Neural Network Metrology for Semiconductor Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor metrology techniques face inefficiencies and reliability issues in accurately measuring critical parameters due to the need for exhaustive grid-based searches and iterative gradient-based methods, which are time-consuming and prone to local minima problems.
Innovation Solution
The implementation of a neural network-based method that predicts near-optimal values of fixed parameters using training data, combining a learned neural network with gradient-based search to enhance efficiency and reliability, allowing for faster and more accurate measurement adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If grid-based search methods are used to optimize fixed parameters for measurement accuracy, then measurement precision is improved, but productivity deteriorates due to exhaustive search requirements
Solution Approach 1:
The patent applies preliminary action by using a neural network to predict near-optimal fixed parameter values before performing the actual optimization search. The neural network is trained offline on a dataset of spectra and corresponding optimal parameters, so that during actual measurement optimization, the system starts from predicted near-optimal values rather than performing exhaustive grid searches from scratch. This preliminary prediction significantly reduces the search space and computational requirements for achieving high measurement accuracy.
2Measurement precision
If iterative gradient-based search methods are used to optimize fixed parameters, then measurement precision is improved, but loss of time increases due to repeated numerical gradient computations
Solution Approach 1:
The neural network performs preliminary computation of gradient information during its training phase, storing learned patterns of parameter relationships. During actual optimization, the system leverages these pre-learned patterns to avoid repeated numerical gradient computations, significantly reducing optimization time while maintaining the ability to achieve high measurement precision through guided search from predicted near-optimal starting points.
3Productivity
If neural network prediction is used to estimate fixed parameter values, then productivity is improved, but reliability deteriorates due to potential prediction errors
Solution Approach 1:
The patent implements feedback by using the neural network's predicted fixed parameter values as initial guesses for iterative optimization algorithms. The actual measurement spectra are then obtained using these predicted values, and the results are fed back to refine the parameter estimates. This feedback loop allows the system to correct any prediction errors made by the neural network, ensuring reliable and accurate final parameter values while maintaining the productivity benefits of starting from near-optimal predictions.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach significantly improves the efficiency and reliability of parameter measurement accuracy by efficiently finding near-optimal values of fixed parameters without repeated numerical gradient computations, ensuring reliable optimization of critical parameter measurements.
Implementation Method 1
The neural network is trained based on one or more of the critical parameters and a low-dimensional real-valued vector associated with a spectrum. The spectrum may be a spectroscopic ellipsometry spectrum or a specular reflectance spectrum.
Implementation Method 2
The spectrum may be a spectroscopic ellipsometry spectrum or a specular reflectance spectrum.
Data Source
AI summary
Machine learning techniques are used to predict values of fixed parameters when given reference values of critical parameters. For example, a neural network can be trained based on one or more critical parameters and a low-dimensional real-valued vector associated with a spectrum, such as a spectroscopic ellipsometry spectrum or a specular reflectance spectrum. Another neural network can map the low-dimensional real-valued vector. When using two neural networks, one neural network can be trained to map the spectra to the low-dimensional real-valued vector. Another neural network can be trained to predict the fixed parameter based on the critical parameters and the low-dimensional real-valued vector from the other neural network.