https://repositorio.ufba.br/handle/ri/7278 Thu, 07 May 2026 02:44:01 GMT 2026-05-07T02:44:01Z https://repositorio.ufba.br/handle/ri/43507 Título: Automação sísmica híbrida: integração de inteligência artificial e métodos determinísticos na análise de velocidades. Autor(es): Luz, Marcos Augusto Lima da Primeiro Orientador: Vasconcelos, Marcos Alberto Rodrigues Abstract: This research presents a hybrid seismic automation methodology that integrates artificial intelligence techniques with deterministic methods for the automatic analysis and estimation of seismic velocity fields. The main goal is to optimize hydrocarbon exploration by enhancing model accuracy and ensuring operational safety throughout geological interpretation. Traditionally, the construction of the velocity field relies on manual picking from semblance panels, a subjective and time-consuming procedure that demands expert interpretation, especially under noisy or geologically complex conditions. The proposed workflow combines statistical and machine learning approaches in a sequential and integrated manner. The process begins with a sample pre-clustering technique, responsible for the preliminary structuring of the data and for automatically determining the optimal number of clusters. Next, the joint application of the K-means++ algorithm and Principal Component Analysis (PCA) enables efficient dimensionality reduction, improving data coherence and representativeness. In the deterministic stage, the Dix equation is employed to convert RMS velocities into interval velocities, which serve as training data for a Multilayer Perceptron (MLP) neural network. This supervised model performs the final adjustment of the velocity field, ensuring physical consistency, smoothness, and monotonic behavior. The hybrid nature of the methodology arises from the synergistic integration between deterministic physical modeling and adaptive artificial intelligence prediction. The proposed approach was validated using both synthetic models and real seismic data from the Gulf of Mexico, demonstrating robustness, stability, and applicability across diverse geological scenarios. The results confirm that the hybrid seismic automation framework provides more realistic and continuous velocity models, substantially reducing human intervention and improving interpretive efficiency in complex exploration environments. Editora / Evento / Instituição: Universidade Federal da Bahia Tipo: Tese Mon, 30 Sep 0009 00:00:00 GMT https://repositorio.ufba.br/handle/ri/43507 0009-09-30T00:00:00Z https://repositorio.ufba.br/handle/ri/41125 Título: Extensão do critério de Barbieri em tomografia de tempos de trânsito não-linear. Autor(es): Vieira, Marcelo Querino e Silva do Prado Primeiro Orientador: Bassrei, Amin Abstract: This work offers two methods to evaluate the quality of traveltime tomography with regularization by derivative matrices and also to improve velocity models. Both methods are based on Barbieri's work, originally developed in medical tomography. The first, Barbieri Criterion (BC), considers forward modeling by straight rays, while the second one, Modified Barbieri Criterion (MBC), is ruled by Fermat's principle. Both use filtering by singular value decomposition to suppress dominant eigenimages by assuming that inversion algorithm errors are randomly located. Simulations with synthetic data showed that both methods had improved the solution in model RMS sense, even if high Gaussian noise was added to data. In general, MBC requires greater regularization than standard inversion. When applied to real data from Dom João Field, Recôncavo Basin, both methods had recovered similar models and with a higher resolution than the standard approach. Real data results were validated by seismic reflection data. Forward modeling was performed by ray tracing, based on analytical solution for differential ray equation, and by graphs, which is a simple application of Fermat's principle. Graph modeling proved to be superior as it always links sources and receivers regardless of velocity model. Numerical inversion was performed by generalized inverse and by a conjugate gradient method, which presented a lower computational cost without losing quality. Solution was stabilized by regularization by derivative matrices. Regularization factors were selected by L-curve and sine-Theta-curve, the latter developed in this work as an extension to the former, or by generalized cross-validation. Editora / Evento / Instituição: Universidade Federal da Bahia Tipo: Tese Thu, 28 Nov 2024 00:00:00 GMT https://repositorio.ufba.br/handle/ri/41125 2024-11-28T00:00:00Z https://repositorio.ufba.br/handle/ri/40427 Título: Imageamento sísmico: aplicação eficiente da RTM causal, FLSRTM com algoritmos SVD estocásticos e WEM baseado na aproximação de Jacobi-Anger. Autor(es): Apraez, Daniel Esteban Revelo Primeiro Orientador: Pestana, Reynam da Cruz Abstract: This thesis addresses the development of techniques related to seismic migration, aiming overcome the limitations and challenges present in traditional methodologies. The three chapters of this work present solutions that enhance computational efficiency, accuracy in seismic imaging, and the ability to handle complex subsurface models. In chapter 1, we focus on the