Semiconductor Die Wafer Mapping Using Partial Measurement Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing complexity and number of processes in semiconductor manufacturing lead to defects and the need for efficient monitoring of semiconductor dies, which is costly and time-consuming, especially with the rise in integration and nano-scale technology.
Innovation Solution
An electronic device employing a machine learning module and arithmetic module to generate predictions about unmeasured semiconductor dies using measured data, process data, and geometric features, reducing the need for exhaustive physical measurements through boosting models and interpolation/extrapolation algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individual measurement and wafer map generation is performed for all semiconductor dies, then measurement precision and quality control are improved, but manufacturing cost and time consumption increase significantly
Solution Approach 1:
The patent applies partial action by performing individual measurements only on a subset of semiconductor dies (first dies) rather than all dies. The system selects representative samples for detailed measurement while using machine learning to infer characteristics of remaining dies, thereby reducing measurement time and cost while maintaining adequate quality control through statistical sampling and predictive analytics
Solution Approach 2:
The patent uses copying by creating virtual representations (wafer maps) of semiconductor dies that were not physically measured. Machine learning models generate predicted wafer maps based on measurements from sampled dies, effectively copying the measurement process results to unmeasured dies through computational inference rather than physical measurement
2Measurement precision
If individual measurement and wafer map generation is performed for all semiconductor dies, then measurement precision is improved, but manufacturing cost increases
Solution Approach 1:
The system performs measurements on only a partial set of dies necessary to achieve statistically significant quality control. By measuring a representative subset rather than all dies, the patent reduces material consumption, equipment usage, and associated costs while maintaining measurement precision through proper sampling methodology
Solution Approach 2:
The patent reduces manufacturing cost by replacing expensive physical measurements with computational copying. Virtual wafer maps are generated through machine learning inference, eliminating the need for costly physical measurement equipment and operations on every die while preserving the essential information needed for quality control
3Adaptability or versatility
If the number of processes and integration degree increase, then manufacturing capability and functionality are improved, but device complexity and computational requirements increase
Solution Approach 1:
The patent applies universality by designing a machine learning-based quality control system that can handle multiple process types and integration levels through a single unified framework. The system processes diverse semiconductor manufacturing data (process parameters, measurement results, wafer maps) using generalizable algorithms that adapt to different device complexities without requiring separate specialized systems for each process type
Solution Approach 2:
The patent introduces machine learning models as intermediary components between physical manufacturing processes and quality control decisions. These intermediary models simplify the complexity by learning patterns from process data and providing predictive insights, thereby managing device complexity through computational abstraction rather than direct handling of raw process variables
4Loss of time
If machine learning prediction is used for unmeasured dies, then time consumption and cost are reduced, but measurement precision may be compromised
Solution Approach 1:
The patent applies partial action by performing exhaustive physical measurements on a carefully selected subset of first dies that provide maximum information content. This partial measurement approach, combined with machine learning inference on remaining dies, achieves an optimal balance where the measured subset provides sufficient precision while the predictive models extend coverage to unmeasured dies with acceptable accuracy
Solution Approach 2:
The patent implements feedback mechanisms where machine learning models are trained and refined using actual measurement data from sampled dies. The system continuously improves prediction precision by learning from the discrepancy between predicted and actual measurements, thereby reducing the precision gap between measured and unmeasured dies over time through iterative model optimization
Data Source
AI summary
A method supporting manufacture of semiconductor dies, the method includes obtaining process data of the semiconductor dies, wherein the semiconductor dies include first dies, second dies, and third dies, obtaining measurement data associated with features of the first dies, encoding the process data to obtain preprocessed process data, generating first prediction data representing features of the second dies based on the measurement data and the preprocessed process data, computing second prediction data representing features of the third dies based on the measurement data and the first prediction data, and generating full-die level information representing features of the semiconductor dies based on the measurement data, the first prediction data, and the second prediction data.