Defect-Map Image Alignment for Cross-Resolution Wafer Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image alignment techniques for semiconductor manufacturing face challenges in aligning low-resolution and high-resolution images without distinctive patterns, leading to insufficient training data for machine learning models and difficulty in distinguishing defects from nuisances.
Innovation Solution
The use of a defect map to align images of different resolutions by identifying common defects and calculating an offset value, allowing for die-to-die alignment even in the absence of distinctive patterns, thereby generating aligned image pairs for training machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution images are captured to distinguish defects from nuisances, then detection precision is improved, but image capture time increases significantly
Solution Approach 1:
The patent uses machine learning models to generate synthetic high-resolution images from low-resolution images. The ML model learns the relationship between low-resolution and high-resolution image characteristics and creates realistic high-resolution copies that preserve defect information without requiring actual high-resolution captures, thus reducing inspection time while maintaining detection accuracy
Solution Approach 2:
The patent replaces expensive and time-consuming high-resolution imaging with faster low-resolution imaging combined with computational generation. The low-resolution images serve as temporary placeholders that are quickly processed through ML models to produce the necessary high-resolution output, eliminating the need to always capture and store expensive high-resolution data
2Device complexity
If conventional alignment techniques are used without distinctive patterns, then device complexity is reduced, but alignment precision deteriorates
Solution Approach 1:
The patent introduces defect maps as intermediary objects that bridge the alignment process. Instead of directly aligning images based on visual patterns, the system first creates defect maps from both low-resolution and high-resolution images, then uses these maps to determine alignment transformations. This intermediary approach enables precise alignment even when images lack distinctive patterns
Solution Approach 2:
The patent replaces traditional visual pattern recognition and mechanical alignment methods with a computational approach using defect maps and transformation models. The system substitutes direct visual alignment with a multi-step process involving defect detection, map generation, and mathematical transformation, achieving precise alignment through computation rather than visual matching
3Quantity of substance
If machine learning models are trained without sufficient aligned image pairs, then training data requirements are reduced, but model accuracy deteriorates
Solution Approach 1:
The patent performs preliminary alignment of low-resolution and high-resolution images using defect maps before using them for training. By pre-aligning the images and generating corresponding defect maps, the system creates properly paired training data that maintains spatial relationships between features in both resolutions, ensuring accurate model training even with limited data
Solution Approach 2:
The patent generates synthetic aligned image pairs by copying and transforming existing image data through machine learning models. The system creates artificial training examples that preserve the relationships between low-resolution and high-resolution images, effectively multiplying the available training data while maintaining accuracy through the generative models
Data Source
AI summary
A method for die-to-die (D2D) image alignment using a defect map associated with an image. The method includes accessing a set of images of a substrate, which correspond to different image capture conditions. The locations of various defects on the set of images are obtained and a defect map indicating relative locations of at least some of the defects is generated. The set of images are aligned with each other using the defect map to generate an aligned set of images.