separation of the wavefield into upgoing and downgoing components, a crucial step in the processing of multicomponent data, wavefield propagation, and seismic imaging. We propose an alternative method for computing the analytical wavefield using the first-order partial equation in time, solving the wave equation only once. This approach allows for the explicit separation of wavefields at each time step, making it more computationally efficient and enabling the application of the causal imaging condition in reverse time migration (RTM). Results from synthetic models indicate that the proposed method achieves a decomposition similar to that obtained by the conventional method, which requires two \-propagations, with potential applications in 3D cases. Moreover, the method effectively removes low-frequency noise present in RTM images that use the cross-correlation imaging condition. In chapter 2, we investigate frequency-domain least-squares RTM (FLSRTM), which is capable of producing high-resolution reflectivity models. However, storing the Green's functions needed for gradient computation and the scattered wavefield via Born modeling may be unfeasible due to their size. We propose a FLSRTM scheme using low-rank Green's functions obtained through randomized (rSVD) and compressed (cSVD) singular value decomposition algorithms. These algorithms allow for efficient storage of Green's functions in memory, using less space and reducing computational time. Evaluations on synthetic models demonstrate that the proposed scheme generates migrated sections identical to those produced by the FLSRTM scheme using the original Green's functions while utilizing less memory and computational time. In chapter 3, we address the limitations of conventional depth migration operators based on downward continuation of the acoustic wavefield, such as the undesired generation of evanescent waves, imaging in media with strong velocity contrasts and steeply dipping reflectors, and the stability of the one-way propagator. We propose a depth migration algorithm based on an one-way wave equation that is both stable and efficient. To achieve this, we use a spectral projector to suppress evanescent modes from the Helmholtz operator and apply the coupled Schulz iterative scheme to compute the square root of this filtered operator. Finally, we introduce the Jacobi-Anger expansion to approximate the exponential matrix operator, enabling the stable construction of the propagator. Impulse response tests, as well as field data applications, demonstrate that our algorithm is more accurate and effective for imaging in media with strong lateral velocity variations, surpassing the quality of images obtained by conventional methods. Editora / Evento / Instituição: Universidade Federal da Bahia Tipo: Tese Thu, 03 Oct 2024 00:00:00 GMT https://repositorio.ufba.br/handle/ri/40427 2024-10-03T00:00:00Z https://repositorio.ufba.br/handle/ri/39304 Título: Solução de sistemas lineares de grande porte e computação de alto desempenho Autor(es): Ariza Ariza, Cristian David Primeiro Orientador: Porsani, Milton José Abstract: This work describes a method for solving large, positive-defined, block-structured, dense linear systems with multiple right-hand sides that uses high-performance parallel computing. The system solution is obtained through a generalized Levinson recursion that uses the linear combination of smaller forward and backward solutions associated with lower order subsystems. The new implementation is described for parallel computing and is based on a partitioned matrix algorithm. The algorithm was separated into two subroutines, the first that computes the backward solution and the error energy matrix for smaller orders, and the second that computes the solution recursively. The algorithm was implemented for three types of systems: shared memory systems, distributed memory systems, and GPU systems. In each case, the lowest order systems were calculated using appropriate libraries. In the first, the OpenBLAS or MKL library was used; in the second, SCALAPACK; and finally, for systems with GPUs, we implemented an OUT-OF-CORE algorithm, in which the lowest order systems were calculated using MAGMA. In all three cases, the final solution is compared with the complete system solution using LAPACK, SCALAPACK, and MAGMA, respectively. In all three cases, the first part of the algorithm proved to be more computationally expensive compared to the Cholesky decomposition. However, the second part that computes the solution proved to be more efficient than the successive solution of two triangular systems when the right side of the system has a significant size, generally a few times the value of N. The error in the estimated model does not present significant variations compared to the reference solution. Finally, we present the use of the algorithm in frequency-domain seismic wave modeling, which involves the solution of large, sparse linear systems. These results show a disadvantage of the algorithm in sparse non-Toeplitz systems, as it increases the computational cost and memory consumption. Editora / Evento / Instituição: Universidade Federal da Bahia Tipo: Tese Mon, 15 Jan 2024 00:00:00 GMT https://repositorio.ufba.br/handle/ri/39304 2024-01-15T00:00:00Z