Scheduling System for Semiconductor Fabrication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional production scheduling optimization in semiconductor fabrication fails to effectively account for processing-time variations and real-time data streaming, leading to inefficiencies and hierarchy delays in dynamic scheduling.
Innovation Solution
A near-optimal scheduling system incorporating a data parser, publish subscribe mechanism, processing time prediction module, and scheduling optimization module, utilizing orthogonal greedy algorithms and recurrent neural networks to predict operation processing times and generate optimized schedules through ordinal-optimization methods, ensuring real-time dynamic scheduling without hierarchy delays.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional linear programming is used for production scheduling optimization, then the scheduling problem can be solved, but the scheduling performance fails to meet expectations due to inability to grasp on-site conditions and time-consuming calculations
Solution Approach 1:
The system performs preliminary actions by predicting processing times using machine learning models (Random Forest, Gradient Boosting, Neural Networks) before scheduling optimization. Historical data is pre-processed and features are pre-extracted, allowing the optimization to proceed faster with more accurate time estimates, thus reducing overall calculation time while improving scheduling performance
Solution Approach 2:
The patent replaces conventional linear programming mechanisms with a hybrid approach that incorporates machine learning prediction models. This substitution allows the system to handle complex, non-linear relationships in processing time data more effectively, improving scheduling optimization performance while reducing computation time through more efficient algorithms
2Productivity
If conventional scheduling optimization is used, then production schedules can be generated, but hierarchy delays occur and real-time dynamic scheduling is not achieved
Solution Approach 1:
The system segments the scheduling process into independent modules: data collection, feature extraction, processing time prediction, and scheduling optimization. Each module operates independently and can process data in real-time, eliminating hierarchy delays. The publish-subscribe mechanism further segments communication, allowing parallel processing without waiting for hierarchical approvals
Solution Approach 2:
The patent implements dynamic scheduling by continuously updating processing time predictions based on real-time data from the manufacturing execution system. The system adapts to changing conditions on the production line and re-optimizes schedules dynamically, achieving real-time responsiveness without hierarchical delays
3Measurement precision
If processing time variations are not considered, then scheduling can be simplified, but scheduling accuracy deteriorates under uncertain processing conditions
Solution Approach 1:
The system performs preliminary feature extraction and selection from historical data, identifying the most relevant factors that influence processing time variations. This pre-processing reduces the complexity of the scheduling system by focusing only on significant features, while maintaining high prediction accuracy through the machine learning models
Solution Approach 2:
The patent changes parameters by using multiple machine learning models (Random Forest, Gradient Boosting, Neural Networks) that can handle variations in processing time. The system adjusts prediction parameters based on different production conditions and tool types, improving accuracy without requiring overly complex scheduling logic
Data Source
AI summary
A near-optimal scheduling system for considering processing-time variations and real-time data streaming includes a data parser module, a publish subscribe mechanism module, a processing time prediction module and a scheduling optimization module. In the processing time prediction module, an orthogonal greedy algorithm and a recurrent neural network are used to extract key features and then predict varying operation processing time. By formulating the scheduling problem in an integer programming form, an ordinal-optimization (OO) embedded decomposition and coordination method is established in the scheduling optimization module to provide dynamic and near-optimal schedules in a computationally efficient manner. Moreover, a publish subscribe mechanism is developed in the publish subscribe mechanism module by using subscribe and publish methods to realize real-time needs. Through this mechanism, the messages (updated production states) are published as specific topics and subscribed by the required modules of a manufacturing execution system without hierarchy delays.