E-Beam Yield Prediction Using ML Before Electrical Test
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor manufacturing processes face challenges in providing reliable yield prediction during in-line examination due to the difficulty in correlating in-line measurements with end-of-line electrical test data, which typically have a long feedback delay and are multi-variant, making it difficult to predict yield accurately before electrical testing is completed.
Innovation Solution
A machine learning-based system that utilizes a trained model to analyze electron beam images in real-time during fabrication, correlating them with historical electrical test data to predict yield before electrical testing is performed, incorporating additional data types such as optical and sensor data, and allowing for layer-specific model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If in-line examination measurements are used for yield prediction, then early yield prediction capability is improved, but correlation accuracy with end-of-line electrical test data deteriorates due to long feedback delay and multi-variant factors
Solution Approach 1:
The patent introduces an intermediary mapping model that bridges in-line examination measurements and end-of-line electrical test data. This mapping model learns the complex relationship between intermediate process parameters and final yield outcomes, enabling accurate yield prediction without direct causal linkage. The mapping model acts as a mediator that translates early process indicators into meaningful yield predictions despite the temporal and causal gaps.
Solution Approach 2:
The system performs preliminary yield prediction during in-line examination before electrical testing is completed. By analyzing process measurements and examination data at intermediate stages, the system predicts yield outcomes in advance, allowing early identification of potential defects and performance issues without waiting for final electrical test results.
2Measurement precision
If multiple data types (optical, sensor, e-beam) are integrated for yield prediction, then prediction accuracy is improved, but system complexity increases
Solution Approach 1:
The patent merges multiple data sources including optical inspection data, sensor measurements, and e-beam examination images into a unified yield prediction framework. By combining these diverse data types, the system captures comprehensive process information from different modalities, enabling more accurate yield prediction through multi-variant analysis while managing system complexity through integrated processing.
3Measurement precision
If layer-specific model training is performed, then prediction accuracy for specific structures is improved, but training time and computational resources increase
Solution Approach 1:
The patent segments the yield prediction task by training separate machine learning models for different structure types or layers within the semiconductor device. Each model is specialized for predicting yield of specific structures (e.g., transistors, interconnects, capacitors), enabling structure-specific accuracy. This segmentation allows targeted training on relevant data subsets, reducing overall training time compared to a single comprehensive model while maintaining high precision for each structure type.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enables early and accurate yield prediction during semiconductor fabrication, reducing feedback delays and improving process control by providing timely insights into potential defects and performance, thereby enhancing manufacturing efficiency and quality.
Implementation Method 1
obtain an electron beam (e-beam) image representative of a given layer of a given structure on the semiconductor specimen, the e-beam image acquired in runtime during in-line examination
Data Source
AI summary
There is provided a system and method of examination of a semiconductor specimen. The method includes obtaining an e-beam image representative of a given layer of a given structure on the specimen in runtime, processing at least the e-beam image using a ML model, and obtaining yield related prediction with respect to the given structure prior to performing an electrical test. The ML model is previously trained using a training set comprising multiple stacks of e-beam images corresponding to multiple sites of the given structure on one or more training specimens, each stack of e-beam images representative of the at least given layer of a respective site; and test data acquired from an electrical test performed at the multiple sites and related to actual yield of the training specimens, the test data respectively correlated with the stacks of e-beam images and used as ground truth thereof.