Introduction

Authors
Piotr W. Mirowski Formerly Schlumberger-Doll Research Now Ph.D candidate, Courant Institute, New York University, NY piotr.mirowski@computer.org Nikita Seleznev Schlumberger-Doll Research, Ridgefield, CT Nseleznev@ridgefield.oilfield.slb.com Michael Herron Schlumberger-Doll Research, Ridgefield, CT MHerron@ridgefield.oilfield.slb.com

Traditional Methods for WellWell-toto-Well Correlation

Manual correlation on gamma ray and resistivity logs
Well-to-well correlation is an inexact art form. Correlations typically use only total gamma ray and resistivity. While lines are drawn connecting wells, the reasoning process for selecting the correlation points is usually unexplained. In the best case, the shapes of the curves are used for matching, but the absolute values are commonly not considered. This figure shows an example of intrareservoir correlation based on electric logs and part of a complex study with a biostratigraphic approach in Venezuela (Rull, 1994).
Samuel D. Fluckiger, Schlumberger, Doha, Qatar sfluckiger@doha.oilfield.slb.com

David S. McCormick Schlumberger-Doll Research, Cambridge, MA DMcCormick@boston.oilfield.slb.com

Abstract
New Software For Well-To-Well Correlation Of Spectroscopy Logs
We have developed a research prototype PC application, Well2Well, which enables improved well-to-well correlation, especially when coupled with spectroscopy log data. The advantage of using spectroscopic log data is the absolute values of the chemical concentrations of the most important rock-forming chemical elements. The full list available includes elements from the Elemental Capture Spectroscopy (ECS) and Natural Gamma Ray Spectroscopy (NGS, HNGS) sondes. This is in contrast to the traditional use of gamma ray/spontaneous potential and resistivity data, which are at least as sensitive to the fluids as to the rock. Elements that are believed to be more geologically important can be included in the analysis and weighted according to the users preference, while less important trace elements such as thorium and uranium, which are major components of the gamma ray, may be excluded. The input logs can be preprocessed and smoothed with edge-preserving algorithms to make the correlation results more stable. These well-to-well depth correlations rely on the Dynamic Programming algorithm, inspired by the genematching algorithms used in biology, which computes a log signature mismatch matrix between a pair of wells and then finds the best correlation path thanks to an optimal path search through that matrix. No commercial plans are yet made for this software. However, tests on multiple datasets from North America demonstrate high-fidelity well-to-well correlations on wells separated by several hundred meters, suggesting making our research prototype software is a very promising geological characterization tool.

Automated wellwell-toto-well correlation on gamma ray logs

Spectroscopy logs
Elemental concentration logs were obtained from the Elemental Capture Spectroscopy (ECS*) sonde. The ECS* sonde uses a standard 16-Ci [59.2 1010-Bq] americium beryllium (AmBe) neutron source and a large bismuth germanate (BGO) detector to measure relative elemental yields based on neutron-induced capture gamma ray spectroscopy. The primary elements measured in both open and cased holes are for the formation elements silicon (Si), iron (Fe), calcium (Ca), sulfur (S), titanium (Ti), and gadolinium (Gd), chlorine (Cl), barium (Ba), and hydrogen (H). Wellsite processing uses the 254-channel gamma ray energy spectrum to produce dry-weight elements, lithology, and matrix properties. The first step involves spectral deconvolution of the composite gamma ray energy spectrum by using a set of elemental standards to produce relative elemental yields. The relative yields are then converted to dry-weight elemental concentration logs for the elements Si, Fe, Ca, S, Ti, and Gd using an oxides closure method. Matrix properties and quantitative dry-weight lithologies are then calculated from the dry-weight elemental fractions using empirical relationships derived from an extensive core chemistry and mineralogy database, and a realtime petrophysical analysis program. * Mark of Schlumberger The algorithm presented here has already been applied for the well-to-well correlation of Gamma Ray, Neutron Porosity, Density, Spontaneous Potential, Spherically Focused Laterolog and Induction Log Medium logs (Le Nir et al., 1998, illustrated on the figure above).

NPLC Detector Acquisition Cartridge

AmBe Source

However, this algorithm has not been used with spectroscopy logs such as the dry weight percentage channel logs of Si, Ca, Fe, S, Ti, Gd or Al elements. The figure on the right illustrates the interface of our prototype software for applied to well-to-well correlation of standard gamma ray logs. As we will see in our poster, these results can be significantly improved.

BGO Crystal And PMT

Boron Sleeve

Electronics Heat Sink

Dewar Flask

Spectroscopy sonde

New Software for Well-to-Well Correlation of Spectroscopy Logs

Piotr Mirowski, Michael Herron, Nikita Seleznev, Samuel Fluckiger, David McCormick
WellWell-toto-well correlation algorithm
To automate the process of determining correlations between pairs of wells, we employ the technique of dynamic programming (Lineman et al., 1987; Doveton, 1994; Le Nir et al., 1998), inspired by the gene-matching algorithms used in biology. This technique consists of computing a mismatch matrix between log values in a pair of wells and then finding the best connection path through that matrix using an optimal path search.

