Simulação de ultrassom aplicada a ensaios não-destrutivos

Abstract

One of the main non-destructive testing (NDT) techniques is ultrasound inspection and imaging, widely used in the oil and gas industry to verify the integrity of subsea equipment, parts, and welds. Inner surface imaging, a primary application, enables the evaluation of metallic structures and pipes. Although seemingly unrelated, seismic imaging shares the same foundation as ultrasound imaging: the propagation of mechanical waves. In seismic imaging, Full Waveform Inversion (FWI) reconstructs images of seabed layers through an iterative method that compares acquired signals with simulated signals. This method updates the simulation’s propagation medium until the differences between simulated and acquired signals meet stopping criteria, resulting in a model characterizing the studied region’s composition. FWI has the potential to produce more detailed images than classical Delay-and-Sum (DAS) methods by considering the physics of wave propagation. However, for FWI to be practical in NDT, fast and accurate wave propagation simulations are essential, as they constitute the most computationally demanding part of the regression. This dissertation examines graphics processing units (GPUs) for acoustic wave simulation with absorbing boundary conditions (ABCs) to prevent artificial reflections at domain borders in simulators for FWI. The theory of acoustic waves, Perfectly Matched Layers (PML), Convolutional PML (CPML) ABCs, and the finite difference method for solving the wave equation are studied. Implementation strategies and testing methodologies are detailed, with comparisons between simulations and analytical solutions, both formulations of the acoustic wave equation, and the speed-up achieved by GPU implementations. Additionally, the reflection absorption by PML and CPML for the same setup and the domain reduction achieved by using ABCs are compared. The results show a mean squared error (MSE) residue between simulations and the analytical solution in the range of 10−4 to 10−3, faster simulations with GPU PML achieving a 70.24x speed-up, and GPU CPML achieving a 77.77x speed-up, FWI regression in 1600 seconds, reflection absorption of PML at 99.80% and CPML at 99.92%, and a domain size reduction of 45.13%. Finally, the simulations are compared to real ultrasound acquisitions, and the challenges and future directions of NDT ultrasound simulation for FWI are discussed.