Spectrum-Based Structure Measurement for IC Metrology Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Measuring complex structures in integrated circuits requires significant time and cost, often necessitating destruction of the circuit, which is inefficient and detrimental to productivity.
Innovation Solution
A method and system utilizing a machine learning model trained on simulation data, with sub-models generated from a first model, to estimate structure dimensions from measured spectra without destructive testing, leveraging transfer learning to enhance accuracy with limited measured data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional measurement methods are used to measure complex structures in integrated circuits, then measurement accuracy can be achieved, but measurement time and cost increase significantly and the circuit must be destroyed
Solution Approach 1:
The patent replaces physical/mechanical measurement methods with optical measurement. Instead of using electron microscopes or other complex physical measurement equipment that require sample preparation and destruction, the invention uses optical spectra analysis to measure structure thickness and dimensions non-destructively, thereby maintaining measurement accuracy while dramatically improving productivity
Solution Approach 2:
The patent introduces machine learning models as an intermediary between optical spectrum measurements and structure parameter determination. The ML models are trained on simulation data and then fine-tuned with limited measured data to accurately predict structure dimensions from optical spectra, enabling accurate non-destructive measurement without requiring complex physical measurement systems
2Reliability
If conventional measurement methods are used, then structure verification can be performed, but time and cost resources are consumed excessively
Solution Approach 1:
The patent performs preliminary training of machine learning models using extensive simulation data before actual measurements. This preliminary action creates a robust model that can quickly and accurately predict structure parameters from optical spectra, eliminating the need for time-consuming conventional measurement procedures while maintaining verification reliability
Solution Approach 2:
The invention substitutes time-consuming physical measurement processes with rapid optical spectroscopy combined with pre-trained machine learning analysis, achieving the same verification reliability in a fraction of the time
3Productivity
If machine learning models are trained only on simulation data, then model generation is fast, but measurement accuracy decreases due to limited measured data
Solution Approach 1:
The patent performs preliminary training of machine learning models using extensive simulation data to establish a robust baseline model. This preliminary action creates a model that captures the fundamental relationships between optical spectra and structure parameters, which can then be efficiently fine-tuned with limited measured data to achieve high accuracy
Solution Approach 2:
The patent changes the training approach from using only one type of data to a two-stage process: first training on simulation data (virtual spectra and structure parameters), then fine-tuning on measured data (actual optical spectra and corresponding structure measurements). This parameter change in training data composition enables the model to achieve high accuracy while maintaining training efficiency
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for measuring a structure based on a spectrum, includes obtaining a first model that includes a first sub-model and a second sub-model following the first sub-model and is trained based on simulation data, generating a second model including a third sub-model identical to the first sub-model, training the second model based on sample spectrum data generated by measuring spectra of sample structures, and estimating, based on the trained second model, the structure from measured spectrum data generated by measuring a spectrum of the structure.