Simulated Specimen Images for Neural Network Defect Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The generation of training data for artificial neural networks to identify defects in crystalline materials is time-consuming, as it requires producing crystalline materials with defects, which can be rare or difficult to produce, and involves capturing specimen images using microscopes.
Innovation Solution
The use of simulated specimen images generated based on data models describing crystalline material characteristics and defect types, allowing for efficient training of artificial neural networks without the need for actual specimen images, enabling focused training on specific defect types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If actual specimen images are used for training, then the training data reflects real defect characteristics, but the data generation process is time-consuming and requires rare defect production
Solution Approach 1:
The patent creates simulated specimen images that copy the essential characteristics of real defect appearances without requiring actual defective specimens. The simulation model generates training images by rendering defect patterns based on material properties and imaging parameters, providing a copy of the visual information needed for training while avoiding the time-consuming process of producing and imaging real defective materials
Solution Approach 2:
The patent performs preliminary generation of training data by creating simulated specimen images before actual defect analysis is needed. The simulation model pre-generates diverse defect patterns with known characteristics, allowing the neural network to be trained in advance on comprehensive defect variations without waiting for rare real defects to occur or be produced
2Reliability
If a large set of training data is generated, then the accuracy of defect identification improves, but the time and effort required for data production increases
Solution Approach 1:
The simulation model efficiently generates large volumes of training data by computationally rendering defect patterns rather than physically producing specimens. This copying approach allows rapid generation of diverse defect scenarios including rare and edge cases that would be impractical to obtain through actual specimen preparation, thereby improving both data quantity and generation efficiency
Solution Approach 2:
The patent varies multiple parameters in the simulation model including material properties, defect characteristics, and imaging conditions to generate diverse training data efficiently. By systematically changing these parameters, the system produces a large set of varied training images without the linear time cost of producing each additional real specimen, thereby improving productivity while maintaining data quality
Data Source
AI summary
Techniques for training an artificial neural network (ANN) using simulated specimen images are described. Simulated specimen images are generated based on data models. The data models describe characteristics of a crystalline material and characteristics of one or more defect types. The data models do not include any image data. Simulated specimen images are input as training data into a training algorithm to generate an artificial neural network (ANN) for identifying defects in crystalline materials. After the ANN is trained, the ANN analyzes captured specimen images to identify defects shown therein.