Banca de DEFESA: RAMON DE OLIVEIRA SILVA JUNIOR

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : RAMON DE OLIVEIRA SILVA JUNIOR
DATE: 24/04/2024
TIME: 11:00
LOCAL: Plataforma Google Meet
TITLE:

Source wavelet estimation with machine learning and generalized seismic wavelets

KEY WORDS:

Seismic inversion, Machine Learning, Wavelets

PAGES: 71
BIG AREA: Ciências Exatas e da Terra
AREA: Física
SUMMARY:

Source wavelet estimation from seismic data is an important task in the subsurface imaging effort. In this study, we intend to test a Machine Learning (ML) approach to obtain the wavelet parameters from seismograms. For the sake of simplicity, it is usually assumed that the Ricker wavelet is a good approximation for the seismic source waveform. However, the attenuation caused by propagation in heterogeneous media and specific water phenomena distort the frequency content of the wavelet, an effect not adequately represented by Ricker. To correct this phenomenon, we work with the Generalized Seismic Wavelet (GSW), which is a recently proposed generalization of the Ricker wavelet. The Ricker has just one parameter, the fundamental frequency. On the other hand, the GSW has two parameters: the same fundamental frequency and the order of the fractional time derivative of the Gaussian function; the latter controls the wavelet’s shape and asymmetry in time domain. For a specific value of this additional parameter, the GSW becomes the Ricker wavelet. Using the seismogram as input, machine learning models are trained to output the generalized wavelet parameters that best explain the data. The source wavelet is then estimated as the WSG whose parameters are those provided by the model. The proposed methodology was tested and validated with synthetic seismic data, in which a marine acquisition in deep waters was modeled, using a simplified velocity model obtained from Gato do Mato, an oil field located in the Santos basin, Brazil. We propagate an acoustic wave using the following acquisition geometry: one source placed at the surface and 50 receivers placed at 2100 m in depth, spaced 100 m apart from one another in the horizontal direction. We performed one thousand numerical propagations using the GSW with different parameterizations as the source waveform. Furthermore, we randomly selected 70% of the data for training, and the other 30% for testing. Three ML algorithms have been considered for regression: Extreme Gradient Boosting (XGBoost), Bayesian Ridge  and the Random Forest. The results were satisfactory, achieving good accuracy in estimating the GSW parameters.

COMMITTEE MEMBERS:
Externo à Instituição - RAMON CRISTIAN FERNANDES ARAUJO - UFRN
Presidente - 1379465 - GILBERTO CORSO
Interno - 2492756 - JOAO MEDEIROS DE ARAUJO
Externo à Instituição - SERGIO LUIZ EDUARDO FERREIRA DA SILVA - POLITO