2021 Virtual Undergraduate Research Symposium

2021 Virtual Undergraduate Research Symposium

Machine Learning for Fracture Mode Classification in Rocks

Machine Learning for Fracture Mode Classification in Rocks

PROJECT NUMBER: 27 | AUTHOR: Bryan Dickson, Mechanical Engineering

MENTOR: Reza Hedayat, Civil and Environmental Engineering

ABSTRACT

In this project, Machine Learning was used to predict the type of fractures associated with the uniaxial compression testing of brittle rocks. Fractures are commonly identified in the forms of Tensile, Shear, and Mixed. The fracture type can be recognized through the waveforms generated during fracturing. The waveform features are used to train machine learning models and then these models are tested with unseen data. The project objective is to identify a machine learning algorithm that can successfully predict the fracture mode based on the waveform features.

PRESENTATION

AUTHOR BIOGRAPHY

Bryan Dickson is a Sophomore in Mechanical Engineering Pursuing a minor in Metallurgical and Materials Engineering as well as Advanced Manufacturing. The research project “Machine Learning for Fracture Mode Classification in Rocks” is done through the Civil Engineering Department to help use Machine Learning to predict fracture types based on their properties. The research was done using Python and add-on tools to train and test Machine Learning Classification models on these experiments. In the future, Bryan wants to be able to work more with materials, design, or manufacturing, but enjoys learning new and potentially relevant subjects and strategies.

4 Comments

  1. Interesting work Bryan! When you performed oversampling, did you experiment at all with different parameter choices in the SMOTE and ADASYN algorithms (for example, changing k=5 in SMOTE to several different values). If so, did this have a significant impact on the results you observed?

    • Hello, we did try changing the parameters for the oversampling techniques. The other changes made very little difference to the algorithm, and the standard yielded the best and most reliable results. Thank you for asking about that because adjustments like that can be easy to overlook since I’m learning a lot of this as well still too.

  2. Hi Bryan!
    Sounds a really interesting technique in order to predict how a rock will fracture. Just for clarification, having the imbalanced data was negative to the machine learning but I wondering if the same data was provided later on after the machine learning was more calibrated would it be accurate?

    • Hi Annaliese,
      Yes, the imbalanced data was causing the predictions to be less accurate. And no, based on what we have done I do not believe that would be the case. The way the training worked without balancing the data set was that the model would only predict Fracture Mode 1. So the testing set shown shows how the model would realistically react to a novel set provided later on. Balancing the data to make it so that the training set has about the same number of each fracture mode actually yields better results across all fracture modes with the unbalanced set still. Sorry if I worded that strangely, all I mean is that the balanced data yielded better results in both cases than using the standard data.

Share This