Optimal Path Search Algorithm

The Dynamic Programming algorithm is an optimal path search through the weighted mismatch matrix. It relies on a cumulative cost matrix cost that is, for a given pair (i, j), the cumulative sum of mismatches on a path going from (1, 1) to (i, j), and on a direction matrix dir.
c1 = cost ( i 1, j ) c = cost i 1, j 1 ( ) 2 i [ 2, M ] c3 = cost ( i, j 1) j [ 2, N ] cost ( i, j ) = min ( c1 , c2 , c3 ) + WM ij dir ( i, j ) = arg min ( c , c , c ) 1 2 3 1,2,3 { }

cost (1,1) = WM 11 dir (1,1) = 2 cost ( i,1) = cost ( i 1,1) + WM i1 i [ 2, M ] dir ( i,1) = 1 cost (1, j ) = cost (1, j 1) + WM 1 j j [ 2, N ] dir (1, j ) = 3

Mismatch matrix
The mismatch matrix is a two-dimensional representation of the difference between chemical signatures (i.e. log values) measured in two different wells. For each possible pair of depth values, it shows where the chemical signatures in both wells are similar (low value) or dissimilar (high value). The optimal path appears as a valley of low values (of mismatch) through the pairwise map. The mismatch can be defined as a distance between log values. For a single log channel, we compute a rectangular mismatch matrix, Mij, using differences between absolute log values l1(di) in well 1 at depth index i, and l2(dj) in well 2 at depth index j. Low and high values of Mij respectively correspond to strong and weak connections for a particular couple of samples.
M ij = l1 ( d i ) l 2 ( d j )

Good Good

Mediocre Mediocre

Bad Bad

To start the process it is necessary to assign initial column and row values of the matrix: cost(1, 1), cost(1, j) and cost(i, 1), as well as the directions dir(1, 1), dir(1, j) and dir(i, 1).

Bad Bad

A perfect well-towell correlation corresponds to a straight optimal path on the mismatch matrix, joining the Top and Bottom markers. The correlation strata are then parallel.

2170

Log curve filtering

1880

Optimal path

1680

To facilitate the connection of wells on their log values, it can be useful to smooth the log values so as to soften the apparently noisy aspect of logs and provide a better continuity on adjacent depth samples. The filters we have used in our study include the Mathematical Morphology filter (Serra, 1986, see the figure on the right), median filter, and the Gaussian filter. We tend to prefer edge-preserving filters such as the latter, because they do not smear out spikes and important changes in log curves that can correspond to geological events. These filters are local, i.e., for a given depth, the new value of the log curve is computed as a function of the log values of a small set of depth samples above and below.

Well A

ll A We

1540

Unfiltered

Filtered

1710

1840

2160

Well B

lB l e W

In this example we show the mismatch matrix computed on Si dry weight channels from two wells A and B. Columns of the matrix correspond to depths indexes of well B, and rows to those of well A. Dark values correspond to a high mismatch or distance between the log values, whereas light values correspond to a good match. On the 3D rendering of the region bounded by the green rectangle, mismatches appear as plateaus and good correlations as valleys. The optimal path is highlighted with dots.

Weighted mismatch matrix

The weighted mismatch matrix is a linear combination of mismatch matrices for the different available log channels, each log cannel given a weight between 0 and 1 that is proportional to its user-assigned relative importance. This weighted mismatch matrix defines a multi-log metric of well-to-well connection and hence enables a multi-dimensional search of optimal path in the well-to-well log distance space.
WM ij = wk l1,k (d i ) l2,k ( d j )
k

Test software prototype implemented

Spectroscopy Well-to-well Well log displays We developed a program with a user-friendly log channel selection correlation display interface. The program allows us to import well log data (including spectroscopy dry weight percentage channels), to specify for each well start and end markers for the well-to-well correlation interval, select and assign user-specified weights to log channels according to their interpreted geological importance, apply filters to log curves, visualize mismatch matrices and well-to-well correlation results, and export the results as text files. Running on Windows platform, it took a few seconds for our software to compute well-to-well correlations in a dataset of 2,500 feet of multichannel logs sampled at 0.5 foot intervals. Such a good performance makes this technique a promising tool for geoscientists. Mismatch matrix display

0.5x
Silicon

+ 1x
Calcium

+ 0.5x
Aluminum

=
Weighted Mismatch Matrix

Illustration of the concept of weighted mismatch matrix, where the mismatch matrices for Silicon and Aluminum dry weight channel logs have weights of 0.5, and the mismatch matrix for Calcium has a weight 1.

