A Deep Learning Framework for the Spatiotemporal Feature Extraction and Statistical Characterization of Terabyte-Scale XCT Datasets

Materials science faces a bottleneck in the ability to analyze large-scale image data. Datasets produced by modern imaging modalities such as X-Ray Computed Tomography (XCT) at synchrotrons output Terabytes of data per sample. Traditional approaches that use commercial software fail to effectively scale, while classical machine learning requires labeled data which can be impossible to obtain depending on the expertise required and volume of features present. As a result, analysis is reduced to hand-crafted features and small subsets of data which introduce human bias and only captures region-specific defect interactions as opposed to sample-wide behavior. To solve this, we have developed a framework that leverages distributed and high performance computing and machine learning (ML) to build an automated pipeline for the translation of 2D XCT images into 3D spatiotemporal graph representations of all microstructural features of interest. This graph-based representation provides a full statistical characterization of all defects within a given sample volume and enables additional downstream analyses.<br/><br/>We apply this spatiotemporal graph (st-graph) framework to XCT scans characterizing stress corrosion cracking (SCC) in Al-Mg alloys during a slow-strain tension test. The tests were conducted with collaborators at the Diamond Light Source on field-retrieved Al-Mg plate material. Samples had experienced 42-years of real world exposure to determine the effects of long-term service on stress corrosion cracking. Our st-graph pipeline 1) segments all fractures, precipitates, and pores on Terabytes of scans 2) provides a complete statistical characterization of all features of interest 3) constructs a spatiotemporal graph representation of the microstructural defect profile. We demonstrate the ability to extract over 150,000 features in a single scan, build a complete microstructural feature profile, and derive novel insights into degradation patterns from the interactions of features across a sample.