Mask Defect Filtering With ML Re-Ranking and User Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current mask inspection methods struggle to efficiently filter a large volume of defect candidates without compromising the detection of true defects, particularly in advanced lithography processes like EUV, leading to increased labor and time consumption, and a high number of false alarms.
Innovation Solution
A computerized system using machine learning (ML) to cluster defect candidates based on attributes, rank them for review, and actively retrain the model on user feedback to prioritize and filter true defects, ensuring a high capture rate of defect of interest (DOI) while reducing the number of candidates for detailed review.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large volume of defect candidates are inspected to ensure high detection rate, then the capture rate of true defects is improved, but the labor and time consumption increases significantly
Solution Approach 1:
The patent segments the large volume of defect candidates into smaller batches or groups that can be reviewed more efficiently. By dividing the inspection task into manageable portions, the system maintains high detection coverage while reducing the cognitive load and time required for manual review of each individual defect candidate.
Solution Approach 2:
The patent introduces an intermediary filtering mechanism (such as automated pre-screening algorithms or prioritization systems) that processes defect candidates before they reach manual review. This intermediary layer filters out obvious false alarms or low-priority candidates, allowing human reviewers to focus only on high-probability true defects, thereby reducing overall review time while maintaining capture rate.
2Measurement precision
If manual review of all defect candidates is performed to minimize false alarms, then the accuracy of defect identification is improved, but the productivity decreases
Solution Approach 1:
The patent applies partial action by performing manual review only on a subset of defect candidates that are most likely to be true defects, rather than reviewing all candidates. Automated systems handle the majority of candidates with high confidence, and human reviewers focus only on ambiguous or high-priority cases, achieving high accuracy without the productivity loss of complete manual inspection.
Solution Approach 2:
The patent dynamically adjusts review parameters such as confidence thresholds, batch sizes, or prioritization criteria based on the specific characteristics of the defect candidates. By changing these parameters adaptively, the system optimizes the balance between accuracy and productivity, ensuring that manual review resources are allocated efficiently to cases where they provide the most value.
3Manufacturing precision
If advanced lithography processes are used to achieve submicron features, then the manufacturing precision is improved, but the sensitivity to mask defects increases
Solution Approach 1:
The patent implements preliminary action by performing enhanced mask inspection and defect detection before the lithography process begins. By identifying and addressing potential mask defects in advance, the system prevents these defects from affecting the high-precision submicron features during manufacturing, thereby protecting the manufacturing precision achieved through advanced lithography.
Solution Approach 2:
The patent establishes a feedback mechanism where inspection results from advanced lithography processes are fed back to the mask inspection system. This feedback loop allows the system to learn from actual manufacturing outcomes and refine its defect detection algorithms, improving its ability to identify defects that could impact submicron feature precision while reducing false alarms.
Data Source
AI summary
There is provided a system and method of defect filtering for a mask, comprising: clustering a group of defect candidates into one or more clusters each comprising a set of defect candidates ranked by a machine learning (ML) model according to respective probabilities of being a defect of interest (DOI); and filtering each cluster to identify a subset of DOIs, comprising: presenting the set of defect candidates on a graphical user interface (GUI) to a user according to the ranking; upon receiving an indication from the user regarding at least one defect candidate, retraining the ML model based on the indication; using the retrained ML model to re-rank one or more defect candidates that are not yet reviewed, and presenting the re-ranked defect candidates on the GUI for the user; and repeating the retraining, the using and the presenting, until meeting a criterion.