Machine learning predictions of crack paths in brittle and ductile media

Despite their modern widespread use, additively manufactured (AM) materials suffer from a frequent problem: defects. Defects, such as “lack of fusion” or “keyhole” defects, result in a porous, inhomogeneous final material containing a distribution of voids. These void defects control the fracture, and thus failure, behavior, limiting the application of AM materials. In this thesis, we first develop machine learning techniques using artificial neural networks (ANNs) to predict how voids affect crack growth in an algorithmically-described brittle-like material and apply this to predict crack paths as a proof of concept. The ANNs are trained using inputs describing the porosity distribution surrounding the crack tip and outputs representing the ensuing crack growth direction. We then extend this technique to porous ductile media, representing defect-filled AM materials. Results showed excellent performance in predicting both individual steps of crack growth and full crack paths for both brittle and ductile media. Individual crack growth directions were predicted with over 98% accuracy for brittle media and accuracies neared 90% for ductile media. Common prediction errors for both brittle and ductile fracture resulted from higher void densities and insufficiencies in the ANN training data for particular types of void configurations. Multi-step crack path predictions were in good alignment with the true paths. The high predictive performance for crack paths allowed for the design of porosity distributions with an evolutionary genetic algorithm to achieve crack propagation along desired crack paths. A limited ability to design for a targeted fracture toughness was demonstrated. However, it was noted that ductile fracture is governed by multiple mechanisms. The mechanism accounted for in this thesis is primarily void-by-void, where the crack tip interacts with a single void at a time. However, it was found that many cracks are governed by multiple void interaction, where multiple voids grow simultaneously, creating multiple damage zones that eventually coalesce. Training of the ANN to account for only void-by-void crack growth thus limited its ability to predict full ductile crack paths. Extensions to more accurately account for multiple void interaction mechanisms are discussed, and is a continuing subject of future work.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Aerospace Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Worthington, Michael

Contributors dc:contributor

• Chew, Huck Beng

• Machine Learning
• Artificial Neural Networks
• Warp3d
• Crack Growth
• Void Defects
• Micromechanics Model
• Crack Patterns
• Crack Paths
• Additive Manufacturing
• Genetic Algorithm
• Brittle Fracture
• Ductile Fracture
• Fracture Toughness

Statement dc:rights

• Copyright 2022 Michael Worthington

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/117718

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F117718