t Special section: Seismic inversion

Porosity inversion by Caianiello neural networks

with Levenberg-Marquardt optimization
Cyril D. Boateng1, Li-Yun Fu2, Wu Yu2, and Guan Xizhu3
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Abstract
Caianiello neural networks (CNNs) incorporated with the Robinson seismic convolutional model are modified
by the Levenberg-Marquardt algorithm to improve convergence. CNNs are extended to the multiattribute domain
for reservoir property inversion, with time-varying signal processing by a frequency-domain block implementation
using fast Fourier transforms. Optimal inversion can be achieved by applying the Levenberg-Marquardt optimi-
zation to multiattribute domain CNNs for convergency improvement due to its ability to swing between the steep-
est-descent and Gauss-Newton algorithms. The methodology is applied to porosity estimation in an oilfield with
six wells in the Bohai Basin of China. Cross-validation results indicate significant correlation between actual
porosity logs and predicted porosity logs. Compared with a traditional method, our technique is robust.

Introduction Lines and Treitel, 1984; Schultz et al., 1994). The former
Optimal seismic inversion for inferring petrophysical requires a deterministic function for relating acoustic
property distribution during reservoir characterization parameters and seismic responses, whereas the latter
involves integrating seismic data and well-log data. Seis- requires a statistical relation between the two data
mic data have good lateral coverage at the basin scale, sources. Examples of deterministic methods are single-
but they are constrained at both ends of the frequency trace recursion, model-based, and traveltime inversion
band, and this limits seismic methodology when de- (Russell, 1988). Statistical methods have the strength of
tecting thin beds, for example. On the other hand, even being data driven, the ability to reconstruct missing
though well logs are at least an order of magnitude bet- frequency components and extending beyond well loca-
ter in terms of vertical resolution when compared with tions (Schultz et al., 1994; Russell et al., 1997; Herrera
seismic data, spatial sampling is usually limited. As et al., 2006).
outlined by Fu (2003), additional challenges that may The mechanisms that create sedimentary basins and
be encountered are inexact data (incomplete informa- reservoirs can be deterministic and statistical. It is ad-
tion, overlapping information, and noise contamina- vantageous to combine the two methods for better in-
tion), leading to ambiguous physical relationships. To version results. Some important recent examples of
tackle the problem of integration, geophysicists have seismic inversion are deterministic (Leite and Vidal,
focused on the application of data-driven methods, e.g., 2011; Kumar et al., 2016), statistical (Calderon and Cas-
artificial neural networks. The problem can be formu- tagna, 2007; Khoshdel and Riahi, 2011; Naeem et al.,
lated as an inverse problem (Tarantola, 1984, 2004; 2015), and a combination of both (Fu, 2004). One stat-
Menke, 2012), in which parameters characterizing the istical tool that has been very successful in many geo-
subsurface are determined from observed data. These physical applications is artificial neural networks (van
inferred model parameters should be meaningful to de- der Baan and Jutten, 2000; Poulton, 2002), and when the
scribe the properties of the subsurface; e.g., assuming appropriate network is deployed, it has the ability to al-
acoustic-wave propagation in the earth, a model param- low deterministic components to be incorporated into
eter will be P-velocity (Sheriff, 2002). the activation functions acting as neurons. Designing
The deterministic and statistical methods have been an appropriate network involves choosing the right
used in inversion applications, and each has its respec- number of layers, choosing the activation unit (the ba-
tive strengths and limitations (Oldenburg et al., 1983; sic unit of the network) that can find the nonlinear

