AI Quality Assurance for Semiconductor Defect Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional process control techniques, such as Statistical Process Control, are inadequate for predicting defects in complex manufacturing processes due to their limitations in handling multiple machine parameters and reliance on linear models, leading to fragmented and unbalanced data, which restricts the adoption of advanced AI solutions like neural networks that require vast amounts of balanced data.
Innovation Solution
An AI-based prediction method using a neural network trained with multi-task learning and transfer learning to model machine operations, integrating non-homogeneous machine parameters, maintenance, and quality data, enabling real-time quality assurance predictions in semiconductor manufacturing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional Statistical Process Control techniques are used, then the system is simple to implement, but it cannot reliably anticipate defects in complex manufacturing processes
Solution Approach 1:
The patent replaces traditional mechanical Statistical Process Control methods with an AI-based neural network system. The neural network models complex relationships between multiple machine parameters and defect outcomes, substituting simple statistical thresholds with sophisticated pattern recognition that can handle non-linear, multi-parameter interactions in semiconductor manufacturing.
Solution Approach 2:
The system transforms the control approach by changing from fixed threshold-based parameters to dynamic, learned parameters through multi-task learning. The neural network learns optimal parameter relationships and thresholds automatically from historical data, adapting to process variations without manual reconfiguration, thereby improving reliability while managing complexity through automated parameter optimization.
2Reliability
If advanced AI techniques like neural networks are implemented, then defect prediction accuracy improves, but vast amounts of balanced data are required
Solution Approach 1:
The patent implements a multi-task learning neural network that performs multiple prediction functions simultaneously - predicting different types of defects, multiple machine parameter outcomes, and various quality metrics from the same input data. This multi-functionality allows the system to learn from diverse data sources and tasks, improving prediction accuracy while reducing the data requirement for each individual task through shared learning.
Solution Approach 2:
The system performs preliminary data processing and feature extraction before the main prediction task. By pre-processing data to handle imbalances and extract relevant features in advance, the neural network can achieve high accuracy with less raw data. The preliminary action includes data normalization, feature selection, and handling of missing values that prepares the data for efficient learning.
3Ease of operation
If manual set-up of thresholds is used, then the system is easy to configure, but it complicates practical use due to fragmented and unbalanced data
Solution Approach 1:
The neural network system performs self-service by automatically learning optimal thresholds and parameter relationships from historical data without manual configuration. The multi-task learning framework enables the system to self-adapt to different machines and processes, automatically handling data fragmentation and imbalances through learned representations rather than manual threshold setting, thereby improving reliability while maintaining ease of operation through automated configuration.
Data Source
AI summary
The subject matter herein provides for AI-based prediction of production defects in association with a production system, such as a semiconductor manufacturing machine. In one embodiment, a method begins by receiving production data from the production system. The production data typically comprises non-homogeneous machine parameters and maintenance data, quality test data, and product and process data. Using the production data, a neural network is trained to model an operation of a given machine in the production system. Preferably, the training involves multi-task learning, transfer learning (e.g., using knowledge obtained with respect to a machine of the same type as the given machine), and a combination of multi-task learning and transfer learning. Once the model is trained, it is associated with the given machine operating environment, wherein it is used to provide quality assurance predictions.