Real-time grain orientation mapping of anisotropic media for improved ultrasonic non-destructive evaluation

Estimating the spatially varying grain orientations of locally anisotropic media non-destructively is important for the accurate detection of flaws. Conventional optimisation algorithms used for solving this inverse problem come with significant computational cost and are therefore not suitable for near-real-time monitoring applications. Here we propose a framework which uses deep neural networks to learn the relationship between ultrasonic travel time data and grain orientation maps. For comparison we also employ an optimisation framework based on surrogate optimisation methods. Following a computationally expensive training process, the deep neural network predicts grain orientation maps in near-real time (0.15 seconds). In comparison, the surrogate optimisation takes approximately 32 minutes per model inversion (for a standard desktop computer). In most cases studied here, the deep neural network also outperforms the surrogate optimisation in terms of reconstruction accuracy,