1
Institute of Geology and Geophysics, Key Laboratory of Petroleum Resource Research, Beijing, China and University of Chinese Academy of
Sciences, Beijing, China. E-mail: cyrilboat@ mail.iggcas.ac.cn.
2
Institute of Geology and Geophysics, Key Laboratory of Petroleum Resource Research, Beijing, China. E-mail: lfu@mail.iggcas.ac.cn;
wuyu19880216@163.com.
3
Formerly Institute of Geology and Geophysics, Key Laboratory of Petroleum Resource Research, Beijing, China; presently CNOOC, Beijing,
China. E-mail: gxz620@163.com.
Manuscript received by the Editor 11 July 2016; revised manuscript received 20 January 2017; published online 17 May 2017. This paper appears
in Interpretation, Vol. 5, No. 3 (August 2017); p. SL33SL42, 11 FIGS., 1 TABLE.
http://dx.doi.org/10.1190/INT-2016-0119.1. 2015 Society of Exploration Geophysicists and American Association of Petroleum Geologists. All rights reserved.

Interpretation / August 2017 SL33

 relationship, and choosing the most efficient optimiza- ance, but to other seismic attributes, the methodology is
tion technique. extended into the multiattribute domain. Convergence
Most artificial neural network applications use the is also improved by applying the Levenberg-Marquardt
McCulloch-Pitts model (McCulloch and Pitts, 1943). optimization instead of the steepest-descent optimiza-
The McCulloch-Pitts model is a very simple logic unit tion. Finally, the methodology is applied to a reservoir
designed to mimic the nerve net of the brain and pro- in the Bohai Basin of China to attempt porosity inver-
duce complex patterns by accepting multiple input and sion with only six wells.
providing binary output. An alternative to this general
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

form is the Caianiello (1961) model, with the following

advantages: the capability for temporal pattern process- Methodology
ing; fast operation with Fourier transforms; and an CNNs
avenue to integrate deterministic, physically meaning- Conventional neural networks are based on the fol-
ful models into the statistical network (Fu, 2003). In Fu lowing McCulloch-Pitts neuron model (McCulloch and
(2002), the Caianiello neural network (CNN) is applied Pitts, 1943):
in reservoir characterization for reservoir property in-
X
N 
version by incorporating deterministic components into
oj f wji oi j ; (1)
the neural network and integrating seismic data, well i1
logs, and geologic data. This methodology was limited
to porosity inversion from impedance. Porosity may, where oi is the input, oj is the output, wji is the connec-
however, not be sensitive to impedance in a particular tion weight, and j is the threshold. These networks
reservoir, but it may respond to other seismic attributes have been widely used in exploration geophysics (van
that are components of the seismic data. der Baan and Jutten, 2000; Poulton, 2002) as a universal
To overcome this limitation, this paper will (1) shift approximator to map an input to any desired output.
the application of the CNN into the multiattribute do- However, the networks may not be appropriate to proc-
main, thereby taking advantage of the neural networks ess time-varying signals because the McCulloch-Pitts
ability to accept multiple inputs and (2) improve con- neuron model with dot-product integration of weights
