Tensor Decomposition for Semiconductor Inspection Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semiconductor inspection techniques suffer from decreased sensitivity due to additional noise from adjacent reference imagery, leading to process variation and alignment errors, which affect the accurate detection of defects on semiconductor wafers.
Innovation Solution
A system and method utilizing tensor decompositions and singular value decomposition (SVD) processes to generate improved reference images for inspection comparison, reducing residual noise and enhancing defect detection sensitivity by performing decomposition processes on target tensors to form reference tensors and identifying differences between target and reference image frames.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional die-to-die inspection techniques are used to compare target die with reference die, then defect detection capability is provided, but process variation errors and alignment errors between target die and reference die introduce additional noise that decreases inspection sensitivity
Solution Approach 1:
The patent extracts and removes the harmful noise components from the difference image through singular value decomposition. By decomposing the difference image into singular vectors and selectively eliminating those corresponding to noise (rather than defect signals), the method isolates and removes only the detrimental process variation and alignment errors while preserving genuine defect information.
Solution Approach 2:
The patent transforms the conventional 2D image comparison into a multi-dimensional tensor analysis problem. By representing images as tensors and applying tensor decomposition, the method adds mathematical dimensions for analyzing variance sources, enabling separation of noise from defects through singular value analysis across multiple image frames and decomposition components.
2Reliability
If die-to-median die or die-to-computed reference die techniques are used to reduce process variation and alignment errors, then reference robustness is improved, but residual noise remains that is insufficient for many inspection processes
Solution Approach 1:
The patent implements a feedback mechanism where the decomposition analysis of multiple difference images provides information about noise characteristics, which then guides the selective modification of singular vectors. This feedback loop enables continuous optimization of noise removal while preserving defect signals, improving upon the static reference image approaches of conventional methods.
Solution Approach 2:
The patent changes the parameter representation from conventional pixel-space image comparison to tensor-space decomposition with singular values. By operating in this transformed parameter space, the method can selectively modify specific singular vectors to remove noise while preserving defect information, achieving superior sensitivity compared to traditional reference image techniques.
3Manufacturing precision
If image subtraction is performed after sub-pixel alignment to remove intrinsic patterns, then background removal is achieved, but additional noise from reference imagery decreases defect detection sensitivity
Solution Approach 1:
The patent performs preliminary decomposition and noise removal on multiple difference images before final defect detection. By pre-processing the images through tensor decomposition and selectively modifying singular vectors to remove noise components, the method prepares cleaner input data for subsequent defect analysis, improving overall detection sensitivity.
Solution Approach 2:
The patent creates a composite processing approach by combining multiple difference images into a tensor structure, applying decomposition to generate modified difference images with reduced noise. This composite methodology merges information from multiple images while systematically removing noise, achieving better signal-to-noise ratio than simple pairwise subtraction.
Data Source
AI summary
A sample characterization system is disclosed. In embodiments, the sample characterization system includes a controller communicatively coupled to an inspection sub-system, the controller including one or more processors configured to execute a set of program instructions stored in memory, the set of program instructions configured to cause the one or more processors to: acquire one or more target image frames of a sample; generate a target tensor with the one or more acquired target image frames; perform a first set of one or more decomposition processes on the target tensor to generate one or more reference tensors including one or more reference image frames; identify one or more differences between the one or more target image frames and the one or more reference image frames; and determine one or more characteristics of the sample based on the one or more identified differences.