Substrate Processing Scheduler Using Neural Networks for Wafer Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current substrate processing systems face challenges in optimizing scheduling parameters for semiconductor wafer processing, leading to inefficiencies such as idle times, reduced throughput, and difficulty in managing complex scheduling constraints, especially in batch processing tools with varying recipe times and wafer wait times.
Innovation Solution
A machine-learning assisted scheduler using a neural network model is trained with data from preventive maintenance operations, recipe times, and wafer-less auto clean times to predict optimum scheduling parameters, minimizing idle times and maximizing throughput by simulating and optimizing wafer routing and processing scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional scheduling methods are used for substrate processing, then system complexity is manageable, but productivity and throughput are reduced due to idle times and inefficiencies
Solution Approach 1:
The patent replaces traditional mechanical scheduling approaches with an artificial intelligence-based system comprising neural networks and machine learning algorithms. The AI system processes scheduling parameters, predicts optimal sequences, and dynamically adjusts processing schedules, substituting conventional control mechanisms with intelligent computational systems that learn from historical data and improve scheduling efficiency automatically
Solution Approach 2:
The system dynamically changes scheduling parameters such as processing sequences, chamber assignments, and timing based on real-time conditions and historical performance data. The AI model adjusts multiple parameters simultaneously to optimize throughput while managing system complexity through intelligent parameter transformation rather than manual configuration
2Productivity
If scheduling parameters are optimized for maximum throughput, then productivity increases, but wafer idle times may increase due to complex coordination requirements
Solution Approach 1:
The system implements continuous feedback loops where the AI model monitors actual processing times, wafer movement durations, and chamber availability in real-time. This feedback is used to dynamically adjust scheduling predictions and reduce idle times by anticipating bottlenecks before they occur, allowing the system to maintain high throughput while minimizing wait periods through adaptive real-time optimization
Solution Approach 2:
The AI system performs preliminary scheduling actions by predicting optimal processing sequences in advance based on historical data and current system state. It pre-coordinates chamber availability, substrate transfer timing, and processing start times to minimize idle periods, ensuring that wafers are ready for processing immediately when chambers become available, thereby reducing wait times while maintaining maximum throughput
3Adaptability or versatility
If a simple scheduling system is used, then ease of operation is maintained, but adaptability to tool-to-tool variations and performance drift is poor
Solution Approach 1:
The AI-based scheduling system performs self-service by automatically adapting to tool-to-tool variations and performance drift without requiring manual reconfiguration. The neural networks learn from historical data specific to each tool's characteristics and continuously update their models to account for performance changes, enabling the system to self-adjust and maintain optimal scheduling for diverse tools while preserving operational simplicity for users
Data Source
AI summary
For etching tools, a neural network model is trained to predict optimum scheduling parameter values. The model is trained using data collected from preventive maintenance operations, recipe times, and wafer-less auto clean times as inputs. The model is used to capture underlying relationships between scheduling parameter values and various wafer processing scenarios to make predictions. Additionally, in tools used for multiple parallel material deposition processes, a nested neural network based model is trained using machine learning. The model is initially designed and trained offline using simulated data and then trained online using real tool data for predicting wafer routing path and scheduling. The model improves accuracy of scheduler pacing and achieves highest tool/fleet utilization, shortest wait times, and fastest throughput.