You are on page 1of 12

SPE-183953-MS

Determination of True Formation Resistivity from LWD Conventional


Resistivity Measurements in a Horizontal Well

S. Fang, D. Boesing, and S. Wessling, Baker Hughes

Copyright 2017, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Middle East Oil & Gas Show and Conference held in Manama, Kingdom of Bahrain, 6-9 March 2017.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
In a horizontal well, apparent resistivity curves often show varied separations and increased values due to
the polarization horn effects near a boundary and other effects such as adjacent beds or anisotropy. The true
formation resistivity (Rt) is uncertain and can be mis-interpreted. As a consequence, saturation estimates are
uncertain so that the reserves cannot be accurately predicted. A full inversion can be used to derive the true
Rt. However, the instability of an inversion due to the lack of measurement variation along the measured
depth makes the Rt derivation very challenging.
To determine a true Rt from the logging-while-drilling (LWD) resistivity measurements, a general full
inversion is usually time-consuming and needs some inputs on formation model. This paper presents an
alternative approach to derive the true Rt without input of any prior information. The alternative approach
uses a simple two-layer model for the removal of horn effects and derives a solution at two steps: a) inverting
resistivities of the two layers at each distance away from the boundary; b) selecting a solution with specific
constraints such as using statistics in a moving window and internally consistent physical constraints to
make the solution more reliable. In the first step, neural networks are developed to calculate tool responses
and derivatives for the savings of computation time and memory needs.
Synthetic examples show that true Rt can be recovered when the relative dip between a borehole and a
layered formation is greater than 85 degrees. The examples show that bed boundaries can be reproduced with
sufficient accuracy by our approach, among others to define pay zone intervals. In addition, the resolution
of the method was studied with models of different layer thicknesses and is discussed in details in the paper.
As an outcome, the derived true Rt reads lower in a resistive thinner layer and reads slightly higher in a
conductive thinner layer. A field example from a North Sea oil field demonstrates very promising and robust
Rt results from the method.
The inversion is fully automatic and can be used in real time and downhole. Physical constraints that are
special to a horizontal well and better strategies make our algorithm robust and very fast.

Introduction
Resistivity is one formation property that is widely used in logging-while-drilling (LWD). At an early stage
of LWD, the resistivity property helps to distinguish one formation from another or determine water-filled
2 SPE-183953-MS

