Using Perfectly Matched Layer in a GPU simulation of ultrasound NDT

Abstract

Full Waveform Inversion (FWI) is a state-of-the-art method of image reconstruction with promising results for seismology and NDT. FWI estimates the velocity of wave propagation along the inspected object by iteratively comparing acquired data with simulated data. This method is computationally demanding as at least three simulations are needed per iteration, which motivates the research effort for fast simulation methods. When simulating acoustic waves within a finite domain, artificial reflections created at the boundaries yield unrepresentative signal acquisition in the simulation. These reflections disturb the computation of residuals in FWI, creating artifacts that result in inaccurate positioning of flaws and discontinuities. Various absorbing boundary conditions (ABCs) have been proposed to avoid these reflections, e.g., Higdon, Reynolds, and Perfectly Matched Layers (PML). Although Reynolds and Higdon have been used in other applications, they are not well suited for FWI for reasons such as weak absorption rate and excessive introduction of amplitude errors. On the other hand, the PML method, an ABC that gradually attenuates boundary reflections with the addition of layers with matched acoustic impedance, provides adequate attenuation for FWI and can be readily extended to elastic waves. In this work, we describe an implementation of PML in a parallel GPU-accelerated CUDA simulator, which is intended to be integrated into an FWI regression in future works. Finally, we provide examples demonstrating that PML can significantly reduce the domain of simulation of an ultrasound NDT while maintaining simulation accuracy.