vergency using the Levenberg-Marquardt optimization, and inputs outputs a single value even when the input
thereby avoiding the disadvantages of the steepest- is a vector. Therefore, the frequency and phase informa-
descent back-propagation optimization. Steepest descent tion of a time signal may be lost.
is a first-order optimization technique that is synonymous In contrast, the CNNs are based on the following
with back-propagation. This technique uses the deriva- Caianiello (1961) neuron equation:
tive of the error (which is a function of the weights).
The major disadvantage of steepest descent (also known
X
N Z
t 
as gradient descent) is that one has to specify a learning oj t f wji oi t d j t ; (2)
rate and a wrong choice may impact the way the neural i1
0
network works. Therefore, it requires a lot of trial and
error to get the right learning rate. where the neurons input, output, threshold, and activa-
Multiattribute transforms for estimating reservoir tion function are represented by oi t, oj t, j t, and
properties are first applied by Schultz et al. (1994) and f :, respectively, and wji t is the time-varying connec-
have since been accepted as an effective inversion tion weight (neural wavelet).
method with applications in different basins (Russell Fu (2003) outlines the details of how inversions in
et al., 1997; Liu and Liu, 1998; Hampson et al., 2001; Pra- the CNN are executed. The computation in the neuron
manik et al., 2004; Herrera et al., 2006; Ahmed et al., model involves dot-product integration of spatial inputs
2010). The Levenberg-Marquardt optimization swings and a convolutional integration of temporal inputs. The
between the steepest-descent and the Gauss-Newton process is as follows: (1) dot-product integration of spa-
algorithms, and it is implemented with back-propaga- tial inputs but convolutional integration of temporal in-
tion to minimize the errors during learning. Several puts, (2) the neural network crosscorrelating with the
studies have shown the effectiveness of the Leven- inputs from other neurons, (3) significant correlation
berg-Marquardt algorithm as an optimization technique coefficients denoting a good match between the input
in other geophysical applications (Lines and Treitel, and the neurons filtering property, and (4) updating of
1984; Zhao et al., 1994; Boadu, 1997; Arenas et al., the neural wavelet. The advantage of this process is its
2001; Pujol, 2007; de Lautour et al., 2009). learning power due to adaptive changes of the neural
The CNN incorporated with the Robinson (1967) wavelet, whereas its spatial summation offers an ability
seismic convolutional model has been modified in this to consider several multiattribute traces simultane-
paper for porosity inversion. The network has been suc- ously, and each parameter becomes a time sequence in-
cessfully applied to impedance inversion (Fu, 2004) and stead of a constant value. Figure 1 shows a schematic of
joint lithologic inversion (Fu, 2002). To overcome the a simple three-layer CNN. In neural networks, the con-
limitation of reservoirs in which petrophysical proper- volutional model is embedded as the forward model.
ties of interest may not necessarily respond to imped- Even though the CNN updates just as the traditional

