Plasma Etching Regression Model Using Sparse Wavelength Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating regression equations in plasma processing apparatuses face challenges in achieving accurate and efficient estimation of etching rates using light emission data, often resulting in overfitting and complex models that are difficult to interpret, especially when data quality and quantity are insufficient.
Innovation Solution
A learning method that preprocesses light emission data, sets constraints for machine learning to generate a regression equation, selects target wavelengths, and incorporates additional sensor data to create a regression equation that accurately represents the relationship between etching rates and light emission data, using sparse modeling and genetic programming to ensure model interpretability and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning is performed without constraints on light emission data, then the model can capture complex relationships, but overfitting occurs and model interpretability decreases
Solution Approach 1:
The patent applies parameter changes by transforming the original light emission data through wavelength selection and preprocessing. Specific wavelengths are selected based on their correlation with etching rates, and the data is transformed into a format suitable for constrained machine learning, thereby improving prediction accuracy while maintaining model interpretability through sparse modeling
Solution Approach 2:
The patent extracts only the most relevant wavelengths from the full spectrum of light emission data. By selecting specific wavelengths that have high correlation with etching rates and excluding irrelevant ones, the model achieves better accuracy with fewer parameters, reducing overfitting while maintaining interpretability
2Measurement precision
If more sensor data is incorporated into machine learning, then prediction accuracy improves, but data processing complexity increases
Solution Approach 1:
The patent merges multiple data sources including light emission data at selected wavelengths, other sensor data from the plasma processing apparatus, and etching rate measurements. This combination of heterogeneous data sources improves prediction accuracy by capturing multiple aspects of the plasma process while the constraint-based modeling keeps processing manageable
3Loss of time
If machine learning is performed with insufficient data, then processing time is reduced, but model accuracy and reliability decrease
Solution Approach 1:
The patent performs preliminary actions by pre-selecting wavelengths and preprocessing data before the actual machine learning process. This preparation ensures that even with limited data, the model learns from the most relevant features, improving reliability without requiring extensive training data or time
Data Source
AI summary
There is provided a learning method. The method includes performing preprocessing on light emission data in a chamber of a plasma processing apparatus, setting a constraint for generating a regression equation representing a relationship between an etching rate of the plasma processing apparatus and the light emission data, selecting a learning target wavelength from the light emission data subjected to the preprocessing, and receiving selection of other sensor data different from the light emission data. The method further includes generating a regression equation based on the set constraint while using, as learning data, the selected wavelength, the received other sensor data, and the etching rate, and outputting the generated regression equation.