Processing Condition Search Using Transfer Learning Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for deriving processing conditions in semiconductor processes, particularly through machine learning, suffer from low accuracy in I/O models, leading to increased process development periods and costs due to the need for extensive verification experiments and the use of inappropriate reference processing data.
Innovation Solution
A processing condition search device and method that employs transfer learning and teachered learning to estimate I/O models, utilizing a target processing database and a reference processing database, with a transferability determination unit to select the appropriate learning method based on similarity and accuracy criteria, thereby predicting optimal processing conditions efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If machine learning is used to derive processing conditions, then processing condition derivation can be automated, but the accuracy of the I/O model is low leading to incorrect predictions
Solution Approach 1:
The patent introduces a transfer learning mechanism that uses a reference I/O model (built from reference processing data) as an intermediary to improve the target I/O model. The reference model acts as a mediator that transfers knowledge from reference processes to the target process, enhancing the accuracy of the target I/O model without reducing automation.
Solution Approach 2:
The patent performs preliminary action by building a reference I/O model using reference processing data before deriving the target processing conditions. This preliminary model construction and transfer learning process prepares the system in advance to improve the target I/O model accuracy, reducing the need for extensive verification experiments later.
2Manufacturing precision
If extensive verification experiments are performed to ensure accuracy, then processing condition accuracy improves, but process development time increases
Solution Approach 1:
The patent implements a feedback mechanism where the transferability determination unit evaluates the quality of transfer learning results and determines whether to use the transferred I/O model or revert to traditional methods. This feedback loop ensures accuracy by selectively applying transfer learning only when it proves beneficial, avoiding unnecessary verification experiments while maintaining processing condition accuracy.
3Measurement precision
If reference processing data is used to improve model accuracy, then I/O model accuracy improves, but inappropriate reference data can degrade performance
Solution Approach 1:
The patent introduces dynamic adaptability by making the I/O model construction process flexible and adjustable. The system dynamically determines whether to use transfer learning based on the evaluated transferability of reference data. This dynamic approach allows the system to adapt to different scenarios - using reference data when appropriate and avoiding it when inappropriate, thereby maintaining both accuracy and reliability.
4Reliability
If traditional processing condition derivation methods are used, then verification experiments are required, but the process development period and costs increase
Solution Approach 1:
The patent uses copying by creating a reference I/O model from reference processing data and copying its knowledge structure to improve the target I/O model. This copying mechanism allows the system to leverage existing knowledge from reference processes, reducing the need for extensive new verification experiments and accelerating process development while maintaining reliability through the transferability evaluation framework.
Data Source
AI summary
Target processing conditions are searched while reducing a process development time and process development costs. A processing condition search device includes a learning database including a processing database storing target processing data for a target process and a reference processing database storing reference processing data for a reference process, a teachered learning execution unit that estimates an I/O model of the target process, which is an I/O model between a target description variable and a target objective variable, a transfer learning execution unit that estimates the I/O model of the target process using a reference I/O mode between a reference description variable and a reference objective variable, and the target processing data, and a transferability determination unit that determines whether the teachered learning execution unit or the transfer learning execution unit is used to estimate the I/O model of the target process.