Substrate Etching Model for Compressed Nozzle Motion Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity of nozzle movement in etching processes for substrate film thickness adjustment leads to increased dimensions in training data, making it difficult to optimize learning models, and the number of suitable nozzle works for a target processing amount is not limited to one, necessitating costly and time-consuming trial and error.
Innovation Solution
A training device and method that acquire experimental data, convert variable nozzle conditions into compressed data, and generate a learning model through machine learning to predict film thickness differences, enabling efficient determination of multiple processing conditions for substrate processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If the sampling intervals are shortened to complicate the nozzle work, then the precision of film thickness adjustment is improved, but the number of dimensions of training data increases exponentially
Solution Approach 1:
The patent extracts and separates the variable condition data from the training data structure. Instead of using raw high-dimensional time-series nozzle position data, the invention extracts compressed representations that capture the essential processing characteristics while reducing dimensionality. This allows maintaining manufacturing precision without the exponential increase in data complexity.
Solution Approach 2:
The patent transforms the data from high-dimensional time-series space to a lower-dimensional feature space by introducing compressed data representations. This dimensionality change allows the learning model to process complicated nozzle work patterns efficiently without requiring exponentially large datasets, resolving the contradiction between precision and complexity.
2Adaptability or versatility
If multiple nozzle works are considered for a target processing amount, then the adaptability of processing conditions is improved, but the time and cost for optimization increases
Solution Approach 1:
The patent performs preliminary action by pre-processing the variable conditions into compressed data representations before training the learning model. This preprocessing step organizes multiple nozzle work patterns into structured compressed data, allowing the model to efficiently learn and evaluate multiple processing conditions simultaneously without requiring time-consuming trial and error optimization.
Solution Approach 2:
The patent replaces the mechanical trial-and-error optimization process with a learning model-based system. Instead of manually testing multiple nozzle work configurations, the invention uses machine learning to automatically evaluate and determine optimal processing conditions from compressed data, significantly reducing optimization time while maintaining adaptability.
3Measurement precision
If the number of data sets for machine learning is increased to handle higher dimensions, then the accuracy of the learning model is improved, but the cost and time for data collection increases
Solution Approach 1:
The patent changes the parameters of the training data by transforming raw variable conditions into compressed data representations. This parameter transformation maintains the essential information needed for accurate learning model training while reducing the apparent dimensionality, thereby achieving high accuracy without requiring an exponentially large number of training datasets.
Data Source
AI summary
A training device includes an experimental data acquirer that acquires a first processing amount indicating a difference between a film thickness obtained before a process for a film and a film thickness obtained after the process for the film, after the process for the film is executed according to processing conditions including a variable condition indicating a relative position of a nozzle with respect to a substrate, with the relative position varying over time, a converter that converts the variable condition into compressed data and a model generator that generates a learning model, with the learning model executing machine learning using training data that includes the compressed data and the first processing amount corresponding to the processing conditions and predicting a second processing amount.