Reusable ML Architecture Libraries for Application-Specific Metrology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for constructing machine learning (ML) architectures in semiconductor manufacturing, such as for wafer inspection and metrology, lack flexibility and adaptability to application-specific characteristics, often using fixed training algorithms and loss functions, leading to suboptimal performance.
Innovation Solution
A system and method for constructing a machine learning library that includes defining reusable architecture blocks and templates, assigning metadata based on input data metrics and performance objectives, and storing them for selecting application-specific ML architectures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If fixed training algorithms and loss functions are used in ML architectures, then the system structure is simple and easy to implement, but the adaptability to application-specific characteristics deteriorates
Solution Approach 1:
The patent implements dynamic selection of training algorithms and loss functions based on application-specific requirements. The system allows switching between different training algorithms (e.g., supervised, unsupervised, reinforcement learning) and loss functions (e.g., MSE, cross-entropy, focal loss) depending on the specific inspection or metrology application, thereby resolving the contradiction between implementation simplicity and adaptability.
Solution Approach 2:
The patent enables parameter changes in the ML architecture by allowing configuration of different hyperparameters, learning rates, batch sizes, and architecture parameters (e.g., number of layers, filter sizes) tailored to specific applications. This parameter flexibility allows the same base architecture to adapt to different semiconductor inspection and metrology tasks without sacrificing implementation ease.
2Adaptability or versatility
If application-specific ML architectures are created for each scenario, then the adaptability improves, but the device complexity increases
Solution Approach 1:
The patent creates a universal ML architecture framework that can serve multiple semiconductor inspection and metrology applications. The system uses a common base architecture with configurable components that can be adapted to different applications (defect detection, overlay measurement, CD measurement, etc.) without requiring completely separate systems, thereby reducing overall complexity while maintaining high adaptability.
Solution Approach 2:
The patent segments the ML architecture into modular, reusable components including architecture blocks (e.g., convolutional layers, pooling layers), training algorithms, loss functions, and preprocessing pipelines. This segmentation allows independent configuration and optimization of each component for specific applications without increasing overall system complexity, as modules can be selectively assembled.
3Adaptability or versatility
If reusable architecture blocks and templates are defined with metadata, then the extendibility improves, but the initial setup time and complexity increase
Solution Approach 1:
The patent performs preliminary action by pre-defining reusable architecture blocks, templates, and metadata structures that can be directly applied to new applications. Common architectural patterns and configurations are established in advance, allowing rapid deployment of new ML models for different semiconductor inspection and metrology tasks without requiring extensive setup time or complex configuration from scratch.
Data Source
AI summary
Methods and systems for constructing a machine learning (ML) library are provided. One method includes defining multiple architecture blocks, each of which is a reusable piece of ML architecture, and defining multiple architecture templates, each of which is a reusable template configurable for including one or more of the multiple architecture blocks. The method also includes assigning metadata to the templates responsive to input data metrics and performance objectives for which the templates are suited. The method further includes storing the blocks, templates, and metadata in a ML library configured for use in selecting one or more of the templates for an application-specific ML architecture based on the input data metrics and the performance objectives specific to the application. Similar steps may be performed for loss functions and hyperparameters. The embodiments provide flexibility and extendibility of ML architectures for application-specific challenging scenarios for applications such as metrology.