Semiconductor Metrology Using ML for Spectral Variability Robustness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor metrology methods face challenges in achieving accuracy, process robustness, precision, and throughput due to the complexity and tight process limits of modern semiconductor manufacturing.
Innovation Solution
The implementation of machine learning (ML) methods based on supervised learning to predict metrology parameters using optical signals, incorporating variability data to minimize loss functions and improve model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If hardware improvements are made to optimize move and measure time and tool stability, then throughput and sampling rates are improved, but device complexity and cost increase
Solution Approach 1:
The patent replaces hardware-based metrology optimization with a software-based machine learning approach. Instead of improving physical measurement tools to increase throughput, the invention uses ML models to predict metrology parameters from existing measurements, achieving high throughput without additional hardware complexity.
Solution Approach 2:
The patent introduces machine learning models as an intermediary between raw optical signals and final metrology parameters. This intermediary layer processes measurements computationally rather than requiring more complex measurement hardware, thereby improving throughput while avoiding hardware complexity increases.
2Measurement precision
If machine learning models are trained with variability data from multiple sources, then measurement precision and robustness are improved, but loss function complexity and training time increase
Solution Approach 1:
The patent performs preliminary actions by collecting and organizing variability data from multiple sources (different tools, measurement conditions, spectral variations) before training the ML model. This pre-processing of variability data allows the model to learn robust patterns without requiring excessively complex architectures, as the complexity is managed in the data preparation phase rather than the model structure.
Solution Approach 2:
The patent changes parameters in the loss function to balance model complexity with measurement precision. By incorporating variability terms into the loss function with appropriate weighting, the model learns to account for multiple sources of variation without becoming overly complex, achieving robust predictions while maintaining tractable model sizes.
3Reliability
If spectral variability is incorporated into the loss function, then process robustness is improved, but computational time and training complexity increase
Solution Approach 1:
The patent applies partial action by incorporating spectral variability terms into the loss function at a controlled level. Rather than exhaustively modeling all possible sources of variability with equal weight, the approach selectively includes the most significant variability sources, achieving improved process robustness without the full computational burden of exhaustive variability modeling.
Data Source
AI summary
A semiconductor metrology system including a spectrum acquisition tool for collecting, using a first measurement protocol, baseline scatterometric spectra on first semiconductor wafer targets, and for various sources of spectral variability, variability sets of scatterometric spectra on second semiconductor wafer targets, the variability sets embodying the spectral variability, a reference metrology tool for collecting, using a second measurement protocol, parameter values of the first semiconductor wafer targets, and a training unit for training, using the collected spectra and values, a prediction model using machine learning and minimizing an associated loss function incorporating spectral variability terms, the prediction model for predicting values for production semiconductor wafer targets based on their spectra.