or oil-filled sand. A typical logging tool for such a purpose is a multiple propagation resistivity (MPR)
tool (Meyer et al. 1994). The MPR measures the attenuation and phase differences of the magnetic field,
which are converted to apparent resistivites assuming a homogenous formation. The MPR tool responses
are mostly smooth in a vertical or near-vertical well. This can be explained by the parallel or near-parallel
eddy currents to any boundary of resistivity contrasts (Doll 1949, Moran and Kunz 1962).
Currently, highly deviated and horizontal wells are often drilled for enhanced exploitation and production.
In a highly deviated or horizontal well, the MPR tool responses are no longer smooth and mostly show
polarization horns near a boundary. The horn effects are useful in identifying a boundary during a landing
section for well placement. However, these effects complicate the tool responses such as causing larger
curve separation, confusing the curve order based on the depth of investigation, and severely deviating the
responses from the true formation resistivity (Rt). The complicated responses create a barrier for making
a real-time decision for geosteering applications. This situation also increases the difficulty for formation
evaluation (FE) and further petrophysical analyses.
To remove the horn effects, a full parametric inversion is typically used with an initial model derived
from the knowledge of a user. The inversion result relies on the choice of the initial model which requires
geoscientific expertise and sufficient understanding of underlying algorithms. The geoscientist will usually
interpret a fleet of FE logs such as neutron-density, gamma ray, images to derive a petrophysically sound
initial model (Sviridov et al. 2014 and Peikert et al. 2016). The model will then enhance to robustness of
inversion algorithms towards inverting for the most appropriate Earth model which is able to explain MPR
log responses. One parameter within the Earth model is the initial resistivity distribution of formations
around the wellbore which is difficult to retrieve from MPR data affected by polarization horns. In addition,
inversion schemes may become processing-intensive so that a robust approach to remove polarization horns
to create an appropriate initial resistivity model is highly beneficial. For such a purpose, a data-driven
inversion scheme is preferrable, especially when there is a lack of knowledge about the formation of interest.
How good is a data-driven inversion scheme? Generally speaking, it depends on the variety of data
independence. In other words, independent data that has a different dependence to the formation model will
make the inversion more robust. Unfortunately, in a horizontal well the MPR tool responses are constant
along the borehole and a window selection is no longer useful. Moreover, the MPR data complexity due to a
near-boundary complicates the interpretation of formation and a simple homogeneous model does not fit. To
better derive the Rt from MPR data in a horizontal well, one practical consideration is to use a simple two-
layer model for the removal of horn effects. More layers are not encouraged due to the increased number of
unknown parameters to solve and the limited available MPR data. Other considerations include the utilizing
of statistics on inversion results in a pre-selected moving window along a horizontal or near-horizontal
borehole and the application of internal consistent physical constraints on formation resistivities of the two-
layer model. In addition to the robustness concern, another concern is the speed of the inversion, especially
for real-time applications. Our proposed approach here is to use neuron networks for calculating the MPR
tool responses and the corresponding derivatives to each model parameter.
In the following sections of the paper, we describe the detailed methodology for the data-driven inversion
scheme, and then verify the algorithm using a few synthetic examples. Further analysis on the resolution and
limitations of the algorithm are discussed through more synthetic examples. Finally, the method is applied
to a field example and the advantages and disadvantages of the method are discussed.

Methodology
In a highly deviated or horizontal well, the horn effects near a boundary create a barrier for determining
the true formation resistivity (Rt). To remove the barrier, a two-layer model with a tool in a borehole (not
shown) parallel to the boundary is assumed (Fig. 1). The borehole is not necessarily horizontal but can be
deviated at an angle, as long as it is still parallel to the boundary.
SPE-183953-MS 3

Figure 1—A two-layer model with a tool in a parallel borehole where Rup represents the
upper layer formation resistivity, Rdown represents the lower layer formation resistivity, and
Dis stands for the distance from the tool to the boundary at a reference position of 0 meter.

In reality, there may be more than one boundary near the tool. This situation could cause errors in the
determination of Rt based on the simple two-layer model. A more complicated model such as a three-layer
model can be assumed, but this increases the complexity of the inversion and could yield an unreliable
solution due to the limited available data. It is also impossible to maintain a perfect parallel borehole to a
boundary; there always exists a small angle variation along a borehole. The errors from multiple boundaries
and from the angle variation are taken as limitations and discussed in the next section.
Under the assumption of a two-layer model, there are three unknowns (Dis, Rup, and Rdown) for the
data-driven inversion. At least three independent data are thus needed to solve the problem. The MPR tool
has even eight compensated measurements including attenuations (Att) and phase differences (PD) at long
and short spacings and high and low frequencies. In the inversion, the attenuation and phase difference data
are converted to real (Re) and imaginary (Im) signal strengths using the formulas below:

Fig. 2 shows profiles of the MPR tool responses for a two-layer model with 1Ωm above (Rup) and 20Ωm
below (Rdown) the boundary along a near-horizontal borehole at the inclination of 89.5 degrees.

Figure 2—The MPR tool responses along an inclination of 89.5 degree borehole for a two-layer model of
1Ωm above and 20Ωm below. The horizontal axis is the distance to a boundary in feet (1ft = 0.3048 m). The
negative distance means the tool is in the upper layer. The first letter in the subscript of legend stands for
low frequency (L) or high frequency (H). The second letter stands for long spacing (L) or short spacing (S).

