ML Hotspot Ranking via Printed Substrate Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current device manufacturing processes, particularly in semiconductor production, face challenges in accurately predicting and correcting patterning defects due to inaccuracies in computational lithography models, leading to inefficiencies in identifying critical hotspots and optimizing process parameters.
Innovation Solution
A method involving machine learning models, such as convolutional neural networks, is trained using simulated patterns and measurement data to correct characteristic inaccuracies and improve hotspot ranking, enabling more precise defect prediction and resource allocation in the patterning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If computational lithography models are used to predict patterning defects, then defect detection capability is improved, but model inaccuracies lead to incorrect hotspot ranking and reduced prediction reliability
Solution Approach 1:
The patent implements feedback by measuring actual printed patterns on substrates and using these measurements to train and refine machine learning models. The measured data from real patterning processes feeds back into the model training process, allowing the system to learn from actual outcomes and improve prediction accuracy iteratively
Solution Approach 2:
The patent introduces machine learning models as intermediaries between computational lithography models and actual defect detection. These ML models are trained on measured data to correct inaccuracies in computational models, serving as a bridge that translates theoretical predictions into more accurate real-world predictions
2Measurement precision
If comprehensive metrology measurements are performed on all patterns, then defect detection accuracy is improved, but measurement time and resource consumption increase significantly
Solution Approach 1:
The patent extracts and identifies only the most critical patterns (hotspots) that are likely to contain defects based on measured data from previous substrates. By taking out only these high-priority patterns for detailed metrology measurement, the system avoids measuring all patterns, thereby reducing measurement time while maintaining high defect detection accuracy
Solution Approach 2:
The patent applies partial action by performing comprehensive metrology measurements only on a subset of critical patterns rather than all patterns. The machine learning model identifies which patterns require detailed measurement, allowing the system to apply excessive measurement effort only where necessary to detect potential defects
3Manufacturing precision
If machine learning models are trained with measured data from printed substrates, then model accuracy is improved, but data collection and processing complexity increase
Solution Approach 1:
The patent creates a universal data collection framework that serves multiple functions: collecting measurement data for model training, identifying hotspots for prioritized measurement, and providing feedback for iterative improvement. This multi-functional approach consolidates what could be separate complex processes into an integrated system
Data Source
AI summary
Methods for training a process model and determining ranking of simulated patterns (e.g., corresponding to hot spots). A method involves obtaining a training data set including: (i) a simulated pattern associated with a mask pattern to be printed on a substrate, (ii) inspection data of a printed pattern imaged on the substrate using the mask pattern, and (iii) measured values of a parameter of the patterning process applied during imaging of the mask pattern on the substrate; and training a machine learning model for the patterning process based on the training data set to predict a difference in a characteristic of the simulated pattern and the printed pattern. The trained machine learning model can be used for determining a ranking of hot spots. In another method a model is trained based on measurement data to predict ranking of the hot spots.