Method and system for training an artificial neural network utilizing physics based knowledge
| DWPI Title: Method for training artificial neural network for material characterization, involves obtaining functional data having phase and amplitude, and performing dimensional reduction on aligned functional data to produce dimensional representation of functional space |
| Abstract: A method for training an artificial neural network including classifiers for material characterization. The method includes obtaining functional data having phase and amplitude, registering functional data by phase-amplitude separation and statistical analysis on the phase-amplitude separated data with an elastic distance to produce aligned functional data, performing dimensional reduction on the aligned functional data to produce a dimensional representation of the functional space, performing, by a computer system, a training operation to train an artificial neural network based on the dimensional representation of the functional space. A method and system for material characterization is also disclosed. |
| Use: Method for training an artificial neural network for material characterization. |
| Advantage: The method achieves a greater than 30% improvement in overall classification accuracy when compared to the purely data driven method using the unprocessed H-CT voxels to train a purely data-driven method using a 1-D VGG like convolutional neural network (CNN). The high-energy spectral CT can enhance performance by generating higher-quality images, while traditional CT methods using Bremsstrahlung radiation sources suffer from noise and artifacts due to lack of penetration or nonlinearities in material absorption (i.e., beam hardening). |
| Novelty: The method involves obtaining (101) functional data having phase and amplitude. The functional data is registered by phase-amplitude separation of the functional data to produce separated phase and the amplitude components with an elastic distance. A statistical analysis is performed on the components to produce aligned functional data. A dimensional reduction is performed to produce a dimensional representation of a functional space of the aligned data. An artificial neural network is trained based on the representation of the space by a computer system e.g. personal computer, where the data is voxel data from an x-ray computed tomography scan. |
| Filed: 1/19/2022 |
| Application Number: US17579324A |
| Tech ID: SD 15591.0 |
| This invention was made with Government support under Contract No. DE-NA0003525 awarded by the United States Department of Energy/National Nuclear Security Administration. The Government has certain rights in the invention. |
| Data from Derwent World Patents Index, provided by Clarivate All rights reserved. Republication or redistribution of Clarivate content, including by framing or similar means, is prohibited without the prior written consent of Clarivate. Clarivate and its logo, as well as all other trademarks used herein are trademarks of their respective owners and used under license. |