Fig. 2 shows that the responses are not monotonic. In other words, there are two distance values
corresponding to one specific response reading in certain cases. This creates a non-unique solution to
the distance if only one reading is available. The use of multiple curves could eliminate this ambiguity
4 SPE-183953-MS

if no noise exists in the data. However, the noise on real data cannot be neglected and makes certain
cases indistinguishable. For example, the noisy data at certain distances from two sides of a boundary are
apparently very close to each other within the noise level.
To solve for three unknowns, it is theoretically possible to invert for all three parameters simultaneously.
Here, a different approach is implemented. First, two formation resistivities for each distance within the
selected parameter range are inverted using a conjugate gradient method (Golub and Van Loan 1996). Then,
one solution among those with a misfit less than a certain percentage (e.g., 50%) above the best misfit is
selected. The misfit is a quantity used in inversion to determine how well the measured data match the
synthetic data. It is calculated as a relative root-mean-square error between the measured data and synthetic
data. The final solution is picked up based on certain physical constraints such as a minimum low resistivity
and a minimum resistivity contrast (high/low). The main benefits of such an approach are: a) each inversion
is more robust by solving fewer unknowns and b) explicit physical constraints can be easily applied.
Fig. 3 plots the inversion misfits for all distances within the tool's depth of investigation (DOI) when the
tool is in the 20Ωm formation near the boundary. The best misfit model is shown on the right in Fig. 3. From
this best misfit model, we obtain 1.15Ωm for the resistivity of the formation in which the tool is located,
and this is incorrect. However, the final model based on the physical constraints but with a slightly higher
misfit is shown on the left in Fig. 3. The final solution is very close to the true situation. This approach is
applied to the MPR data measured in a single depth. The resistivities obtained at multiple depths along a
horizontal borehole vary due to the random noise, even within the same formation layer. A jump to the other
layer resistivity is possible when the tool is close to a boundary. A post-process with a moving window
along the borehole is used to filter out the abnormal values. The length of the moving window is selected
to include certain amount of depth levels (e.g., 20 levels) to apply some statistics. A logarithm mean and
standard deviation are calculated with all values within the moving window and then values within the three
sigma range are kept for further analysis. A final resistivity value for the central of the window is finally
interpolated from the most popular retained values. We verify the entire algorithm and check its limitations
by applying them to synthetic examples in the next section.

Figure 3—Inversion misfit versus distance to a boundary for a case the tool is sit in 20Ωm
formation and the other resistivity is 1Ωm. The best misfit model is shown on the right and the
final model based on the minimum low resistivity and minimum contrast is show on the left.

In the inversion, we can simulate the synthetic MPR tool responses directly with the 1D code and
their derivatives with the corresponding differences from two forward simulations. Alternatively, we can
interpolate the responses from a lookup table (three parameter dimensions for eight quantities) that are
simulated offline. The direct simulation takes more CPU time and requires a forward-modeling engine while
the interpolation through a lookup table needs more memory to hold the table. In the paper we further
SPE-183953-MS 5

optimize the needs of CPU time and memory using pre-trained neural networks (Gonçalves et al., 1995;
Zhang et al., 1999). By doing so, responses and derivatives can be quickly obtained through the neural
networks with a minimum memory requirement. This is purticularly important for a real-time downhole
processing. A detailed description on the neural network approach can be found in Appendix A.

Synthetic examples
To verify the method, we apply the new method to synthetic examples. The first check is to invert the MPR
synthetic data generated from a two-layer model at three borehole inclinations of 89.5, 88, and 85 degrees.
The MPR synthetic data are generated using the 1D code and adding some random noise (10 nano-volts on
individual receivers from individual transmitters). Fig. 4 shows the inversion results and the true model. It
is clear that the results match the true model except the position shift (about 1.4m along measured depth or
0.12m in true vertical depth) of the boundary at 85 degrees. The position can be more accurately determined
by an azimuthal cross-component measurement or by a gamma ray curve.

