An NSF project has been awarded to the Solid Mechanics group entitled “CAREER: A Multichannel Convolutional Neural Network Framework for Prediction of Damage Nucleation Sites in Microstructure”
Abstract: This Faculty Early Career Development (CAREER) program will aim to answer the longstanding question: What causes materials to fail? Structural materials are the building blocks of modern lifestyle, supporting applications ranging from infrastructure to national security, yet the mechanisms underlying their failure are not well understood. The first stage of catastrophic failure is often the initiation of small pores, through a complex and multifaceted process that has so far evaded simplified models. This project will leverage modern machine learning methods, which have the ability to identify subtle trends in large datasets, to unravel the nuances of pore nucleation. The result will be a computer model that is able to rapidly screen materials to determine damage susceptibility. The educational part of this project is twofold. First, the project will make advances in machine learning-based materials science accessible to budding and amateur scientists through an internet-based application, called “Solid Genius.” Solid Genius will be freely available, allowing direct manipulation and exploration of the model and its predictive ability through a user-friendly yet educational interface. Second, the project will develop a new graduate course to provide training for rising researchers in machine learning-based damage mechanics.
Experimental evidence indicates that grain boundaries are preferential sites for pore nucleation, but no meaningful correlations between grain boundary properties and failure likelihood have yet been conclusively established. A multi-channel convolutional neural network (MCCNN) machine learning framework is proposed that will be able to identify potential pore nucleation sites in pristine microstructure. The framework will simultaneously account for both nonlocal properties (microstructure, grain texture, etc.) and local properties (pointwise curvature, inclination, etc.), synthesizing them against a training dataset to produce a reliable estimator of failure likelihood. Training data will consist of reconstructed experimental micrographs and EBSD data, divided into “failure” and “no-failure” partitions. The raw experimental data will then be enriched with secondary calculations and supplemental mechanics simulations to supply non-visible channels such as grain boundary energy and mechanical stress. The trained MCCNN framework will then be used in concert with a damage mechanics model to further probe the early-time behavior of pore initiation and growth. Development will include an emphasis on determining physical interpretability of each aspect of the MCCNN model, such as formalizing the connection between individual convolutional layers and feature segmentation in microstructure, in order to facilitate a more rigorous application of the framework to the problem of damage assessment.