Session: 03-03-01: Applications of AI

Paper Number: 152112

152112 - Deep-Learning-Based Ultrasound Computed Tomography for Material Characterization 

Ultrasound imaging is a significant technique in non-destructive testing (NDT) and structural health monitoring (SHM). Methods such as reverse-time migration (RTM) and total focusing method (TFM) are some ultrasound imaging modalities that are widely used in detecting defects such as cracks, voids, and foreign materials. In addition, inversion-based ultrasound computed tomography (USCT) can identify the material properties (i.e., wave speeds, attenuation, density). Full waveform inversion (FWI) is one of the advanced inverse processes designed to quantitatively determine the properties (model parameters) of a target specimen by minimizing the discrepancy (misfit) between the acquired waveform signals and the synthetic waveform signals. The centerpiece of FWI is the adjoint method, which enables the calculation of cross-correlations between forward and back-propagated waveforms, facilitating the formulation of an iterative optimization problem to determine the optimal model parameters. As a result, FWI can produce high-resolution inverted images of the specimen. However, this process is computationally expensive and time-consuming due to its ill-posed nature and inherent nonlinearity. To mitigate these challenges and enhance the efficiency of FWI, an alternative approach using artificial intelligence (AI) is explored in this study. It aims to implement a neural network as a data-driven regularizer tailored to the problem. Applying the GAN to obtain the final models requires much less computation and is less time-consuming than classical FWI. However, one of the major challenges in such a data-driven approach is the inadequate amount of data to train the neural network. In order to address this bottleneck, 2D numerically generated samples are used in this study. The neural network is trained with paired images consisting of early-stopped-inverted (EFI) images and ground truth images of the sample. For evaluation, the structural similarity index measure (SSIM) and mean squared error (MSE) metrics are used to compare the predicted images with the ground truth images. Finally, the training model will be validated with experimentally acquired data. This study uses a cylindrical phantom with multiple inclusions of varying mechanical properties as the experimental specimen with the data acquisition carried out using a multi-channel data acquisition (DAQ) system from the Verasonics (Vantage system). The deep learning-based adjoint tomography architecture for USCT presented in this work has a strong potential for material characterization, providing a more efficient and accurate method for detecting internal defects and variations in material properties across a wide range of industrial applications.

Presenting Author: Shoaib Anwar The University of Alabama

Presenting Author Biography: Shoaib Anwar is a 4th year Ph.D. student in the Aerospace Engineering and Mechanics department at The University of Alabama. His research includes novel ultrasonic imaging methods for material/structures/medical imaging, and machine learning. He completed his undergrad in mechanical engineering at Bangladesh University of Engineering and Technology. He also has three years of working experience in an automotive company in Bangladesh