Figure 4—Inversion results of synthetic data from a two-layer model


at three different borehole inclinations of 89.5, 88, and 85 degrees.
6 SPE-183953-MS

The second check is to apply to profiles on a three-layer model at three middle layer thicknesses of 0.3,
0.9, and 1.5 meters with the same borehole inclination of 88 degrees. Fig. 5 shows the inversion results and
the true model. Because we use a two-layer model as a basis for our method, it is understood that the results
will have some errors as long as the additional boundary is within the tool's DOI. The obtained resistivities
of the central layers and the positions of the boundaries are all shown some errors. In other words, the horn
effects have not been fully removed from the process. However, the results are still looking very promising.

Figure 5—Inversion results of synthetic data from a three-layer model at three different layer thickness (0.3m, 0.9m, and 1.5m).

The final synthetic example includes six groups of models with multiple thin layers. The layer thickness
in each group randomly varies near 0.1m, 0.9m, 0.2m, 0.8m, 0.1m, and 1m. The resistivity of each layer also
randomly varies near 1Ωm, 10Ωm, 100Ωm, and 1000Ωm. The synthetic MPR tool responses in a borehole
with an inclination of 85 degrees were generated with the same noise settings as before. Fig. 6 shows the
inversion results and the true model. For the very thin multiple layers such as in groups 1, 3, and 5, the
SPE-183953-MS 7

resulting Rt curve cannot resolve the individual layers. This is mainly due to the limited resolution of the
measurement, which is defined by several parameters such as the frequency and antenna spacing of the tool
or the resistivity contrast. The other layer groups are well resolved, proving that the method is generally
capable of detecting layer boundaries accurately. It is obvious that larger resistivity errors happen on those
more resistive layers, which can be explained by the decreased sensitivity of the MPR tool in high-resistivity
formation. It is not the purpose for the current data-driven process to obtain the high-resolution resistivity
profile. The high resolution can be achieved from the model-based inversion with constraints from other
measurements such as image data (Peikert et al. 2016).

Figure 6—Inversion results of synthetic data from multiple layers with random varied
thicknesses near 0.1m, 0.9m, 0.2m, 0.8m, 0.1m, and 1m. The resistivity of each layer randomly
varies near 1Ωm, 10Ωm, 100Ωm, and 1000Ωm. The borehole inclination is 85 degrees.

Based on the synthetic examples, the method described in the paper works in a horizontal well and is
still acceptable in a borehole with an inclination of 85 degrees. For a laminate formation, the method yields
an averaged resistivity for the formation with very thin layers (0.1m and 0.2m in the last example) but still
can resolve the boundary positions for the laminate formation with larger layer thickness (0.8m ~ 1m in
the last example).

A Field Example
Fig.7 shows the inversion results from a horizontal section of a log in a North Sea oil field. Rh and Rv
are from a model-based full inversion of memory data including MPR data and azimuth cross-component
measurement and Rt is obtained from the MPR data only using the method described in the paper. From
the comparison, the following can be observed:

• The derived Rt is typically a geometrical mean of Rh and Rv such as in the layer around X950m.
In a horizontal well the Rt is closer to Rv.
• The thin layers could be missed from the data-driven method such as those possible layers shown
between X900m and X960m. This is due to the resolution limitation from the measurement.
Without other information the layers would probably not be resolved, even in the model-based full
8 SPE-183953-MS

inversion. The model-based full inversion method uses additional information such as other FE
data (with higher resolution) to build a start model for the inversion.

Figure 7—Inversion results of field data. Rh and Rv are from a model-based full inversion of
memory data including both MPR data and azimuth cross-component measurement. In isotropic
formation Rh curve is mostly covered by the Rv curve. Rt is from the method described in the paper.

