Seismic data processing using artificial neural networks

PhD-Candidate Thomas de Jonge

Picture of Thomas de Jonge
Thomas de Jonge

Main content

I am an industrial Ph.D. student that conducts research in cooperation with CGG Services and the University of Bergen. In CGG Oslo’s R&D group, my colleagues and I work with seismic processing and how to improve it. The main goal of seismic processing is to alter the data to suppress noise and create an accurate image of the subsurface.

Personally, I work with seismic processing and machine learning. More specifically using deep neural networks to find new ways of processing seismic data, and hopefully improve the processing results. Deep neural networks are trying to find complicated patterns and connections in the data in a similar way that the neurons in a brain do. Much like a brain, an artificial neural network needs to learn from experience. If a neural network has been trained, it will be able to perform a specific task. It is possible to train an artificial neural network to identify the behavior for different types of seismic noise and to remove it from the data.

In a marine seismic survey an array of air-guns is often used as a seismic source. An air-gun releases over-pressured air into the water. The air expands like a bubble in the water. As the bubble expands, the pressure in the bubble decreases and the expansion will eventually come to a halt and the bubble starts to collapse. As the bubble collapses, the pressure in the bubble increases and the bubble will eventually start to expand again. The bubble produced by all the air-gun will oscillate but is damped with each oscillation. Depth and air-gun volume will change the bubble behavior. In a marine seismic acquisition this differential bubble behavior can be used as an advantage such that the bubble constructively interferes for the first expansion and destructively interfere everywhere else. However, this destructively interference will still leave some energy behind, this is considered noise and is often referred to as the bubble. Figure 1 show a source signature estimated from near-field hydrophones with a sharp peak followed by bubble energy.

Illustration of source signature

Figure 1: The source signature estimated from near-field hydrophone measurements. The first sharp peak is considered to be signal and the bubble which comes afterwards is noise.

Thomas de Jonge

A big challenge in marine seismic data is to find a good estimation of the source signature. A source signature estimation is an important step in seismic processing because some of the processing steps are dependent on the source signature (de-ghosting, de-bubble, de-multiple, etc.). The problem is that the source signature can be challenging to estimate and a bad estimation will negatively affect the processing results. My goal is to test an artificial neural network’s ability to do some of these processing steps without using an estimated source signature and to understand how the network operates. I am comparing the results from an artificial neural network to modern industrial processing techniques using synthetic and real data. At the moment I am teaching a network to remove the bubble from the data using a convolutional neural network.

My main supervisor is associate professor Einar Iversen from the University of Bergen. My other supervisor is Vetle Vinje from CGG Services in Oslo. They both have high expertise and competence within the theoretical aspects of seismic wave propagation, seismic inversion and seismic signal processing. 

Illustration of neural network

Figure 2: A simple conceptual illustration of how an artificial neural network is trained to remove the bubble from common shot gathers. The input is given to the network that predicts an output. The output is compared to the truth which is used to update the network. This is repeated until the network predicts satisfactory results.

Thomas de Jonge