Lithography Predictive Maintenance Model for Sparse Configuration Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Lithographic systems face challenges in predictive maintenance due to limited data availability and system heterogeneity, leading to low prediction accuracy and inefficiencies in maintenance strategies.
Innovation Solution
A computational framework and method for predictive maintenance that involves a prediction model comprising a first component optimized for a population of tools with different configurations and a second component based on knowledge of configuration relationships, using a cost function that incorporates a configuration-specific component and a regularization term to account for similarities across datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a prediction model is optimized for a specific configuration using only configuration-specific data, then prediction accuracy for that configuration is improved, but the model cannot generalize to other configurations and data availability is limited
Solution Approach 1:
The patent combines configuration-specific data with data from other configurations to train the prediction model. The loss function integrates both configuration-specific loss and data from other configurations, allowing the model to learn general patterns while maintaining configuration-specific accuracy. This merging of data sources resolves the contradiction by enabling the model to generalize across configurations while still achieving high prediction accuracy for specific configurations.
Solution Approach 2:
The prediction model is designed to be universal across multiple configurations by training it on diverse configuration data. The model structure and loss function are configured to handle multiple configurations simultaneously, making the model adaptable to different system configurations while maintaining prediction accuracy. This universality approach allows a single model to serve multiple configurations without requiring separate models for each.
2Adaptability or versatility
If data from multiple configurations is used to train a single prediction model, then model generalization is improved, but configuration-specific prediction accuracy may deteriorate due to data heterogeneity
Solution Approach 1:
The loss function incorporates a configuration-specific component that ensures each configuration receives appropriate attention during training. By weighting the loss contribution from different configurations, the model maintains high prediction accuracy for specific configurations while still learning from diverse data. This local quality approach prevents the model from being overly generalized at the expense of configuration-specific performance.
Solution Approach 2:
The patent uses parameter optimization to adjust model weights and hyperparameters based on configuration-specific characteristics. By dynamically adjusting parameters during training based on the specific configuration being evaluated, the model can adapt its behavior to maintain high accuracy for each configuration while benefiting from the generalization provided by multi-configuration training data.
3Measurement precision
If separate prediction models are trained for each configuration, then configuration-specific accuracy is maximized, but device complexity and maintenance overhead increase
Solution Approach 1:
The patent employs a single universal prediction model that can handle multiple configurations through configuration identification and adaptive parameter adjustment. This eliminates the need to develop, deploy, and maintain separate models for each configuration, significantly reducing device complexity and maintenance overhead. The model achieves configuration-specific accuracy through its ability to adapt to different configurations rather than through separate model instances.
Solution Approach 2:
The prediction model incorporates dynamic parameter adjustment capabilities that allow it to adapt its behavior based on the detected configuration. Rather than being a static model trained for a single configuration, the model dynamically adjusts its parameters and processing approach based on the input configuration, enabling a single model to replace multiple static models while maintaining configuration-specific accuracy.
4Productivity
If predictive maintenance is implemented with limited data availability, then maintenance strategies can be deployed quickly, but prediction accuracy remains low
Solution Approach 1:
The patent merges data from multiple configurations and sources to compensate for limited availability of configuration-specific data. By combining diverse data sources in the training process, the model achieves sufficient prediction accuracy even when individual configuration datasets are small. This approach enables rapid deployment of predictive maintenance while maintaining acceptable prediction accuracy through data augmentation from related configurations.
Data Source
AI summary
A method of tuning a prediction model relating to at least one particular configuration of a manufacturing device. The method includes obtaining a function including at least a first function of first prediction model parameters associated with the at least one particular configuration, and a second function of the first prediction model parameters and second prediction model parameters associated with configurations of the manufacturing device and/or related devices other than the at least one particular configuration. Values of the first prediction model parameters are obtained based on an optimization of the function, and a prediction model is tuned according to these values of the first prediction model parameters to obtain a tuned prediction mode.