MSE Faculty Candidate Seminar: “Unsupervised Learning of Dislocation Motion”

High-performance designs that utilize metallic alloys are driving a need to quantify deformation in-situ at the finest length scales in order to reduce weight, increase operating temperatures, and improve fatigue life. With improvements to data reconstruction algorithms, brighter X-ray sources, and more efficient detectors, these in-situ studies of microstructural and micromechanical evolution in 3-D (nm-µm length scales) and at rapid time scales (<ms) are now possible. As numerous projections are often required for inversion of 3-D physics-based scattering models, trade-offs typically must be made between microstructural detail and the time scale probed. Instead, utilization of unsupervised learning, specifically locally linear embedding (LLE), is proposed to analyze in-situ diffraction data and find lower-dimensional embeddings that characterize microstructural transients, thus by-passing the need for a scattering model chosen a priori and enabling material understanding to be recovered with sparser data sets. The approach is applied to diffraction data gathered during uniaxial deformation of additively manufactured Inconel 625. The evolution of the lower-dimensional representation of microstructure is directly connected to the evolution of the defect densities that dictate strength and plastic flow behavior using a well-established material model. The implications of the findings for future constitutive model development and wider applicability to the study of material evolution during processing, particularly additive manufacturing, will be discussed.