aggressive nuisance filtering during actual production to maximize throughput.
Data Source
AI summary
A method of defect discovery can include providing a nuisance bin in a nuisance filter, partitioning the defect population into a defect population partition, segmenting the defect population partition into a defect population segment, selecting from the defect population segment a selected set of defects, computing one or more statistics of the signal attributes of the defects in the defect population segment, replicating the selected set of defects to yield generated defects, shifting the generated defects outside of the defect population segment, creating a training set, and training a binary classifier. This method can be operated on a system. The method can enable a semiconductor manufacturer to determine more accurately the presence of defects that would otherwise have gone unnoticed.