SL34 Interpretation / August 2017

 back-propagation, its activation functions will find non- wl1 wl J Tl J l I1 J l el : (4)
linear relationships between the data that are peculiar
to the embedded deterministic model of a specific res- We will use the Levenberg-Marquardt algorithm (equa-
ervoir. The cost function for this problem is defined as tion 4) as an update equation for neural wavelets to im-
the following mean-square error performance function: prove the back-propagation learning technique in CNNs.
1XX 2 1XX
E ek t dk t ok t2 ; (3) Joint deterministic and data-driven inversion
2 k t 2 k t
An essential component of the joint-inversion
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

where dk t is the desired output and ok t is the actual scheme (Fu, 2002) is the incorporation of a determinis-
output from the output layer of the network. The appli- tic model. As stated earlier, the major challenges with
cation of the back-propagation learning algorithm to re- seismic inversion for reservoir characterization include
duce the cost function leads to an updated equation for instability, nonuniqueness, and uncertainty, which arise
neural wavelets. In contrast to the forward propagation from inexact measured data and ambiguous physical re-
that is characterized by temporal convolution opera- lationships. For example, deterministic inversions suf-
tions, the back propagation for errors and the update fer from uncertainty and their inability to hold for
equation for neural wavelets are carried out by cross- different environments, whereas statistical inversions
correlation operations. This allows a block frequency- have the potential to generate spurious results, which
domain implementation with fast Fourier transforms, im- may be physically untenable due to a small number of
proving the effectiveness for geophysical applications sampling locations and the low signal-to-noise ratio. A
(from Fu, 2004). For mathematical details of CNNs, refer combination of the two methods should be optimal for
to Fu (2002, 2003, 2004). In this paper, we extend CNNs enhanced inversion results. Significant advances have
to the multiattribute domain for a many-to-one mapping, been made in addressing the relationship between the
which may be more meaningful and sensitive to porosity physical properties of rocks and geophysical observa-
(Schultz et al., 1994; Russell et al., 1997; Hampson tions (Mavko et al., 2009). In deterministic models, a
et al., 2001). forward relation is assumed for the reservoir property
inverse problem (Gelfand and Larner, 1984; Lines and
Levenberg-Marquardt optimization Treitel, 1984; Martinez, 1985; He and Reynolds, 1995).
The popular back-propagation technique (Rumelhart Finding the relationship between compressional
et al., 1986) is actually characteristic of a least-squares velocity and porosity has been the topic of much re-
algorithm, the simplest version of all steepest-descent search, and examples of some meaningful relationships
optimization methods. Convergence and nonunique- between reservoir acoustic properties and porosity are:
ness problems are commonly inherent in the algorithm (1) Wyllies time-average equation (1p 1 m 
in many geophysical applications (Lines and Treitel, f , where vm is the P-wave velocity of the rock ma-
1984), which requires more exact treatments in the ini- trix and vf is the velocity of the pore fluid), (2) Gardners
tial parameters. Particularly, an adaptive change of equation ( 0.23V 0.25 ), and (3) Raymers equation
learning rates becomes very important to assure algo- (p 1 2 m f ), which was proposed as a mod-
rithm convergence, but it is often ineffective for seismic ification of the time-average equation by suggesting dif-
inversions involved with vast amounts of data. Recent ferent laws for different porosity ranges (Wyllie et al.,
implementations with other nonlinear optimization tech- 1958; Gardner et al., 1974; Raymer et al., 1980; Nur et al.,
niques demonstrate faster convergence and more stable 1998). Figure 3a shows a schematic diagram of some of
operations. In this paper, we replace the least-squares these well-known empirical relations. Variations in res-
algorithm by more advanced optimization methods ervoir acoustic properties are a result of different geo-
to improve the back-propagation learning technique in logic variables, such as lithology, porosity, clay content,
CNNs. The Levenberg-Marquardt algorithm was devel- fluid saturation, pore pressure, temperature, etc., and
oped by Levenberg (1944) and Marquardt (1963), and it these may change depending on the reservoir being
is a second-order optimization technique. It is a combi- studied. Due to the variation and the random nature of
nation of the steepest-descent and the Gauss-Newton the subsurface, it is often the case that these relation-
methods, using the stability of the former and the speed ships are modified to suit a particular basin. It is chal-
advantage of the latter. The derivation is done by starting
from the steepest-descent algorithm through to the New-
ton method, the Gauss-Newton method, and finally the
Levenberg-Marquardt algorithm (Allred and Kelly, 1990;
Hagan and Menhaj, 1994; Hagan et al., 1996; Poulton,
2001; Yu and Wilamowski, 2011; Sandham and Leggett,
2013). Figure 2 shows a schematic diagram of how the
steepest-descent optimization and Levenberg-Marquardt
optimization techniques work to update a back-propaga-
tion neural network. The update rule for the Levenberg-
Marquardt algorithm is Figure 1. Three-layer CNN.

Interpretation / August 2017 SL35

 lenging to derive a single global empirical relationship We then define the architecture of the CNN using a
between compressional wave velocity and porosity for three-layer network (see Figure 1). The CNN is de-
siliciclastic sediments. This is especially due to the sen- ployed to estimate the nonlinear factor in equation 5.
sitivity of velocity to other factors such as consolidation Artificial neural networks are essentially statistical
history and clay content (Erickson and Jarrad, 1998). tools that find the nonlinear relationship between a tar-
An empirically derived, relatively flexible expression re- get data set and a predictor data set by mimicking the
lating velocities and porosity in vertical time mode to fit way the brain works. Neural networks are an effective
well-log data for continental sandstone sediments in tool for deriving nonlinear relationships because they are
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

eastern China, which was applied in this case, is shown able to capture patterns in complex multidimensional
in equation 5 systems and follow the cloud of data points closely. They

