Developing neural networks to rapidly map crystallographic orientation using laser ultrasound measurements

Abstract

Rapid measurement of crystal orientation is critical in the materials discovery process as it facilitates real-time decision-making and quality control. Acoustic inspection methods rapidly characterise microstructure without the need for extensive infrastructure or expense – the laser ultrasonic method known as Spatially Resolved Acoustic Spectroscopy (SRAS) has been developed with this intent and accurately characterises crystal orientation by leveraging a combination of forward modelling and an exhaustive brute force process to obtain the best-fit orientation. While effective, this method is computationally demanding and time-intensive. We introduce a novel approach that utilises neural networks to classify measured acoustic signals into orientation planes to significantly expedite the characterisation process and demonstrate classification on real-world Inconel 617 and CMX4 specimens. A reduction in the orientation determination time from around 10 hours (brute force search) down to 15 seconds (neural network) was achieved while exhibiting an average plane angle difference of between 5.3^∘ and 13.8^∘.