Inspection Mode Selection Using Masked Deep Learning Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining inspection modes in semiconductor manufacturing require extensive training of convolutional neural networks for every possible combination, leading to high computational costs and inefficiencies, and the Greedy search approach does not guarantee the best mode combination.
Innovation Solution
A deep learning-based system that uses a setup convolutional neural network model to separately perform defect detection for each mode, generate masked outputs, and identify candidate modes with significant differences, reducing the need for exhaustive training and resource consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Brute Force approach is used to train CNN models for every possible mode combination, then the best mode combination can be guaranteed, but computational time and resources increase exponentially
Solution Approach 1:
The patent segments the exhaustive mode combination search into individual mode evaluations. Instead of training CNNs for all possible combinations (2^N-1), the system evaluates each mode independently by training separate CNNs for each mode and selecting the top K modes based on individual performance metrics, reducing computational complexity from exponential to linear scale
Solution Approach 2:
The patent applies partial action by training CNN models for only the necessary individual modes rather than all possible combinations. The system trains N individual CNNs (one per mode) and selects top K modes, performing partial training work that suffices for achieving good mode selection without the excessive computational burden of complete exhaustive search
2Productivity
If Greedy search approach is used to select top K modes, then training time is reduced, but the best mode combination is not guaranteed
Solution Approach 1:
The patent incorporates feedback mechanisms where each mode's CNN model is evaluated based on its defect detection performance, and this feedback is used to rank and select the top K modes. The system uses sensitivity metrics and detection accuracy as feedback signals to iteratively refine mode selection, ensuring both efficiency and reliability
3Measurement precision
If multiple CNN models are trained for every mode combination, then comprehensive defect detection is achieved, but resource consumption increases substantially
Solution Approach 1:
The patent extracts and evaluates the contribution of each individual mode to defect detection sensitivity. By separating the evaluation of each mode's defect detection capability and selecting only the top K contributing modes, the system reduces resource consumption while maintaining comprehensive defect detection through the combined strength of selected modes
Data Source
AI summary
Methods and systems for determining information for a specimen are provided. One system includes a computer subsystem and one or more components executed by the computer subsystem that include a setup deep learning (DL) model configured for separately performing defect detection for a specimen based on output generated for the specimen by each of two or more modes of an inspection system, respectively, and separately re-performing defect detection for the specimen based on masked output generated for each of the modes, respectively. The computer subsystem determines a difference between results of separately performing and separately re-performing the defect detections for each of the modes, respectively, and identifies a subset of the modes for which the difference is larger than other modes as candidate mode(s) for inspection of the specimen.