Improving the seismic image in reverse time migration by analyzing of wavefields and post processing the zero lag cross correlation imaging condition
Author: Juan Guillermo Paniagua Castrillón.
Research project with Ecopetrol and Colciencias: develop algorithms of seismic migration using wave field extrapolation in the direction of time (RTM, Reverse Time Migration), evaluating the preservation of amplitudes and frequencies as well as the conditions of stability, numerical dispersion and computational cost.
An approach to the study of time, time-frequency and time-scale transformations for seismic migration problems
Author: Juan Guillermo Paniagua Castrillón. Research project with Ecopetrol and Colciencias: seismic pre-stack migration in depth by extrapolating wave fields using high performance computing for massive data in complex areas.
Laguerre Gaussian filters in Reverse Time Migration image reconstruction
Author: Juan Guillermo Paniagua Castrillón.
Reverse Time Migration (RTM) solves the acoustic or elastic wave equation by means of the extrapolation from source and receiver wavefield in time. A migrated image is obtained by applying some criteria known as imaging condition. The zero lag cross-correlation between source and receiver wavefields is the commonly used imaging condition. However, this imaging condition produces lowspatial-frequency noise or artifacts, due to the strong contrasts in velocity field (Pestana et al., 2014). Several imaging techniques have been proposed to reduce the artifacts occurrence. Derivative operators as Laplacian are the most frequently used. In this work, we propose the usage of a technique based on a spiral phase filter ranging from 0 to 2π, and a toroidal amplitude bandpass filter, known as Laguerre-Gauss transform. Through numerical experiments we present the application of this particular filter on Seg Eage salt model and Sigsbee 2A model. We also present evidences that this method improves RTM images by reducing the artifacts and notably enhance the reflective events.
On the relation between big data and machine learning
Author: Leandro Fabio Ariza Jiménez.
Recently the term “Big Data” was coined to capture the meaning of a data-explosion trend from diverse sources and domains, which society has been exposed due to technological advances since the second half of the 20th century. Machine Learning (ML) is a sophisticated analytical technology that can be used to provide us with intelligent analysis of Big Data. In this talk the relation between Big Data and ML is discussed, and two approaches of ML are presented: Clustering, and Deep Learning. The fundamental concepts of, and problems with Clustering will be discussed, followed by a description of some traditional algorithms and the presentation of experimental results in artificial datasets. With regard to Deep Learning, an introduction is given and future work in relation with big data is described.
Transfer learning on an Autoencoder-based Deep Network
Author: Leandro Fabio Ariza Jiménez.
It is widely known that deep neural networks can be difficult to train in practice, since in order to obtain state-of-the-art results we need a great amount of data and computing power. However, we can overcome this issue either using autoencoders as way to “pre-train” deep neural networks or following a “transfer learning” approach. In particular, here we carried out several experiments to study how both approaches can benefit the training of deep networks.
Riemannian wave-field extrapolation: thesis proposal
Autor: Hector Roman Quiceno E.
The earth is at least a visco elastic medium, in which absorption losses give rise to attenuation and dispersion effects. The elastic wave equation is framed in terms of tensor operators acting on vector quantities. It is also true that a proper treatment of anisotropy fundamentally demands an elastic viewpoint, even when only P-waves (quasi-P waves) are contemplated. Different representations for the same physical law can lead to different computational techniques in solving the same problem, which can produce different and new numerical results, so this new but accurate representation should lead us to new results and descriptions of the phenomena.