RNN Process Parameter Prediction in High-Mix Semiconductor Manufacturing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor manufacturing processes face challenges in adopting AI/ML technologies due to the complexity of building and deploying deep learning models, especially in high-mix environments, limiting their effectiveness in predicting processing parameters.
Innovation Solution
A recurrent neural network (RNN) deep learning model is trained and deployed to predict and set processing parameters in semiconductor manufacturing, utilizing context information, feedback loops, and subject matter expertise for feature selection and hyperparameter tuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a linear process model is used to calculate processing parameters, then the calculation process is simple, but the prediction accuracy is limited
Solution Approach 1:
The patent transitions from a linear process model with fixed parameters to a deep learning model where parameters are dynamically adjusted based on learned patterns from historical data. The model learns optimal processing parameters by training on datasets containing process conditions and outcomes, enabling adaptive parameter selection that improves prediction accuracy while maintaining computational efficiency through pre-trained weights.
2Measurement precision
If AI/ML technologies are adopted in semiconductor manufacturing, then prediction accuracy improves, but the complexity of building and deploying models increases
Solution Approach 1:
The patent segments the complex model development process into distinct phases: data collection and preparation, model training with historical data, validation and testing, and deployment. This segmentation allows each phase to be managed independently, reducing overall complexity. The use of transfer learning further simplifies deployment by enabling models trained on one process to be adapted to similar processes without complete retraining.
Solution Approach 2:
The patent performs preliminary actions by pre-training models on extensive historical manufacturing data before deployment. This pre-training establishes a solid foundation of learned patterns that can be quickly adapted to specific production scenarios. The model architecture and feature engineering are also performed in advance, so that when deployment occurs, the complex work of building the model is already complete, reducing on-site complexity.
3Ease of operation
If traditional linear models are used for high-mix manufacturing, then the system is easier to operate, but the adaptability to different process conditions is limited
Solution Approach 1:
The patent implements dynamic adaptability by training separate deep learning models for different process conditions, product types, and manufacturing scenarios. The system dynamically selects the appropriate pre-trained model based on the current production context, enabling adaptability across high-mix environments. This dynamic model selection maintains ease of operation because the system automatically chooses the right model without requiring operator intervention, while simultaneously providing versatility across diverse process conditions.
Data Source
AI summary
Disclosed techniques for applying a neural network deep learning model in a fabrication strategy for high-mix semiconductor manufacturing, such as deposition, chemical-mechanical polishing (CMP), etching, photolithography, plating, etc. Training and normal operation modes of the fabrication strategy are described.