Semiconductor ML Model Training Using Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current algorithms used in semiconductor applications, such as inspection and metrology, require hand-crafted features and tend to rely on ad hoc and lossy methods, limiting their ability to fully utilize the capabilities of metrology and inspection equipment.
Innovation Solution
A system and method for training a machine learning-based model that accelerates learning by using information from nominal instances of specimens, with the option to re-train using non-nominal instances through transfer learning, enabling efficient training even with limited samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current algorithms are used in semiconductor applications, then hand-crafted features can be extracted, but the training process is slow and requires extensive manual feature extraction
Solution Approach 1:
The patent replaces manual hand-crafted feature extraction with automated machine learning-based feature extraction. The system uses trained models to automatically identify and extract relevant features from semiconductor inspection images, eliminating the need for manual feature engineering and significantly reducing training time while maintaining or improving detection accuracy.
Solution Approach 2:
The patent implements pre-training of machine learning models using synthetic defect data before deployment on real semiconductor inspection data. This preliminary training phase allows the model to learn general defect patterns from synthesized examples, reducing the time required for fine-tuning on actual production data and accelerating the overall training process.
2Measurement precision
If extensive data sets are used for training, then model accuracy improves, but data acquisition and processing time increases
Solution Approach 1:
The patent uses synthetic defect data generated by copying and transforming real defect patterns. The system creates artificial defect images by applying various transformations (scaling, rotation, noise addition) to real defect examples, generating large training datasets without requiring additional physical defect samples. This copying approach maintains model accuracy while eliminating the time-consuming process of acquiring extensive real defect data.
Solution Approach 2:
The patent employs data augmentation techniques that change parameters of existing defect images, such as adjusting brightness, contrast, rotation angles, and scale factors. By systematically varying these parameters, the system generates diverse training samples from limited original data, improving model accuracy without requiring proportional increases in data acquisition time.
3Productivity
If hand-crafted features are used, then feature extraction can be performed, but false alarm rates increase and efficiency decreases
Solution Approach 1:
The patent replaces manual hand-crafted feature extraction algorithms with automated machine learning-based feature extraction. The trained models learn optimal feature representations directly from data, automatically identifying relevant patterns while filtering out noise and irrelevant information. This substitution significantly reduces false alarm rates by distinguishing true defects from artifacts, while simultaneously improving inspection efficiency through automated processing.
4Measurement precision
If rigorous process simulation DOEs are employed, then process window analysis can be performed, but training speed decreases and additional data requirements increase
Solution Approach 1:
The patent performs preliminary training of machine learning models using synthetic defect data that encompasses various process conditions and defect types. This pre-training phase prepares the model to handle diverse scenarios, reducing the need for extensive post-deployment process simulation and DOE experiments. The model learns robust feature representations that generalize across different process windows, maintaining simulation accuracy while accelerating training speed.
Data Source
Figure 1
Figure 1a~2
Figure 3
AI summary
Methods and systems for accelerated training of a machine learning based model for semiconductor applications are provided. One method for training a machine learning based model includes acquiring information for non-nominal instances of specimen(s) on which a process is performed. The machine learning based model is configured for performing simulation(s) for the specimens. The machine learning based model is trained with only information for nominal instances of additional specimen(s). The method also includes re-training the machine learning based model with the information for the non-nominal instances of the specimen(s) thereby performing transfer learning of the information for the non-nominal instances of the specimen(s) to the machine learning based model.