Semiconductor Defect Detection with Orthogonal Convolution Context Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect detection and classification methods in semiconductor wafers are affected by process variations, leading to variations in optical image properties and context attributes that do not correlate with defectivity, resulting in undesirable noise and misclassification of defects.
Innovation Solution
Calculating context attributes by convolving the pattern of a semiconductor die's layer with orthogonal kernels representing the optical imaging system's response, using a machine-learning system to classify defects and identify care areas, thereby isolating defects of interest from nuisance defects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If context attributes are calculated from optical images of semiconductor wafers, then defect detection and classification can be performed, but process variations cause intensity and contrast changes that do not correlate with defectivity, leading to detection and classification variations
Solution Approach 1:
The patent transforms the optical image data from intensity/contrast domain to frequency domain by calculating the Fourier transform modulus spectrum. This parameter transformation converts process variation-induced intensity and contrast changes into a representation where they appear as background patterns, enabling defect detection algorithms to focus on anomalies rather than being confounded by process variations. The Fourier transform modulus spectrum preserves the spatial frequency information while eliminating the direct impact of intensity scaling and contrast changes.
2Productivity
If traditional optical image properties are used for defect classification, then classification decisions can be made, but variations in intensity and contrast from process variation cause classification decisions to vary
Solution Approach 1:
The patent applies Fourier transform to convert spatial domain image properties into frequency domain characteristics. This parameter change allows the classification algorithm to work with spectral features that are invariant to intensity and contrast variations caused by process changes. The modulus spectrum provides a stable feature set that maintains classification precision across different process conditions while enabling high-throughput processing.
Data Source
AI summary
Context attributes for optical imaging of a patterned layer of a semiconductor die are calculated. Calculating the context attributes includes calculating convolutions of a pattern of the patterned layer with respective kernels of a plurality of kernels, wherein the plurality of kernels is orthogonal. Defects on the semiconductor die are found in accordance with the context attributes.