Virtual Metrology Modeling with Autoencoders and Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning-based automatic virtual metrology (AVM) algorithms require large amounts of paired data for modeling and lack intelligent self-learning capabilities for online model refreshing, limiting their application in high-tech factories like semiconductor and TFT-LCD production.
Innovation Solution
A virtual metrology method integrating convolutional autoencoders with transfer learning to perform modeling and prediction, enabling effective initial-model-creation with insufficient data and incorporating a dual-phase scheme for intelligent self-learning to maintain prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep learning-based AVM algorithms are used, then prediction accuracy can be improved, but large amounts of paired data are required for modeling
Solution Approach 1:
The patent applies preliminary action by using unsupervised pre-training with autoencoders on unpaired process data before supervised fine-tuning. The autoencoder learns feature representations from available unpaired data in advance, so when paired data becomes available, the model can be quickly adapted with less paired data required, thus resolving the contradiction between high accuracy and large data requirements
Solution Approach 2:
The patent introduces transfer learning as an intermediary mechanism between unsupervised learning on unpaired data and supervised learning on paired data. The pre-trained model serves as an intermediary that transfers learned features from the unlabeled data to the final prediction task, reducing the amount of paired data needed while maintaining prediction accuracy
2Adaptability or versatility
If conventional AVM models are deployed, then initial modeling can be performed, but they lack intelligent self-learning capability for online model refreshing
Solution Approach 1:
The patent implements dynamics by enabling the model to transition from a static pre-trained state to a dynamic online learning state. The model can continuously adapt to new data distributions through online fine-tuning when paired data becomes available, making the system dynamic and adaptable rather than fixed, thus providing intelligent self-learning capability while managing complexity through automated updating mechanisms
3Productivity
If sampling inspection is used, then measurement resources are conserved, but metrology delay occurs and real-time total inspection cannot be achieved
Solution Approach 1:
The patent creates a virtual copy of the metrology measurement through the AVM model. Instead of physically measuring every workpiece (which would consume resources), the model generates virtual measurement predictions that replicate the information obtained from actual metrology, enabling real-time total inspection without the time delay of physical measurements while conserving measurement resources
Data Source
AI summary
A virtual metrology method based on convolutional autoencoder and transfer learning includes performing a data alignment operation, a modeling operation and a calculating operation. The data alignment operation includes performing a data-length adjusting operation onto a plurality of sets of process data. The modeling operation includes classifying paired data and unpaired process data; creating a pre-trained model by using the unpaired process data, and then inputting the paired data to the pre-trained model to create a virtual metrology model based on convolutional autoencoder. The virtual metrology model based on convolutional autoencoder includes at least one convolutional neural network model. In addition, the calculating operation includes executing one of a predicting step and a transfer learning step according to whether the actual metrology data is obtained, thereby calculating one of a phase-one virtual metrology value and a phase-two virtual metrology value.