Session: 16-01-01: Poster Session: NSF-Funded Research (Grad & Undergrad)

Paper Number: 99580

99580 - A Multichannel Convolutional Neural Network Framework for Prediction of Damage Nucleation Sites in Microstructure 

Structural failure mechanisms are a major aspect in the project design of high performance components for many industrial, military, and medical applications.  The void nucleation caused by high load impact or high temperature and pressure variations is a problem that remains an open challenge, as its development occurs in seemingly unpredictable locations in the materials' microstructure. Fully understanding this nucleation process is essential to developing safer and more efficient structures. Increasing the performance of these materials, will be possible to use lighter and safer elements in engineering applications reducing costs and creating new commercial possibilities. It has been shown that the pore formation is more likely to occur on the grain boundaries, although a specific correlation between the grain boundaries and pore formation has not been found. Current computational methods that simulate material behavior use big datasets to describe materials and require high computational power. In those models, Molecular Dynamics (MD) are used to simulate material response stress, such as, heat and pressure, and a mapping of the grain boundaries can be created, analyzing the aspect of the grain movements and formations. In this work, we propose the use of machine learning (ML) along with characterization datasets to identify void formation locations. Combined with computer vision methods, our models allow for rapid and efficient analysis of components. In ML, deep learning has improved the learning capacity of neural networks, for the reason that there are no visible layers. A model that has been widely used is Multi-Channel Convolutional Neural Networks (MCCNN). MCCNN is a kernel based method with relatively low computational cost that is efficient for capturing small feature changes. MCCNN showed good accuracy in predicting characteristics in different fields, such as face recognition and cancer detection. MCCNN uses unsupervised learning to establish convolutional layers and their connections identifying relevant features without human interference. This ML method combines similar representations and sparse connections to classify the input image, synthesizing them against a training dataset to produce a reliable estimator of failure probability.  MCCNN is used to identify features in the beginning state of materials that converge in failure formation. Training data was EBSD data and experimental reconstructed micrographs, divided into “failure” and “no-failure” clusters. We used 2D images of grain boundary dataset of FCC, BCC, and HCP materials . The accuracy was good in train and test applications. Also we are determining the physical interpretability of each aspect of the MCCNN model. 

Presenting Author: Bruno Manoel Dobrovolski University of Colorado Colorado Springs

Presenting Author Biography: Bruno holds a computer engineering degree from the Federal Technological University of Paraná - Brazil. He has worked as a software developer and with project optimization for the past 5 years. He is currently working on his Ph.D. in mechanical and aerospace engineering at the University of Colorado Colorado springs

Authors:

Bruno Manoel Dobrovolski University of Colorado Colorado Springs
Brandon Runnels University of Colorado Colorado Springs