A physics-informed clustering approach for ultrasonics-based nondestructive evaluation

We propose a physics-informed clustering (PIC) algorithm tailored for ultrasonic non-destructive evaluation. Ultrasonic pulse-echo testing is used to measure the wave speed and wave amplitude decay of additively manufactured AlSi10Mg samples with programmatically induced pores and varying total volumetric porosities. The standard k-means clustering algorithm is coupled with the Independent Scattering Approximation (ISA) model to group together samples of similar porosity based on their ultrasonic response. The performance of the proposed PIC algorithm across varying seeds and numbers of clusters is compared to that of the standard k-means algorithm with random and k-means++ initializations. We demonstrate that the proposed PIC algorithm yields a more favourable clustering in terms of Pearson correlation coefficient and mean squared error given the ground-truth porosity labels. Our case study suggests that using a physics equation to inform a clustering algorithm can improve the clustering outcome.