Nondestructive Evaluation of Material Properties using Physics-Informed Neural Networks and Mechanical Waves
Advisor: Professor John S. Popovics
Abstract
Characterizing in-place material properties is essential for quality control and condition assessment in construction and manufacturing. Recent advances in machine learning tools have increased the popularity of data-driven models to predict complex phenomena. Artificial neural networks (ANNs), which are popular machine learning tools, have achieved impressive results in many different fields and tasks. However, data-driven models have limitations such as the large size and complexity of ANNs needed to capture realistic behaviors and the substantial amount of training data required, which significantly increase computational cost. Physics-informed neural networks (PINNs) offer a promising alternative that combine the strength of physics-based models with the flexibility of data-driven approaches. This dissertation presents a method for characterizing material properties using PINNs and mechanical wave-based data. The effectiveness of the proposed PINN models is evaluated using mechanical wave data from a variety of specimens, each with unique defects, material properties, and 1-D (rod), 2-D (plate) and 3-D (large slab) geometries that contain simulated defects. Bar waves, Lamb waves and Rayleigh surface waves are considered. Material properties such as wave velocity, quality factor, Young’s modulus and shear modulus are then predicted using PINN models. Mechanical wave propagation data are collected using a contactless sensing method, and the PINN models predict space-dependent wave velocity or Young’s modulus from these data, and the defects are indicated as low material property regions. Based on the material characterization work, several challenges to real-world applications were identified and addressed. One such challenge is the lift-off distance variation between a receiver and a targeted sample with contactless sensing. A laser displacement sensor method was introduced to correct this variation and obtain precise wave velocity profiles. Additionally, a PINN algorithm was proposed that can be used for sparse wavefield data, which is beneficial for reducing data collection time in the field. To enhance PINN performance with limited training data, additional residual points were incorporated. Finally, a PINN architecture capable of processing multiple datasets was also developed to overcome field data quality and accessibility issues. Throughout this study, the developed PINN models show superiority over traditional signal processing methods and purely data-driven methods because the models are able to predict multiple inhomogeneous properties such as wave velocity, Young’s modulus, etc. using a single measurement dataset on different types of structures.
