Hidden-Layer Fusion for Wafer-Aware Semiconductor Defect Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor defect classification methods, including decision trees and neural networks, face challenges such as the inability to utilize wafer-level information, require substantial labeled data, and are computationally complex, leading to inaccurate defect classification that can cause manufacturing delays and increased costs.
Innovation Solution
A system combining a deep learning model with hidden layers and random forest classifiers, utilizing hidden layer outputs in conjunction with specimen-specific inputs, enhances defect classification accuracy by leveraging the strengths of both methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current defect classification methods (decision trees, neural networks) are used, then the classification process can be performed, but the classification accuracy is insufficient due to inability to utilize wafer-level information and requirement for substantial labeled data
Solution Approach 1:
The patent combines multiple information sources including hidden layer outputs from deep learning models, specimen-specific inputs, and wafer-level information into a unified classification framework. This merging of previously separate data streams enables comprehensive defect classification that leverages both local defect characteristics and global wafer context, thereby improving classification accuracy while reducing reliance on substantial labeled data.
Solution Approach 2:
The patent introduces an additional component that acts as an intermediary between the deep learning model's hidden layer outputs and the final classification decision. This intermediary component processes and integrates hidden layer outputs with specimen-specific inputs and wafer-level information, enabling effective information fusion that improves classification accuracy without requiring extensive labeled training data.
2Loss of information
If deep learning models with hidden layers are used, then more information can be extracted from detector output, but the computational complexity increases
Solution Approach 1:
The patent extracts and utilizes only the relevant hidden layer outputs from the deep learning model, rather than processing the entire model output. By selectively extracting specific hidden layer representations and combining them with specimen-specific inputs through an additional component, the system achieves effective information extraction while managing computational complexity through focused processing of key features.
3Measurement precision
If more labeled data is collected for training, then classification accuracy may improve, but the time and resources required for data preparation increase
Solution Approach 1:
The patent enables the classification system to leverage wafer-level information and specimen-specific inputs that are already available in the manufacturing process, rather than requiring separate data collection and labeling efforts. The additional component processes readily available information including hidden layer outputs and contextual data, allowing the system to achieve high classification accuracy using existing process data without substantial additional time investment for data preparation.
Data Source
AI summary
Methods and systems for determining information for a specimen are provided. One system includes one or more components executed by a computer system including a deep learning (DL) model configured for determining information for a specimen from output generated for the specimen by at least one of one or more detectors of an output generation subsystem. The DL model includes hidden layers configured for generating hidden layer output. The one or more components also include an additional component configured for determining additional information for the specimen from the hidden layer output generated by at least one of the hidden layers in combination with input specific to the specimen. In some embodiments, the information of the first DL and its hidden layer are used as inputs to a second network that then also uses non-image based information of the defects to further distill the purity of DOI vs nuisance separation.