Neural Network Wafer Parameter Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional models for predicting wafer parameters, such as overlay error in semiconductor manufacturing, face limitations due to dependence on unpredictable variables and difficulty in identifying complex relationships between inputs and outputs, leading to inefficiencies in lot dispositioning and quality assurance processes.
Innovation Solution
A neural network-based method is employed to predict wafer parameters by training the network with target quality metrics, alignment metrics, and process metrics, allowing it to adapt and improve its predictive capabilities, thereby reducing reliance on optical properties of lithography tools and enhancing the accuracy of overlay error estimation across the wafer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional models are used to predict overlay error, then the prediction process can be performed, but the accuracy is reduced due to dependence on unpredictable variables and inability to identify complex relationships
Solution Approach 1:
The patent replaces conventional mechanical/mathematical modeling approaches with a neural network-based system. The neural network learns complex non-linear relationships from training data without requiring explicit mathematical models, thereby eliminating dependence on unpredictable variables and improving both accuracy and reliability of overlay error predictions.
Solution Approach 2:
The neural network performs self-learning and self-adjustment through the training process. It automatically identifies complex relationships between process parameters and overlay errors by adjusting its internal weights and biases, without requiring manual model specification or intervention, thus improving prediction reliability.
2Productivity
If conventional models are used for wafer parameter prediction, then the process can proceed, but time is lost due to inefficiencies in lot dispositioning and quality assurance
Solution Approach 1:
The patent replaces time-consuming conventional modeling and measurement processes with a neural network system that can rapidly predict multiple wafer parameters simultaneously. This substitution dramatically reduces the time required for lot dispositioning and quality assurance while improving productivity.
Solution Approach 2:
The neural network is trained in advance on historical data to learn the relationships between process parameters and wafer outcomes. During actual production, this pre-trained network can immediately provide accurate predictions without requiring time-consuming real-time analysis, thus reducing loss of time in quality assurance processes.
3Measurement precision
If target structures are used to measure overlay error at specific locations, then localized measurements can be obtained, but approximations at other locations introduce additional error
Solution Approach 1:
The neural network serves multiple functions simultaneously: it predicts overlay error at measured locations with high accuracy while also providing reliable approximations at unmeasured locations. This multi-functional capability ensures both localized measurement precision and overall wafer precision without the trade-off present in conventional methods.
Solution Approach 2:
The patent replaces the physical measurement system (target structures and localized measurements) with an intelligent prediction system. The neural network uses process parameters and training data to predict overlay error across the entire wafer surface, eliminating the need for physical target structures and providing comprehensive precision without localization limitations.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
Apparatus and methods are provided for predicting a plurality of unknown parameter values (e.g. overlay error or critical dimension) using a plurality of known parameter values. In one embodiment, the method involves training a neural network to predict the plurality of parameter values (114, 700, 800, 900). In other embodiments, the prediction process does not depend on an optical property of a photolithography tool. Such predictions may be used to determine wafer lot disposition (114).