Semiconductor Inspection Recipe Feedback for Unmatched Defectivity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor inspection methods struggle to accurately detect defects in real-time due to high false alarm rates and missed defects, with post-fabrication electrical tests providing delayed and insufficient feedback for improving in-line inspection.
Innovation Solution
A computerized system that correlates in-line inspection data with post-fabrication test data using machine learning to optimize the inspection recipe, identifying unmatched defectivity and training a model for improved defect detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If in-line inspection is performed at high-speed with low-resolution, then productivity is improved, but measurement precision deteriorates
Solution Approach 1:
The inspection process is divided into two phases: a first phase for high-speed screening that identifies potential defect locations, and a second phase for high-resolution verification at sampled locations. This segmentation allows the system to maintain high productivity while achieving accurate defect detection through targeted high-resolution imaging only where needed.
Solution Approach 2:
Instead of applying high-resolution inspection across the entire wafer surface, the system applies it partially only to sampled locations where defects are suspected. This partial action approach maintains productivity while achieving sufficient measurement precision at critical points.
2Measurement precision
If machine learning model is trained with sufficient annotated training images, then measurement precision is improved, but loss of time increases
Solution Approach 1:
The system performs preliminary actions by collecting and annotating training images during normal production inspections. Defect images are automatically annotated with defect information during the inspection process, building up a training dataset over time without requiring separate dedicated annotation sessions, thus reducing the time loss for model training.
Solution Approach 2:
The system serves itself by automatically generating annotated training data from its own inspection outputs. The defect detection algorithm and review process automatically provide labeled defect information that can be used to retrain and improve the machine learning model, creating a self-improving system that reduces external annotation time requirements.
3Measurement precision
If post-fabrication electrical tests are performed, then measurement precision of functional defectivity is improved, but loss of time occurs due to delayed feedback
Solution Approach 1:
The system implements a feedback mechanism where post-fabrication electrical test results are fed back into the inspection system. The defectivity information from electrical tests is used to retrain the machine learning model, which then improves future in-line inspection accuracy. This closed-loop feedback continuously improves measurement precision while reducing the effective feedback delay by updating the model for subsequent inspections.
Solution Approach 2:
The machine learning model acts as an intermediary that bridges the gap between in-line inspection and post-fabrication electrical tests. It translates the delayed electrical test results into immediate improvements for in-line inspection by updating the defect detection algorithms, effectively mediating between the two measurement stages.
4Productivity
If in-line inspection is performed during fabrication, then productivity is improved, but reliability deteriorates due to high false alarm rates and missed defects
Solution Approach 1:
The system merges in-line inspection data with post-fabrication electrical test results to create a more reliable defect detection system. By combining the high-speed capability of in-line inspection with the functional accuracy of electrical tests through machine learning integration, the system achieves both productivity and reliability improvements that neither method could achieve alone.
Solution Approach 2:
The system dynamically changes inspection parameters based on feedback from electrical tests and defectivity correlations. The machine learning model adjusts inspection sensitivity, resolution requirements, and sampling rates based on learned patterns, allowing the system to maintain high productivity while improving reliability by adapting to actual defect characteristics observed in functional testing.
Data Source
AI summary
There is provided a system and method of optimizing an inspection recipe for inspecting a semiconductor specimen. The method includes obtaining test data from a test performed after inspection, the test data indicative of functional defectivity of the specimen with respect to at least one structural feature at a suspected layer; retrieving inspection data of the suspected layer including a set of inspection images and a set of defect maps of the plurality of processing steps of the suspected layer; correlating the test data and the set of defect maps of the suspected layer to identify one or more structural features of the suspected layer with unmatched defectivity; for each of the identified structural features, including at least part of the inspection images corresponding to the structural feature in a training set; and using the training set to train a machine learning (ML) model in the inspection recipe.