Although the results (Rt) from the 370m field log are not perfectly matched with the results (Rh and Rv)
from a model-based full inversion, they still look very promising and good enough in identifying most layer
boundaries. It is worth of mentioning that there is no prior information about the formation used in current
method. The current method is very fast. It takes only 30 seconds to process the 370m log with 4150 depth
levels on my laptop (Dell Precision 2.3GHz). For FE applications, we can use the results (Rt) as an initial
model for a further model-based full inversion with proper input constraints. This could speed up the full
inversion and provide an intuitive knowledge about the formation.

Conclusions
In a horizontal well, the horn effects that appear in the MPR tool responses can be partially or fully removed.
The method described in the paper uses a two-layer model and two-step approach for the determination of
true Rt. The two steps are the inversion of two-layer formation resistivities at each distance to the boundary
within the tool's DOI and the post processing to apply certain physical constraints on the inversion results
from the first step.
The synthetic examples show that a true Rt is possible to obtain if adjacent boundaries are separated
enough. Otherwise, the derived Rt reads lower in a resistive formation and slighly higher in a conductive
formation. The higher the formation resistivity is, the larger is the error in the obtained Rt. Multiple adjacent
layers also degrades the accuracy of the resistivity in resistive layers. However, the boundary positions are
well determined as long as they are not too close to each other and the well is not significantly deviated
from horizontal. The determined boundary position is still acceptable when the borehole is inclined 85
degrees. The layers are not resolvable if multiple layers exist with very small thickness due to the resolution
limitation of the measurement. A field example from the North Sea oil field demonstrates very promising
Rt results from the method.
SPE-183953-MS 9

The inversion is fully automatic and can be used in realtime and downhole. Physical constraints that are
special to a horizontal well, the two-step strategy, and the neural network approach for the response and
derivative calculations make our algorithm more robust and very fast.

Acknowledgments
The authors would like to thank Ms. Ulrike Peikert and Mr. David Larsen for providing the field data full
inversion results. In addition, Fang wants to thank Dr. Hans-Martin Maurer and Dr. Zhiqiang Zhou for their
useful discussions.

References
Doll, H. G., 1949. Introduction to induction logging and application to logging of wells drilled with oil based mud. Journal
of Petroleum Technology, 1, 148-162.
Golub, G. H., and Van Loan, C. F. 1996. Matrix Computations. 3rd Edition, Baltimore and London, The Johns Hopkins
University Press. ISBN 0-8018-5414-8.
Gonçalves, C. A., Harvey, P. K., and Lovell, M. A., 1995. Application Of A Multilayer Neural Network And Statistical
Techniques In Formation Characterisation. SPWLA 36th Annual Logging Symposium, 26-29 June, Paris, France.
SPWLA-1995-FF.
Meyer, W. H., Thompson, L. W., Wisler, M. M., and Wu, J. Q., 1994. A New Slimhole Multiple Propagation Resistivity
Tool. SPWLA 35th Annual Logging Symposium, 19-22 June, Tulsa, Oklahoma. SPWLA-1994-NN.
Moran, J. H. and Kunz, K. S., 1962. Basic theory of induction logging and application to study of two-coil sondes.
Geophysics, 27, 6.
Sviridov, M. V., Mosin, A., Antonov, Y., Nikitenko, M., Martakov, S., & Rabinovich, M. (2014). New Software for
Processing of LWD Extradeep Resistivity and Azimuthal Resistivity Data. SPE160257-PA, SPE Reservoir Evaluation
& Engineering, 17 (2), pp. 109-127
Peikert, U., Ritzmann, N., Larsen, D. S., Wessling S., Rommetveit, B., Mele, M., and Galli, M. T. 2016. Improved Post-
Well Reservoir Characterization Using Image-Constrained Resistivity Inversion – An Example from HA/HZ Well in
Channelized Sand. SPWLA 57th Annual Logging Symposium, 25-29 June, Reykjavik, Iceland. SPWLA-2016-R.
Zhang, L., Poulton, M., Zhang, Z. Y., Mezzatesta, A., and Chakravarthy, S., 1999. Fast Forward Modeling Simulation of
Resistivity Well Logs Using Neural Networks. 1999 SEG Annual Meeting, 31 October-5 November, Houston, Texas.
SEG-1999-0124.

