Accelerating Material Discovery and Analysis Using Machine Learning
The big data revolution is only just beginning in the materials science and engineering field, offering the promise to enable high-throughput workflows and accelerate material development. For this to be realized, a new set of tools capable of using this data for identifying better material candidates and assisting in the analysis of samples must be developed. Currently, most material development projects require manually searching vast composition space while relying primarily on expert domain knowledge. While this strategy has been reasonably effective throughout history, emerging technologies are placing rigorous demands on material performance. Furthermore, many of the simplest combinations of elements (i.e. typically one primary element and one to three minor alloying elements) have been thoroughly evaluated. Within the last two decades, the materials science community has become interested in high entropy materials, typically containing five or more cations in near-equimolar amounts. This compositional space is largely unexplored and challenging to model with traditional tools, making it an excellent use case for machine learning. In the first part of the dissertation, machine learning tools are developed and demonstrated for predicting the relative synthesizability of high entropy and ultrahigh ceramics as well as the prediction of crystal structure for alloys. These machine learning approaches were validated by experiments and comparison with computational predictions where possible. The second part of the dissertation details progress in accelerating post-fabrication aspects of material design frameworks. For material analysis routines, deep neural networks were constructed to analyze diffraction patterns and determine symmetry and/or structure of the phases present. The electron backscatter diffraction (EBSD) platform was used to demonstrate the ability for such algorithms to not only identify symmetry, but that neural networks are also a potential solution to some of the grand challenges state of the art EBSD software cannot solve. This includes space group identification and the phase mapping of materials with similar crystal structures. Further validation is performed to demonstrate the model’s reliability with changing acquisition parameters. Combined, these machine learning-based tools represent an opportunity to reduce the time and expertise required for material development and are likely to become valuable components in high-throughput workflows.