Spectroscopy logs carry complementary information

SingleSingle-log correlation

MultiMulti-log multimulti-well correlation

Perfect correlation between wells 1 and 2
Silicon Sulfur Two wells, Well_1 and Well_2, from a dataset from carbonate/evaporate dataset in North America, showing a very good correlation of sequence stratigraphy despite being distant of 100 meters. We filtered the logs before well-to-well correlation in order to provide with a better continuity on adjacent depth samples of log curves.

Very good correlation points on Si and Ca Isolated event on Si Silicon Calcium

Good overall correlation with Ca

Parallel correlation strata Isolated events Fe and S bad correlation channels Iron Sulfur

Optimal path = straight line through mismatch matrix

very good correlation between wells 3 and 4

Geological event There is a similarly good correlation between Well_3 and Well_4, from the same carbonate/ evaporate dataset from North America. A sand bed that did not exist in wells 1 and 2 appears at an approximate depth 4800ft, and is visible on the Silicon dry weight percentage channel logs.

Ti and Gd are trace elements in wells A and B

Very good correlation points on Al Isolated events

Parallel marker beds Intervals between markers are heteroheterogeneous

Silicon

Sulfur

appearance of a sand bed between wells 1 and 3

Sand bed Well_1 and Well_3 have a different sequence stratigraphy. There is a missing sand bed in well 1. It has to be noted that the sand bed appears only on the Silicon dry weight channel logs. The correlation strata above the sand bed are leveled out by the Top correlation marker. The optimal path search is indeed done only between Top and Bottom markers.

Titanium

Gadolinium

Aluminum

Gamma Ray

Poor correlation results between two siliclastic wells A and B from a dataset from North America containing shale, sandstone and thin marine limestone horizons. One single mismatch matrix, derived from one single log, was used at a time. Equally poor results were obtained with Gamma Ray logs only.

vs. MultiMulti-log correlation

Perfect wellwell-toto-well correlation when using a weighted mismatch matrix

Silicon

Sulfur

Unconformity minimized on Si Unconformity minimized on Al 0.5 x Silicon + 1 x Calcium + 0.5 x Aluminum

References
Doveton, J.H., Lateral Correlation and Interpolation of Logs, Geologic Log Analysis Using Computer Methods, AAPG Computer Applications in Geology No. 2, 1994, AAPG, Tulsa, pp.127-150. LeNir I., Van Gysel N., Rossi D., Cross-Section Construction From Automated Well Log Correlation: A Dynamic Programming Approach Using Multiple Well Logs, SPWLA 39th Annual Logging Symposium, May 26-29, 1998. Lineman D.J., Mendelson J.D., Toksos M.N., Well To Well Log Correlation Using Knowledge-Based Systems and Dynamic Depth Warping, SPWLA 28th Annual Logging Symposium, June 29-July 2, 1987. Rull, V. High-Impact Palynology in Petroleum Geology: Applications from Venezuela (Northern South America), AAPG Bulletin, v. 86, no. 2, February 2002, pp.279-300. Herron S., Herron M., Application of Nuclear Spectroscopy Logs to the Derivation of Formation Matrix Density, SPWLA 41st Annual Logging Symposium, June 4-7, 2000. Serra J., "Introduction to mathematical morphology", Computer Vision, Graphics, and Image Processing, Volume 35 , Issue 3, September 1986, pp.283-305.

Conclusions
No commercial plans are yet made for this software. However, tests on multi-channel geochemical log datasets from North America demonstrate precise, consistent well-to-well correlations on wells separated by several hundred meters, suggesting that this prototype software using multiple geochemical logs is a very promising geological characterization and interpretation tool.

Correlation results between siliclastic wells A and B obtained for a weighted set of Silicon, Calcium and Aluminum dry weight channel logs. Si and Al logs were given a weight of 0.5, whereas Ca logs were given a weight 1. Unlike the single-channel correlation, the multi-log correlation reveals a very similar sequence stratigraphy in both wells, yet horizontally distant by a distance of a few hundred meters. Contributions of Silicon logs Silicon (in green) and Sulfur (in red) dry weight percentages channel logs vehicle complementary information for the correlation of these two wells from a dataset from North America. These are carbonate/evaporite sequences and the silicon represents sand, mainly quartz, and some clay. In this environment, the sulfur is due to anhydrite, and many anhydrite beds are chronostratigraphic as well as lithostratigraphic surfaces. Silicon Sulfur Contributions of Sulfur logs

Acknowledgements
We would like to thank Rick Kear, Philippe Marza and Romain Prioul for useful suggestions on the software usability. Mohamed Aly, Jim Bristow, Alexis Carrillat, Selim Djandji, Roy Dove, Tamir ElHalawani, Ahmed Elsherif, Sherif Farag, Melissa Johansson, Lucian Johnston, Rick Lewis, Richard Netherwood, John Philips, Raghu Ramamoorthy, Frank Shray, Nneka Williams, Michael Wilson, Rachel Wood, and Meretta Qleibo helped test the software.