Scale Bridging in Ductile Fracture via Kernel-based Machine Learning

点击翻译按钮获取中文摘要

Ductile metals such as aluminum and steel are used ubiquitously across a variety of industries such as energy, defense, and transportation, and thus impact everyone’s daily life. Unfortunately, the mechanisms by which ductile metals fail or crack are poorly understood, making it difficult to accurately predict when fracture will occur. These knowledge gaps drive up the costs of products, decrease their energy efficiency, and limit our ability to develop stronger, more fracture-resistant materials needed for future energy and defense applications, e.g., high-temperature materials. The goal of this project is to formulate a physics-informed, machine-learning-enabled model for ductile fracture that will lead to reduced product costs and higher energy efficiency. For example, the ability to produce lighter vehicle structures would reduce vehicle carbon emissions while simultaneously reducing fuel costs for consumers. The project will promote science and engineering to undergraduates through research projects and enriched classroom instruction, and to high school students through the development of new modeling- and machine learning-focused modules for high school classrooms. In collaboration with a high school educator, these learning modules will be piloted in a high school physics classroom. Models of ductile fracture via nucleation, growth and coalescence of voids are largely phenomenological with many uncertain parameters that are difficult to determine. Under this project, a predictive, micromechanically informed model for void nucleation will be developed by coupling molecular dynamics simulations with kernel-based machine learning. Molecular dynamics simulations will reveal the fundamental mechanics underlying void nucleation at hard particles while quantifying the void nucleation rate. This rate is affected by a large set of features, including stress, temperature, and defects such as vacancies, solutes, and dislocations. To render model development tractable, a set of kernel-based machine learning models and algorithms will be formulated with two objectives: (i) to extract and engineer a minimally sized set of physical features for quantifying nucleation rates in microstructures, and (ii) to probabilistically establish an accurate closed-form statistical mapping between this microstructural feature set and the nucleation rates. Finally, the resulting machine learning model will be implemented into the finite element method as part of the commonly used Gurson-Tvergaard-Needleman model for ductile fracture. This effort will serve as a prototype for the use of machine learning towards upscaling of nano- and microscale information to engineering scale models.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.