Multi-Stage Gaussian Process Modelling for Semiconductor Metrology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semiconductor device manufacturing process control methods focus on single stages, which are insufficient for complex multi-stage processes, leading to inaccurate predictions and delayed detection of outcomes outside process specifications, resulting in re-work or scrapping of products.
Innovation Solution
The implementation of predictive multi-stage modelling using a multiplicative kernel Gaussian process to collect geometrical data from select stages and make outcome probability predictions, enabling real-time probabilistic predictions for wafer disposition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If single stage disposition is used for process control, then the disposition decision is simple and quick, but the prediction accuracy is insufficient and outcomes are detected too late for meaningful intervention
Solution Approach 1:
The patent segments the complex multi-stage process into individual stage models, where each stage has its own prediction model trained on locally available data. This allows accurate predictions at each stage without requiring a single complex model for the entire process, resolving the contradiction between prediction accuracy and model complexity.
Solution Approach 2:
The patent introduces a temporal dimension by making predictions at multiple stages along the manufacturing process rather than just at the final stage. This multi-stage prediction approach enables early detection of potential failures while maintaining manageable model complexity at each individual stage.
2Reliability
If comprehensive multi-stage modelling is implemented, then prediction accuracy improves, but computational tractability decreases due to too many variables
Solution Approach 1:
The patent divides the comprehensive multi-stage modeling problem into smaller, manageable stage-specific models. Each model processes only the variables relevant to its stage, maintaining computational efficiency while collectively providing reliable predictions across the entire manufacturing process.
Solution Approach 2:
The patent implements partial modeling by creating models for selected critical stages rather than attempting to model every single stage in detail. This selective approach maintains prediction reliability for key outcomes while preserving computational tractability by focusing resources on the most important process stages.
3Measurement precision
If metrology measurements are taken at all intermediate stages, then complete data is available for accurate prediction, but measurement cost and process time increase
Solution Approach 1:
The patent extracts and utilizes only the necessary metrology measurements from intermediate stages that are most critical for prediction accuracy. By selecting key measurement points rather than measuring at all stages, the system achieves accurate predictions without adding excessive process time or measurement costs.
Solution Approach 2:
The patent implements partial measurement by collecting data at select stages rather than all intermediate stages. This selective data collection approach provides sufficient information for accurate outcome prediction while minimizing the additional time and cost associated with comprehensive metrology measurements.
Data Source
AI summary
Predictive multi-stage modelling for complex semiconductor device manufacturing process control is provided. In one aspect, a method of predictive multi-stage modelling for controlling a complex semiconductor device manufacturing process includes: collecting geometrical data from metrology measurements made at select stages of the manufacturing process; and making an outcome probability prediction at each of the select stages using a multiplicative kernel Gaussian process, wherein the outcome probability prediction is a function of a current stage and all prior stages. Machine-learning models can be trained for each of the select stages of the manufacturing process using the multiplicative kernel Gaussian process. The machine-learning models can be used to provide probabilistic predictions for a final outcome in real-time for production wafers. The probabilistic predictions can then be used to select production wafers for rework, sort, scrap or disposition.