Plasma Chamber State Quantification for Dynamic Substrate Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing substrate processing systems struggle to quantify the state inside a plasma processing chamber effectively, leading to inconsistencies in the quality of plasma-processed substrates.
Innovation Solution
A substrate processing system that includes an acquiring unit, a learning unit, and a quantification unit to analyze time-series data, generating learned abnormal value detection models and calculating divergence degrees to adjust control knobs based on chamber conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a plurality of types of time-series data are acquired and analyzed to quantify chamber state, then substrate quality homogeneity is improved, but system complexity increases
Solution Approach 1:
The system segments the complex quantification task into distinct phases: a learning phase where abnormal value detection models are trained on historical data, and a quantification phase where these models are applied to current data. This segmentation allows the system to handle complexity through structured, modular processing rather than attempting to analyze all data simultaneously.
Solution Approach 2:
The system performs preliminary actions by pre-training abnormal value detection models during a learning phase using historical time-series data from multiple sensors. These pre-trained models are then reused during subsequent quantification phases, eliminating the need to reprocess all historical data and reducing computational complexity while maintaining high substrate quality homogeneity.
2Measurement precision
If learned abnormal value detection models are generated by calculating data density of multiple time-series data types, then chamber state quantification precision is improved, but processing time increases
Solution Approach 1:
The system performs data density calculations and model training in advance during a learning phase, storing the resulting abnormal value detection models for rapid reuse. During actual production quantification phases, the pre-trained models are applied directly to new data without repeating the computationally intensive training process, thus maintaining high measurement precision while significantly reducing processing time.
Solution Approach 2:
The system creates simplified representations (copies) of complex chamber states through trained detection models. Instead of processing raw time-series data from multiple sensors each time, the system uses these model copies to quickly assess chamber conditions, maintaining quantification precision while reducing the computational burden and processing time.
3Reliability
If divergence degree is calculated from multiple time-series data types using learned models, then abnormal value detection accuracy is improved, but computational load increases
Solution Approach 1:
The system performs computationally intensive model training and data density calculations during a preliminary learning phase, then stores the resulting detection models. During subsequent abnormal value detection operations, the system only needs to apply these pre-trained models to calculate divergence degrees, dramatically reducing computational load while maintaining high detection accuracy through the use of the pre-processed model knowledge.
Data Source
AI summary
A substrate processing system includes: an acquiring unit that acquires a plurality of types of time-series data for each time when a plasma processing is performed on a substrate; a learning unit that generates a number of learned abnormal value detection models corresponding to a number of the plurality of types of the time-series data by calculating a data density of each of the plurality of types of time-series data acquired in a first phase; and a quantification unit that quantifies a state in a processing space in a second phase by inputting the plurality of types of time-series data acquired in the second phase into the corresponding learned abnormal value detection models, respectively, and calculating a divergence degree from the plurality of types of the time-series data acquired in the first phase.