Semiconductor Metrology Using End-to-End Image Measurement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor examination processes face challenges in achieving high precision and uniformity of device features due to the complexity of multi-stage metrology systems, where tuning and optimization rely heavily on user judgment, leading to suboptimal performance.
Innovation Solution
Implementing an end-to-end (E2E) learning model to directly process specimen images, replacing multiple processing modules, and optimizing the system based on metrology benchmarks, thereby simplifying the architecture and reducing user dependency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple processing modules are used in the metrology system, then the system can perform comprehensive examination functions, but the system complexity increases and user dependency rises
Solution Approach 1:
The patent combines multiple separate processing modules (image preprocessing, image segmentation, and measurement modules) into a single integrated end-to-end learning model. This consolidation maintains the comprehensive examination capabilities while reducing system complexity and eliminating the need for separate tuning of each module, directly resolving the contradiction between versatility and complexity.
2Adaptability or versatility
If traditional multi-module metrology systems are used, then various examination tasks can be performed, but tuning and optimization rely heavily on user judgment leading to suboptimal performance
Solution Approach 1:
The patent implements feedback mechanisms where the end-to-end learning model is trained using ground truth measurement data and cost functions that evaluate precision, correlation, and matching. This automated feedback loop replaces subjective user judgment with objective performance metrics, enabling optimal tuning while maintaining comprehensive examination capabilities.
Solution Approach 2:
The patent transforms the system from requiring manual parameter tuning by users to automated parameter optimization through machine learning training. The model learns optimal parameters during training using cost functions that measure precision, correlation with ground truth, and matching across different tools, thereby improving measurement performance while maintaining versatility.
3Adaptability or versatility
If multiple processing modules are implemented, then the system can handle diverse metrology tasks, but the tuning process becomes more difficult and time-consuming
Solution Approach 1:
The patent performs preliminary action by pre-training the end-to-end learning model on comprehensive datasets that cover diverse metrology tasks. This preliminary training embeds the knowledge for handling various examination tasks within the model itself, eliminating the need for time-consuming manual tuning when deploying the system for different metrology applications.
Data Source
AI summary
There is provided a system and method for examining a semiconductor specimen. The method includes obtaining a runtime image of the specimen, and providing the runtime image as an input to an end-to-end (E2E) learning model to process, thereby obtaining, as an output of the E2E learning model, runtime measurement data specific for a metrology application. The E2E learning model is previously trained for the metrology application using a training set comprising a plurality of training images of the specimen and respective ground truth measurement data associated therewith, and one or more cost functions specifically configured to evaluate, for the plurality of training images and corresponding training measurement data outputted by the E2E learning model, one or more metrology benchmarks from a group comprising precision, correlation, and matching.