| dc.description.abstract | In an engineered composite material, the behavior of different phases and their interactions determine the response of the material as a whole. Data-driven machine learning techniques have emerged as a promising tool for mapping the morphology of a composite material to the loading response of the material using data from experiments or simulations. Such an approach allows for identification of patterns in high-dimensional material descriptor spaces, which is particularly useful for analysis of composites that exhibit randomness or nonlinear behavior. In this work a machine learning framework is laid out to classify composite structures and to predict their behavior. To demonstrate the flexibility of the proposed approach, it was implemented in problems from two unique fields. In the first problem, the response of a random short-fiber reinforced composite subjected to static uniaxial tension was predicted by generating random microstructures, simulating their response using an extended finite element method developed for this purpose, and learning the composite response using Gaussian process regression. The second problem considers the vibrational characteristics of wound electric guitar strings with a range of geometries and material compositions. An experiment was designed to measure the response of strings from different classes. Using the test results a support vector machine was trained to classify strings based on the metrics of the resulting signals. | |
