Etching Condition Learning for Uniform Film Thickness Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity of the etching process in semiconductor manufacturing, where the nozzle movement is time-series data with increasing dimensions, makes it difficult to optimize learning models for film thickness uniformity, and the number of nozzle works suitable for a target processing amount is not limited to one, requiring costly and time-consuming trial and error.
Innovation Solution
A training device that acquires and reduces the dimensionality of time-varying processing conditions using an autoencoder, generating a learning model to predict film thickness differences before and after etching, and determines optimal nozzle movements through machine learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If the work of moving a nozzle is complicated to achieve uniform film thickness, then the processing quality is improved, but the number of dimensions of training data increases exponentially, making it difficult to optimize the learning model
Solution Approach 1:
The patent extracts and separates the time-series nozzle movement data into individual time points, treating each time point as an independent input feature. This extraction approach reduces the complexity of the training data by breaking down the complex time-series workflow into manageable discrete elements that can be processed more efficiently by the learning model.
Solution Approach 2:
The patent segments the continuous nozzle movement process into discrete time points, where each time point represents a specific position of the nozzle. This segmentation transforms the complex continuous workflow into discrete, manageable units that can be used as training data features, reducing the overall dimensionality while preserving the essential information needed for controlling film thickness uniformity.
2Measurement precision
If the number of sampling intervals increases to capture complicated nozzle movement, then the accuracy of process control is improved, but the number of data sets required for machine learning increases exponentially
Solution Approach 1:
The patent performs preliminary processing of the nozzle movement data by extracting key time points before machine learning training. By pre-processing the time-series data to identify and extract critical position information, the system reduces the amount of data needed for training while maintaining the accuracy required for precise nozzle control and film thickness uniformity.
Solution Approach 2:
The patent extracts essential position information from the time-series nozzle movement data at specific time points. This extraction process isolates the critical features needed for control accuracy while discarding redundant information, thereby reducing the total number of data sets required for machine learning training.
3Manufacturing precision
If trial and error by an engineer is used to determine optimal nozzle work, then the processing conditions can be optimized, but the process becomes costly and time consuming
Solution Approach 1:
The patent replaces the mechanical trial-and-error process with an automated machine learning system. The learning model automatically determines optimal nozzle movement patterns by training on historical data, eliminating the need for manual trial and error experiments. This substitution significantly reduces both the time and cost associated with optimizing processing conditions while maintaining or improving film thickness uniformity.
Solution Approach 2:
The system enables self-service optimization where the machine learning model automatically determines optimal nozzle work parameters without requiring engineer intervention for trial and error experiments. The model learns from past data and autonomously identifies the best processing conditions, freeing engineers from time-consuming manual optimization tasks.
Data Source
AI summary
A training device includes a hardware processor, wherein the hardware processor acquires a processing amount indicating a difference between a film thickness obtained before a process and a film thickness obtained after the process in regard to a film formed on a substrate, after a substrate processing apparatus is driven according to processing conditions including a variable condition that varies over time, converts the variable condition such that a dimension count is reduced, and generates a learning model, which executes machine learning using training data and predicts a processing amount indicating a difference between a film thickness obtained before a process and a film thickness obtained after the process, with the training data including a conversion result of the variable condition converted by the compression device, and the processing amount corresponding to the processing conditions.