Additively Manufactured Component Characterization by Machine Learning from Resonance Inspection Techniques

The lack of reliable, nondestructive part qualification for additively manufactured (AM) parts hinders their adoption in key industries of national interest such as aerospace and defense. Resonant ultrasound spectroscopy (RUS) is a relatively low-cost and nondestructive method for accurately determining material properties. In this work, we explored potential applications for using machine-learning techniques to computationally speed up RUS in deriving the material properties of AM parts, as well as identifying print quality of parts post-build. We performed mode identification on 226 cylinders manufactured via laser powder bed fusion (LPBF) from an A20X alloy. A lack of visual separation in the data lead to the use of statistical and dimensionality reduction techniques with the resonance peaks as well as examining the overlaid resonance spectra. We then performed classic RUS using a genetic algorithm to find the density, Young’s modulus, and Poisson’s ratio. By constraining these three parameters according to porosity relations relating all three material properties and training a random forest regression from finite element analysis simulations within a range of representative values, we could predict the material parameters with a lower mean RMS than compared to those values resulting from the less-constrained genetic algorithm.