Tool Clustering for Semiconductor Prediction Model Complexity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In semiconductor processing, the random dispatching of wafers to multiple tools leads to complex and difficult-to-maintain models, making precise prediction challenging due to the large number of tool combinations and limited dataset for each processing route, resulting in poor prediction results and impractical virtual wafer acceptance testing.
Innovation Solution
Clustering processing tools at each stage using algorithms like k-mean, ANOVA, and top-down/bottom-up tree methodologies based on parameters such as sensor values and metrology results, reducing the number of prediction models needed and enabling adaptive virtual metrology by grouping tools into clusters, which allows for more efficient dispatching and improved prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If random dispatching is used to process wafers through multiple tools, then flexibility in tool utilization is improved, but the number of required prediction models increases exponentially, making the system difficult to maintain and adapt
Solution Approach 1:
The patent merges multiple individual tool models into a single cluster model by grouping tools with similar processing characteristics. Instead of maintaining separate models for each tool (which would require 72 models for 4 stages with 3 tools each), the system clusters tools into groups and creates one prediction model per cluster, reducing the total number of models while maintaining prediction accuracy across all tools in the cluster.
Solution Approach 2:
The cluster model serves multiple functions: it can predict outcomes for any tool within the cluster, adapt to new tools by assigning them to appropriate clusters, and provide virtual metrology across multiple processing routes. This universal model replaces the need for multiple specialized individual models, enabling the system to handle diverse tool combinations with a single adaptable model.
2Measurement precision
If separate models are created for each tool combination, then prediction accuracy for specific routes is improved, but the difficulty of maintaining and adapting models increases significantly
Solution Approach 1:
The system implements dynamic cluster assignment where tools can be moved between clusters based on their current processing state or performance characteristics. This dynamic approach allows the model to adapt to changing conditions without requiring complete model recreation, maintaining prediction accuracy while reducing maintenance complexity compared to static individual models.
Solution Approach 2:
The patent changes the parameter space from individual tool identifiers to cluster identifiers, fundamentally altering how models are structured and maintained. By transforming the problem from tracking 72 individual tool combinations to managing a few clusters, the system maintains prediction precision while dramatically simplifying model maintenance and adaptation to new tools or processing conditions.
3Productivity
If a new tool is added to the system, then processing capacity is improved, but a completely new model must be created, increasing maintenance burden
Solution Approach 1:
The system performs preliminary clustering of tools based on their processing characteristics before actual wafer processing begins. This preliminary organization creates a framework where new tools can be quickly integrated by assigning them to existing clusters based on their parameters, rather than requiring complete model recreation. The clustering structure is established in advance to facilitate easy adaptation.
Solution Approach 2:
When a new tool is added, the system copies the modeling approach from existing clusters rather than creating a completely new model. By identifying which existing cluster the new tool belongs to based on its processing characteristics, the system can reuse the established model structure and parameters, significantly reducing the effort required to incorporate new processing capacity into the system.
4Reliability
If virtual wafer acceptance testing is implemented with multiple tool combinations, then quality assurance is improved, but the complexity of realizing VWAT becomes impossible
Solution Approach 1:
The patent segments the complex multi-tool VWAT problem into manageable cluster-level assessments. Instead of attempting to simulate and validate all possible tool combinations (which would be impossible), the system performs VWAT at the cluster level, where each cluster model is validated independently. This segmentation makes quality assurance feasible while maintaining confidence in predictions across all tools within validated clusters.
Data Source
AI summary
A first embodiment is a method for semiconductor process control comprising clustering processing tools of a processing stage into a tool cluster based on processing data and forming a prediction model for processing a semiconductor wafer based on the tool cluster. A second embodiment is a method for semiconductor process control comprising providing cluster routes between first stage tool clusters and second stage tool clusters, assigning a comparative optimization ranking to each cluster route, and scheduling processing of wafers. The comparative optimization ranking identifies comparatively which cluster routes provide for high wafer processing uniformity. Further, wafers that require high wafer processing uniformity are scheduled to be processed along one cluster route that has a high comparative optimization ranking that identifies the one cluster route to have a highest wafer processing uniformity, and wafers that do not require high wafer processing uniformity are scheduled to be processed along another cluster route.