Interpretable Wave-Informed Machine Learning for Defect Detection

Improving Machine Learning interpretability through physical principles is pertinent to a wide variety of areas in physics research and applications. We present a novel algorithm for wave-informed learning, currently applied to defect detection. SiCf-SiCm is a leading candidate for next generation nuclear reactor fuel cladding due to its structural stability in extreme environments. However, heterogeneities formed during the complex manufacturing process can cause uncertainty in performance. Therefore, reliable, efficient, and interpretable defect detection is crucial to ensure cladding quality. We introduce a wave-informed machine learning framework to separate spatially and temporally heterogeneous modes from ultrasonic waves induced into the specimen of interest. We can then identify which modes differ from the baseline wavefield. These variations highlight and characterize heterogeneities in the material. In this presentation, we outline the mathematical and algorithmic framework. We include initial results in flat, isotropic, and homogeneous media, such as metal plates. Finally, we discuss future plans and further applications of our method in other areas of physics.