Semiconductor Language Model Training for Actionable Failure Summaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for summarizing semiconductor equipment data using artificial neural networks are limited in generating summaries that can assist with failure actions, as they primarily focus on overall context rather than actionable insights.
Innovation Solution
A method for training an artificial neural network using semiconductor data, which involves preprocessing the data, generating tokens, and performing unsupervised learning to create a semiconductor language model. This model is then updated using supervised learning with correction data to improve the accuracy of failure data summarization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If existing methods for summarizing semiconductor equipment data using artificial neural networks are used, then overall context can be captured, but actionable insights for failure actions cannot be generated
Solution Approach 1:
The patent segments the summarization task into multiple specialized neural network models, each trained to extract specific types of actionable information (failure causes, failure actions, maintenance actions) from semiconductor equipment data. This segmentation allows each model to focus on specific aspects of failure analysis, thereby capturing actionable insights that a single general-purpose model would miss.
Solution Approach 2:
The patent changes the training parameters and objectives of the neural network models by using supervised learning with specifically curated datasets containing labeled failure actions and maintenance actions. This parameter change enables the models to generate actionable insights rather than just general summaries, directly addressing the information loss problem.
2Measurement precision
If a language model is trained using unsupervised learning on semiconductor equipment data, then a semiconductor language model can be generated, but accuracy in failure data summarization is limited
Solution Approach 1:
The patent performs preliminary unsupervised learning to generate a semiconductor language model that captures general patterns in equipment data. This preliminary action creates a foundation model that can then be efficiently refined through supervised learning on smaller, labeled datasets, reducing the overall training time while improving accuracy for failure data summarization.
Solution Approach 2:
The patent implements a continuous training process where the language model undergoes unsupervised learning first, then supervised learning, and can be iteratively refined. This continuity allows the model to maintain and improve its failure data summarization accuracy over time without requiring complete retraining, thus managing training time effectively.
3Adaptability or versatility
If semiconductor equipment data is preprocessed and tokens are generated using a mapping table, then data can be prepared for language model training, but data processing complexity increases
Solution Approach 1:
The patent introduces a mapping table as an intermediary component that bridges raw semiconductor equipment data and the language model's token representation system. This intermediary handles the complexity of data preprocessing by providing a standardized conversion mechanism, thereby enabling model training capability without directly increasing the overall system complexity.
Solution Approach 2:
The patent creates a tokenized copy of the semiconductor equipment data through the mapping table, which serves as a simplified representation suitable for language model training. This copying approach allows the model to work with processed data while the original data structure remains intact, managing processing complexity through representation transformation.
Data Source
AI summary
A language model training method includes: obtaining first data comprising semiconductor equipment data; performing preprocessing on the first data to generate second data; generating a first token from the second data using a mapping table; and generating a semiconductor language model by performing unsupervised learning on a language model using the first token.