End-to-end ultrasonic evaluation of grain size in aeroengine disk using a spatiotemporal attention network

Abstract The safety of dual performance aeroengine disks depends on their internal grain size, requiring accurate nondestructive evaluation (NDE). A key challenge is interpreting the rich physical information in the backscattered ultrasonic wavefield, specifically the frequency dependent features of Rayleigh scattering and the wavefield’s spatial coherence. Conventional methods using simplified 1D or 2D data fail to capture this critical spatiotemporal information. This study proposes the Scatter Spatiotemporal Attention Network (SSTAN), a hybrid deep learning architecture designed to map raw 3D ultrasonic data directly to grain size. Its multi scale 3D CNN component acts as a learned filter bank for frequency dependent scattering, while its transformer encoder models the non-local correlations in the coherent backscattered wavefield. Tested on a nickel base superalloy dataset, the framework achieved 85.3 % classification accuracy. This result shows a notable improvement over deep learning benchmarks using lower-dimensional data (78.8 %) and an advanced feature engineering method (83.5 %). The results indicate that the SSTAN architecture is a promising, physically-informed approach for challenging ultrasonic NDE tasks.