AbstractsEngineering

The use of quantified microstructures as inputs to neural network models for property prediction has been pioneered by Center for the Accelerated Maturation of Materials (CAMM) at The Ohio State University. Through microstructure-property correlations, neural network models provide predictive tools for mechanical properties in titanium alloys while concurrently developing phenomenological models. The output accuracy (mechanical property prediction) of such models is therefore dependent on the variance and distribution of the input data (quantified microstructures). An estimation of the true variance and distribution can be calculated if a sufficiently large sampling volume of quantifiable microstructural features is available; however, current manual image processing and segmentation techniques made attainment of large-dataset image-processing unfeasible. In this work, a new generation of automated tools has been developed by CAMM which have reduced the total analysis time, including image capture, processing, and characterization to less than 30 seconds per micrograph for optically captured micrographs. Using a comparable SEM-based technique requires less than 6 minutes per micrograph due to extended image capture times. Serial sectioning of a meso-scale 3D volume (mm3) of a+ß processed Ti-6Al-4V was collected for direct 3D quantification. Images were captured two ways: (1) using a Leica optical microscope in conjunction with Clemex image analysis software and (2) using a FEI Sirion SEM. In both cases, CAMM developed image processing package MIPAR was used to calculate the spatial variation in globular a area fraction. Comparisons between the two image capturing methods reveal similar trends in spatial variation indicating SEM-based imaging is only necessary if required by the scale of the particular microstructural feature of interest. A total of 37,800 micrographs were captured and processed. The large number of micrographs allows for accurate quantification of the variance and distribution of globular a volume fraction in a+ß processed Ti-6Al-4V via application of bootstrap confidence intervals. The globular a volume fraction will be combined with other quantified microstructural features in the future to provide inputs for neural network models.