Inhomogeneous Material Property Characterization using Physics-Informed Neural Networks

Abstract

For quality control and evaluation of the concrete infrastructure, in-place characterization of material properties is important. There is a need for reliable and universal nondestructive test methods that can be applied to characterize structural members in situ in a rapid and efficient manner. However, large amounts of test data are normally needed to achieve such characterization goals. Furthermore, construction materials such as concrete, rock, and asphalt are challenging materials because of their inhomogeneity and complexity. Advances in machine learning have helped to solve many problems across different fields that rely on large data sets. More recently, physics-informed neural networks (PINN) have appeared, and they can overcome traditional problems of conventional machine learning methods. PINN is a particular form of artificial neural networks (ANN) and portend notable advantages over traditional measurand analysis or purely data-driven approaches, where physics-based equations are embedded within an ANN structure in order to regularize the outputs during the training process. Here, we explore the potential of in-place concrete characterization using physics-informed neural networks (PINN) and ultrasonic wave data. First ultrasonic wave data are obtained from experiments on long rod-shaped mortar sample and PINN is used to predict spatially varying wave velocity. Then, the proposed method was applied to two concrete slab samples: one is pristine and the other one has inclusion defects. PINN was able to predict spatially varying wave velocity including inclusion defects. The presented results demonstrate the promise of PINN to assist with inhomogeneous material (i.e., concrete) characterization methods.