Weighted Feature Training for Stable Semiconductor ML Predictions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning models used in semiconductor manufacturing exhibit fluctuations in predictive data, leading to inaccurate predictions and the production of defective substrates due to factors like learning rate, number of layers, and validation set size.
Innovation Solution
A system that determines and applies weights to training data based on defined or inferred relationships between manufacturing parameters, generating a characteristic sequence and using correlation matrices to optimize the machine-learning model, thereby reducing fluctuations in predictive data without compromising accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine-learning models are trained without weighted features, then training is simpler and faster, but the output data exhibits fluctuations leading to prediction errors
Solution Approach 1:
The patent applies parameter changes by introducing weights to training features based on their correlation with manufacturing parameters. This transforms the standard training approach by modifying the input data parameters (adding weights) to improve prediction reliability and reduce output fluctuations, directly resolving the contradiction between simplicity and accuracy.
2Reliability
If resource-intensive operations are used to optimize the machine-learning model, then model accuracy improves, but system downtime increases
Solution Approach 1:
The patent implements preliminary action by pre-calculating correlation matrices between manufacturing parameters and features, and determining optimal weights before training begins. This preparation work is done in advance during periods of lower demand, allowing the actual training to proceed more efficiently with reduced downtime, thus improving system availability while maintaining accuracy.
3Manufacturing precision
If standard training methods are used, then the training process is more straightforward, but fluctuations in predictive data lead to defective substrate production
Solution Approach 1:
The patent employs feedback by using correlation matrices to determine weights based on the relationship between manufacturing parameters and features. This feedback mechanism identifies which features most strongly correlate with manufacturing parameters and adjusts their weights accordingly, improving predictive data consistency and reducing defects through data-driven optimization.
Data Source
AI summary
An electronic device manufacturing system configured to receive, by a processor, input data reflecting a feature related to a manufacturing process of a substrate. The manufacturing system is further configured to generate a characteristic sequence defining a relationship between at least two manufacturing parameters, and determine a relationship between one or more variables related to the feature and the characteristic sequence. The manufacturing system is further configured to determine a weight based on the determined relationship and apply the weight to the feature. The manufacturing system is further configured to train a machine-learning model in view of the weighted feature.