Banca de QUALIFICAÇÃO: ALESSANDRO JOSÉ SOARES DANTAS

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : ALESSANDRO JOSÉ SOARES DANTAS
DATE: 20/07/2026
TIME: 09:00
LOCAL: Videoconferência
TITLE:

Use of AI Tools in the Application of FWI Techniques to Seismic Data with Variable  Topography

KEY WORDS:

FWI; WECI; GCN; Seismic processing

PAGES: 50
BIG AREA: Ciências Exatas e da Terra
AREA: Geociências
SUMMARY:

Land seismic data and OBN data have always been acquired in locations with variable  topography. Performing inversion with these data using FWI (Full Waveform Inversion) has always been a challenge, requiring static corrections to make the seismic data appear as if it were acquired  on a flat topography, thereby removing the topographic effect. Therefore, in this work, we apply  artificial intelligence tools, specifically the automatic differentiation available in the PyTorch  framework, to the inversion of seismic data with variable topography, without the need for static  corrections. To make this possible, we applied several inversion techniques, zeroing the gradient of  the objective function above the topography at each optimization iteration. The inversion method  used is FWI, an advanced seismic processing technique that consists of iteratively adjusting an  initial velocity model until the synthetic data resembles the observed data. Due to the high nonlinearity of FWI, the optimization process can get stuck in local minima, a phenomenon known  as cycle skipping. To mitigate this problem, four strategies were evaluated and compared:  conventional FWI, FWI with gradient smoothing, ECI (Envelope Correlation Inversion), GCN  (Global Correlation Norm), and WECI (Weighted Envelope-Correlation Inversion). The  experiments were conducted on the 1994 BP synthetic velocity model with variable topography,  using a 2.5 Hz Ricker wavelet as the seismic source and an array of 795 receivers distributed along  the topographic surface. The wave propagator used was DeepWave, integrated into PyTorch, which  enabled the automatic calculation of the objective function gradient for updating the velocity model  over 250 iterations. The results show that the envelope-based techniques (ECI and WECI) and the  global correlation technique (GCN) exhibit lower sensitivity to the cycle skipping problem compared to conventional FWI, demonstrating the potential of these approaches for seismic inversion in  scenarios with complex topography.

COMMITTEE MEMBERS:
Presidente - 1451214 - ADERSON FARIAS DO NASCIMENTO
Externo ao Programa - 1347984 - GERMAN GARABITO CALLAPINO - nullExterno à Instituição - REYNAM DA CRUZ PESTANA - UFBA