Predictive Model Selection for Semiconductor Wafer Failure Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The semiconductor manufacturing industry faces challenges in detecting and classifying wafer failures due to the high volume and complexity of sensor data, with high correlation between process variables and non-linearity, making it difficult to build robust machine learning models that generalize across disparate equipment clusters.
Innovation Solution
A method is developed to collect and process input data from sensors and other manufacturing sources using machine learning algorithms, involving data cleansing, feature engineering, and model selection to generate predictive models for failure detection and classification, which includes anomaly detection, hierarchical feature selection, and multi-class classification to identify root causes of excursions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are built using sensor data from disparate equipment clusters, then failure detection capability is improved, but model robustness deteriorates due to high correlation between process variables and non-linearity
Solution Approach 1:
The patent segments the complex sensor data into multiple feature categories (temporal features, spectral features, statistical features) and processes them through separate analytical pathways before integration. This segmentation allows the model to handle different aspects of the data independently, reducing the impact of high correlation between variables and improving overall model robustness across disparate equipment clusters.
Solution Approach 2:
The patent transforms the sensor data from its original high-dimensional space into a reduced feature space by extracting key characteristics across multiple dimensions (time, frequency, statistics). This dimensional transformation reduces non-linearity by projecting complex relationships into a more manageable feature space that maintains discriminative power while improving model generalization across different equipment clusters.
2Measurement precision
If comprehensive sensor data is collected from multiple sources, then detection accuracy is improved, but data processing complexity increases
Solution Approach 1:
The patent extracts only the most relevant features from the comprehensive sensor data through systematic feature engineering. By selecting and extracting key temporal, spectral, and statistical features rather than processing all raw sensor data, the system maintains high detection accuracy while significantly reducing data processing complexity and computational burden.
Solution Approach 2:
The patent performs preliminary data processing and feature extraction before feeding data into the machine learning model. This preliminary action includes data normalization, feature selection, and transformation, which simplifies the subsequent modeling process and reduces the complexity of handling comprehensive sensor data from multiple sources.
3Adaptability or versatility
If machine learning pipelines are generalized towards disparate equipment clusters, then applicability is improved, but performance deteriorates due to process variable correlations
Solution Approach 1:
The patent develops a universal feature extraction framework that can be applied across disparate equipment clusters (etch, CMP, lithography, deposition, wets). The same feature engineering and modeling pipeline is used for all equipment types, enabling generalization while maintaining performance through the robust handling of process variable correlations inherent in semiconductor manufacturing processes.
Data Source
AI summary
A robust predictive model. A plurality of different predictive models for a target feature are run, and a comparative analysis provided for each predictive model that meet minimum performance criteria for the target feature. One of the predictive models is selected, either manually or automatically, based on predefined criteria. For semi-automatic selection, a static or dynamic survey is generated for obtaining user preferences for parameters associated with the target feature. The survey results will be used to generate a model that illustrates parameter trade-offs, which will be used to finalize the optimal predictive model for the user.