Deep learning-assisted prediction of mean grain size of polycrystalline materials from ultrasonic wave response

Abstract

This study introduces a novel approach for the non-destructive automated prediction of mean grain size in polycrystalline materials, using ultrasonic testing combined with deep learning (DL) techniques. The proposed approach involves a 1D convolutional neural network (CNN) regression model designed to analyze the ultrasonic longitudinal wave responses of Inconel-600 specimens, with the goal of predicting their mean grain size. These wave responses are generated through Hanning tone burst load excitation. Initially, simulated longitudinal wave responses are obtained through numerical simulations for eight distinct mean grain sizes (ranging from $150\mu m$ to $500\mu m$). Neper software is utilized to generate simulated microstructures with varying mean grain sizes, followed by finite element (FE) simulation using the commercial tool ANSYS-APDL. Subsequently, experimental wave responses are captured for Inconel-600 specimens with three distinct mean grain sizes ($20\mu m$, $67\mu m$, and $107\mu m$), prepared through the annealing process. The variation in mean grain sizes among the specimens is observed through optical microscopic images. The experiments are conducted on an in-house experimental setup, with piezoelectric wafer transducers used to generate and sense the experimental wave responses. In addition to distinct mean grain sizes, wave responses are captured by varying locations, frequencies, and noise levels to create a comprehensive and diverse database. The complete database comprises 1155 experimental and simulated wave responses across 11 different mean grain sizes. 80% of the complete database is randomly chosen for training a 1D-CNN regression model, while the remaining 20% is used for testing. The model’s architecture is optimized for predictive accuracy, incorporating convolutional layers, activation functions, and a Huber loss metric. Training and validation demonstrate the model’s ability to learn complex patterns and generalize to unseen data effectively. Testing on unseen datasets yields promising results in predicting mean grain size values, with the model achieving an average relative error (ARE) of 6.24%.