Calibrating an automated seismic interpretation tool from human

Document Sample
Calibrating an automated seismic interpretation tool from human Powered By Docstoc
					CALIBRATING AN AUTOMATED SEISMIC
INTERPRETATION TOOL FROM HUMAN EXPERT KNOWLEDGE

Jef Caers1 and Andre Haas2
1

Stanford Center for Reservoir Forecasting, STANFORD UNIVERSITY, Stanford, CA 94305-2220, USA. jef@pangea.stanford.edu  1 650 723 1774  1 650 725 2099
2

TotalFinaElf Pau CEDEX 64000, France
Keywords: statistical pattern recognition, geostatistics, turbidite systems, channels

This paper shows that statistical pattern recognition and a novel neural network approach can learn to mimic expert seismic data interpretations and successfully apply it to massive seismic data cubes of turbidite channels on the coast of West Africa and the North Sea. Also, it is shown how the method is integrated into geostatistical simulation methods to constrain models to seismic amplitude data.

1

Abstract
In the development phase of a reservoir, seismic amplitude data is a prime source for constraining facies geometry and heterogeneity. Seismic data in the form of post-migrated, poststack amplitude data are interpreted by expert geologist/geophysicist. Such interpretation is often tedious, time consuming and is based on certain geological and rock-physics rules emanating from years of experience. These rules are therefore not always well defined or easily translated into logical, mathematical or quantitative tools. The challenge in this work is not to understand the rules of the human expert but to design an automated device, often termed “artificial intelligence”, that can learn from the expert interpretations. Learning from expert data is an exercise in statistical pattern recognition consisting of two phases: a device learns from expert data, in this case the device should recognize the relation between seismic data and facies geometry; then, the trained device applies that knowledge in a stochastic mode, i.e. on a new seismic data set, a facies classification and the uncertainty about that classification is provided. The quantification of uncertainty is important for three reasons: first, the facies classifications are never certain, hence uncertainty needs to be quantified; secondly, human expert interpretations are prone to inconsistencies, hence a probabilistic device is needed in order to model such inconsistencies; thirdly, a probabilistic interpretation of facies from seismic allows an easy integration with information at different scales using existing geostatistical techniques. In this paper we use a novel neural network methodology to design the automated seismic interpretation tool and apply the method to actual reservoir rocks on turbidite systems in West Africa and the North Sea. To train the neural network, a training dataset needs to be established. Therefore, on a small part of a large seismic amplitude data cube, a manual interpretation of the human expert is performed, resulting in a given facies geometry. In the particular West Africa and North Sea case studies, a CAD tool (gOcad) is used to interpret channel elements from the seismic information and manually create and discretize the channels on a regular grid. To interpret the entire seismic data cube manually in this fashion is evidently a tedious and time consuming task since possibly hundreds of channels are present with the entire reservoir. The aim is to train a neural network to recognize, from the small training example, the relation between seismic training data sT(u), u  T and facies geometry iT(u), u  T1, then apply that neural network to the un-interpreted seismic data cube s(u), u  D of the reservoir to obtain an automatic facies geometry modeling, without any human interaction. To establish our neural network methodology, we need to recognize the following: the human expert is not likely to make his/her interpretation by scanning the seismic data cube pixel by pixel. The expert tries to detect salient features or patterns within the seismic. A local window of seismic data can represent such patterns: sT(u) = {sT (u), sT (u+h1), sT (u+h1),……, sT (u+hnt) }

1

In this case we deal with only two facies (channel or not), iT(u)=1 if pixel u is within channel, zero else.

2

the lag-vectors h define the geometry of the window or template. From the training datasets sT(u) and iT(u) a database of joint local patterns of seismic and facies is obtained: { sT(u) , iT(u) } = {sT (u), sT (u+h1), sT (u+h1),……, sT (u+hnt); , iT(u) }  u  T We develop a novel probabilistic neural network that can automatically model, using this database, the conditional probability of observing a channel facies at u, given a local window of the seismic data: Pr { I(u) = 1 | sT (u), sT (u+h1), sT (u+h1),……, sT (u+hnt) } A neural network is chosen instead the more conventional probability models because it takes into account the inherent non-linear physical relationship between facies and seismic amplitude data. After neural network training, we apply the neural network to the un-interpreted seismic data cube s(u) to obtain for each pixel u within D, the actual reservoir, the probability of observing channel facies. A double-blind test shows that the neural network can accurately predict the presence of channels using the seismic dataset s(u). In the double-blind test, the human expert interprets manually, based on s(u), the channels using the same technique as applied on the training data set sT(u), see the Figure 1 below.

training: seismic amplitudes
80.000 80.000

training: facies geometry
channel

non-channel

0.0 0.0 118.000

0.0 0.0 118.000

0.8000

actual reservoir: seismic amplitudes
80.000 80.000

act. resvr.: neural net cond. prob.
0.6000

0.4000

0.2000
0.0 0.0 130.000 0.0 0.0 130.000

0.0

act. resvr.: double blind human intrpt.
80.000

channel

non-channel
0.0 0.0 130.000

Figure 1: West Africa reservoir: (a) Horizontal section of the training seismic data cube s T(u), (b) same horizontal section with human interpreted channel, (c) Part of a horizontal section of the actual reservoir seismic, (d) horizontal section of the neural net automated interpretation of that section, no human interaction, (e) same section of the human interpretation performed independent from the result in (d).

3


				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:9
posted:1/1/2010
language:English
pages:3