Author Biographies
Sheng Fang is currently a scientific advisor with Baker Hughes in Houston. He joined Western Atlas as a
scientist in 1998 and continued on after the company merged to Baker Hughes in 2001. Sheng has worked
on the feasibility study of deep-looking multiple propagation resistivity, azimuthal propagation resistivity,
and deep-looking azimuth propagation resistivity, and deep transient resistivity tools for the last ten years.
He has also been involved in projects on data processing and data interpretation in highly deviated and
horizontal wells. He has published 40+ papers and 20+ patents. Sheng hold his BS and M S degrees in
applied geophysics from China University of Geosciences and his PhD degree in geophysics from the
University of Utah.

David Boesing is currently working as a scientist in Research and Development with Baker Hughes and
is based at the Technology Center in Celle, Germany. He joined the company in 2013 and is involved in
extra-deep azimuthal resistivity product development and high-angle/horizontal well projects. David holds
a M.Sc. in Applied Geophysics from the Technical University of Delft, Netherlands.

Stefan Wessling is a Geoscience Applications Development manager with Baker Hughes. He joined the
company in 2008 as a scientist and has been involved in different development projects, including high-
angle / horizontal log correction processing. Stefan holds a doctor degree in natural sciences after finishing
his diploma in Geophysics.
10 SPE-183953-MS

Appendix A

Neural network design for the simulations of MPR tool responses and their
derivatives
To design a neural network, we need to create training data sets for all possible model parameters. For the
two-layer model setup (Fig. 1), three parameters (Rup, Rdown, and Dis) control the model. We sweep each
parameter a certain range to cover all possible cases in reality. Table A1 shows the range of each parameter
we select. Note that the ratio Rup over Rdown is taken as a formation contrast. The distance range is limited
by the tool's DOI and the negative Dis indicates the tool is sitting in a more resistive formation. The MPR
tool responses are obtained by forward modeling in the MCWD software. (Sviridov et al. 2014)

Table A1—Selected ranges of model parameters

Parameter Range

Rdown (Ωm) 0.2 ~ 20

Rup/Rdown 1.0 ~ 200

Dis (m) -3 ~ 3

Fig. A1 shows the architecture of a two-layer neural network with three input neurons, five hidden
neurons, and one output neuron. Each input neuron represents one input parameter and each output neuron
represents one output quantity. Fig. A2 shows a general representation of the two-layer neuron network
architecture. The training process is to determine those coefficient matrices and bias vectors shown in Fig.
A2 as IW, LW, b1, and b2 using all training data sets. Fig. A3 shows the specific settings for the training.
Once the network is obtained, we can calculate the tool responses using the trained network for any given
input.

Figure A1—Architecture of a two-layer network with three input neurons, five hidden neurons, and one output neuron.
SPE-183953-MS 11

Figure A2—A general representation of the two-layer neural network architecture. p is the input vector contains
all input parameters, a is the output vactor from the hidden layer, and y is the output vector. IW is weighting
coefficient matrix for input, LW is a weighting coefficient matrix for the output of the hidden layer, and b1 and
b2 are bias vectors. tansig and purelin are network functions. y=tansig(x)=2/(1+exp(-2x))-1, y=purelin(x)=x.

Figure A3—The specific input and output settings for the network training

To overcome the well-known overfit or underfit problem, we first train the two-layer neural network with
fewer hidden neurons (say, 5) and then train the second two-layer network with more hidden neurons (say,
200) on the data residuals after the first training. The first neural network covers the main features of the
data while the second neural network covers the remaining details of the data. If needed, more neurons or
more neural networks can be used. Fig. A4 shows the flowchart for the process.
12 SPE-183953-MS

Figure A4—Flowchart of multiple neural network training process

You might also like