EUV Stochastic Defect Prediction With Threshold-Based Location Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing stochastic defect prediction models in semiconductor manufacturing are impractical for complex design layouts, require extensive experimental data for calibration, and are prone to under/over prediction of defect-free process windows, making them time-consuming and costly to implement.
Innovation Solution
A method and system that utilize a model to predict stochastic defects by sorting locations based on initial probability, determining cumulative expected defect counts, and selecting subsets above a threshold, allowing for efficient calibration and inspection of high-probability defect locations without extensive experimental data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If generalized empirical models are built with extensive experimental data for calibration, then prediction accuracy for stochastic defects is improved, but the time and cost required for implementation increases significantly
Solution Approach 1:
The patent extracts only the essential calibration parameters from the complex experimental data, separating the critical defect probability predictions from the redundant measurement information. This allows the model to achieve accurate predictions without requiring extensive calibration data, thereby reducing calibration time and cost while maintaining prediction accuracy for stochastic defects in complex design layouts
Solution Approach 2:
The patent performs preliminary model validation using a limited set of representative experimental data before full implementation. By pre-calibrating the model with carefully selected calibration data that captures the essential defect formation mechanisms, the system achieves accurate predictions without requiring comprehensive experimental calibration, thus reducing overall implementation time and cost
2Measurement precision
If generalized empirical models are built with extensive experimental data, then prediction accuracy is improved, but the complexity of the calibration process increases
Solution Approach 1:
The patent extracts and isolates the critical calibration parameters that drive stochastic defect formation, separating these from the complex experimental measurement process. This extraction simplifies the calibration process by focusing only on the essential parameters needed for accurate defect probability predictions, thereby reducing calibration complexity while maintaining high prediction accuracy across diverse design layouts
Solution Approach 2:
The patent segments the calibration process into distinct modular steps: (1) identifying critical defect formation mechanisms, (2) selecting representative calibration data, (3) calibrating specific parameters, and (4) validating predictions. This segmentation breaks down the complex calibration process into manageable tasks, reducing overall complexity while achieving accurate predictions for stochastic defects
3Reliability
If existing stochastic defect prediction models are used, then defect detection is possible, but they are prone to under/over prediction of defect-free process windows
Solution Approach 1:
The patent incorporates feedback mechanisms that continuously compare model predictions with actual inspection results, particularly in defect-free process windows. This feedback loop allows the model to learn from discrepancies between predicted and actual defect counts, adjusting the calibration parameters to improve prediction accuracy in these critical regions where existing models typically under/over predict
Solution Approach 2:
The patent dynamically adjusts calibration parameters based on the specific process conditions and design layouts being analyzed. By changing the calibration parameters adaptively rather than using fixed values, the model achieves more accurate predictions across varying process windows, particularly improving reliability in defect-free regions where traditional models fail to capture the nuanced behavior of stochastic defects
Data Source
AI summary
An initial probability of occurrence of a stochastic defect over an inspection area of a workpiece is received. All locations of the stochastic defects are sorted by the initial probability of occurrence. A cumulative expected defect count is determined and the cumulative expected defect count is normalized to be a fraction of a total expected defect count. A number of defect locations is determined to capture potential stochastic defects above a threshold of total stochastic defects.