Semiconductor Yield Prediction via Machine Learning Feature Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing yield prediction methods in semiconductor manufacturing are inadequate, relying on univariate analysis and assuming normally distributed and independent data, which is not always the case, and fail to effectively utilize the growing amount of test data, leading to inefficiencies in inventory management and process optimization.
Innovation Solution
Implementing machine learning, data mining, and predictive analytics using parallel processing to analyze large datasets, incorporating robust regression algorithms, tree-based algorithms, neural networks, and support vector machines to predict yield and optimize tests, while identifying important features and variables for improved yield prediction and inventory management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If univariate analysis and traditional statistical methods are used for yield prediction, then the methodology is simple and easy to implement, but the prediction accuracy is insufficient and cannot effectively utilize growing test data
Solution Approach 1:
The patent transitions from univariate analysis to multivariate analysis by changing the fundamental parameter of data analysis. Machine learning algorithms process multiple variables simultaneously (test data, process data, defect data) to improve prediction accuracy while maintaining computational feasibility through automated feature selection and model training procedures.
Solution Approach 2:
The patent replaces traditional statistical mechanical methods with machine learning algorithms. Instead of relying on assumptions of normal distribution and independence, the system uses computational models (neural networks, support vector machines, random forests) that can capture complex non-linear relationships in the data without requiring explicit mathematical formulations.
2Measurement precision
If machine learning and multivariate analysis are implemented, then yield prediction accuracy improves, but the computational complexity and data processing requirements increase
Solution Approach 1:
The patent segments the complex data processing task into distinct phases: data collection from multiple sources, data preprocessing and cleaning, feature selection and extraction, model training, and prediction. This segmentation allows each component to be optimized independently and facilitates parallel processing of different data streams and model training tasks.
Solution Approach 2:
The patent performs preliminary actions by collecting and preprocessing data before the actual prediction is needed. Historical test data, process data, and defect data are gathered and prepared in advance, allowing the machine learning models to be trained on comprehensive datasets. Feature selection and engineering are performed beforehand to reduce the dimensionality of the input data, simplifying the subsequent prediction process.
3Measurement precision
If comprehensive test data is collected and analyzed, then yield prediction accuracy improves, but the time and resources required for data processing increase
Solution Approach 1:
The patent implements periodic action by updating yield predictions at key stages in the manufacturing process rather than continuously analyzing all data. Models are trained periodically on accumulated data and applied at decision points such as wafer sort, assembly, and final test, allowing time for data aggregation while maintaining timely predictions for process optimization.
Solution Approach 2:
The patent performs preliminary data processing and feature extraction before the actual prediction is needed. Historical data is preprocessed, cleaned, and transformed into meaningful features in advance, reducing the computational burden during real-time prediction. This allows comprehensive data analysis without proportionally increasing processing time during critical decision points.
Data Source
AI summary
A method for predicting yield for a semiconductor process. A particular type of wafer is fabricated to have a first set of features disposed on the wafer, with a wafer map identifying a location for each of the first set of features on the wafer. Data from wafer acceptance tests and circuit probe tests is collected over time for wafers of that particular type as made in a semiconductor fabrication process, and at least one training dataset and a least one validation dataset are created from the collected data. A second set of “engineered” features are created and also incorporated onto the wafer and wafer map. Important features from the first and second sets of features are identified and selected, and using those important features as inputs, a number of different process models are run, with yield as the target. The results of the different models can be combined, for example, statistically.