Qualificação de Doutorado

Programa: CPGEI
Aluno: Daniel Rossato 
Orientador: Daniel R. Pipa
Co-orientador: Thiago A. R. Passarin
Quando: 02/jun/2023 às 09:00
Link: https://meet.google.com/yxx-ssso-trx

Título: Parallel Full Waveform Inversion in Ultrasound Non-Destructive Testing

Abstract:
In Non-Destructive Testing (NDT), ultrasound testing is widely used as a way to obtain details about subsurface structures in specimens. The state-of-the-art technique for ultrasound image reconstruction using transducer arrays is the Total Focusing Method (TFM), which needs some a priori information about the specimen, and does not model non-linear phenomena such as multiple reflections. In seismic prospecting, surveys are conducted in a similar way to ultrasound testing: multiple geophones capture the waves reflected and refracted through the interest area. For image reconstruction, one promising method is the Full Waveform Inversion (FWI), which aims to match the observed data to the synthetic data generated from a guess model. Although FWI was proposed decades ago, its computational cost did not allow it to be applied in practical settings until recently. This work proposes the application of FWI to ultrasound imaging in NDT using transducer arrays, as well as analyzing and solving some of the difficulties that arise from the differences in scales and methods between ultrasound testing and seismic surveys.  This work utilizes the advent of modern Graphical Processing Units (GPUs) to parallelize both the FWI algorithm and the simulations needed for model evaluations, at the same time occupying less memory than the original approach, using the CUDA API. To mitigate cycle-skipping, a phenomenon that severely degrades FWI performance, the Wasserstein distance is used as the misfit metric. The results obtained using synthetic data outperform TFM in most cases, reconstructing images with good details in their subsurface structures even when materials with high acoustic impedance differences are present, such as water and steel. This work concludes that FWI can be successfully applied to NDT imaging, potentially obtaining much more detailed models of tested specimens.