 do this by using mathematical transforms as artificial
m tm t 2t V p t V f t neurons; modeled after the neuron structure of mammals
t ln ; (5)
tm t t V m t V p t although at a much smaller scale and have the ability to
accept multidimensional input in the form
where is a nonlinear factor, t is the porosity curve in
vertical time, V p is the P-wave velocity curve, m t is the Px; y; z f A1 x; y; z; : : : ; Am x; y; z; (6)
maximum sandstone porosity, V m t is the rock-matrix
velocity, and V f t is the pore-fluid velocity. Even though where P is the reservoir property of interest, A is the seis-
a neural network can map multiinput to multioutput re- mic attribute, m is the number of seismic attributes, and
lationships, it will be quite difficult to find a deterministic f is the nonlinear function. In Fu (2004), impedance data
model that can take into account all the variables that were used as input. Although the inversion was success-
influence reservoir acoustic properties and predict more ful, it may not work efficiently in areas where the reser-
than one simultaneously. To go around this problem, we voir property does not respond to a single seismic
limit the rock parameter to be estimated to only porosity. attribute. To enhance the methodology, we extend it into
The deterministic model in equation 5 can include vari- the multiattribute domain by using attributes of the seis-
ous lithologies, and it is applied as the ac-
tivation function in the neurons because
the nonlinear factor can adjust the func-
tional form of the equation to an appro-
priate shape based on empirical data
points. Figure 3b is a schematic showing
how different nonlinear factors may af-
fect the shape of the velocity-porosity re-
lationship. This model is very useful in
circumstances in which the field data
do not closely follow a particular rock-
physics transform for relating porosity
and P-velocity, which is often the case in Figure 3. Schematic diagrams of rock-physics relationships: (a) well-known
field applications. rock-physics relationships and (b) flexible nonlinear rock-physics model.

Figure 2. A schematic diagram of back-propagation network with (a) steepest-descent optimization and (b) Levenberg-Marquardt
optimization.

SL36 Interpretation / August 2017

 mic data as input and taking advantage of the neural and 1300 ms. Reservoir rocks from the Guantao member
networks ability to accept multidimensional input. are fluvial sandstones comprising individual elongate
Attributes of poststack seismic data are defined as mea- bodies (Ryder et al., 2012). The predominant porosity
surements derived from seismic data (Sheriff, 2002). type is intergranular porosity. The lithology of the mem-
Seismic data are a convolution of a source wavelet and ber comprises of conglomerates, sandstones, and mud-
the earths reflectivity series at a particular location. The stones (Yang and Xu, 2004; Tian et al., 2014).
impedance contrast between reservoir properties has Figure 4 shows the areal distribution of wells used for
varied effect on the data collected. Seismic attributes al- the inversion. Two wells, W3 and W1 are used for vali-
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

low us to mathematically transform seismic data into the dating the inversion. Wells W1 and W6 are at the extreme
component parts (e.g., amplitude, frequency, and phase) ends of the area of interest. Figure 5 shows the workflow
and find the specific component of the seismic data that for the inversion described in this paper.
correlates with a particular reservoir property. Seismic
attributes were computed from seismic
data ensuring that each had a physical re-
lationship with the reservoir property of
interest. In addition, seismic data itself
and nonlinear transforms, such as expo-
nent, square root, inverse, and natural
log, were applied to the derived attrib-
utes. Due to the incorporation of the
deterministic component, the final condi-
tion for the input attributes to satisfy is
to honor the model in the neurons. This
enhances uncertainty reduction by inte-
grating prior knowledge into the inver-
sion. Once the input and output data
sets are defined, we apply a feature
selection technique (Draper and Smith,
1966) to reduce the feature space to find
attributes that actually correlate with the
target data and to avoid redundancy Figure 4. Timeslice 1250 ms showing distribution of wells in the Shengli oilfield.
(Barnes, 2007).
Finally, to ensure that the inversion is reliable, we
follow Qin and Fu (2013) methodology of cross-valida-
tion supplemented with seismic-to-well correlation. The
methodology allows prior seismic-to-well analyses to
be used to test seismic inversions before the implemen-
tation of actual inversion. This entails ensuring an exist-
ing relationship between seismic velocity and porosity
at log scale and tackles the problem of measuring the
fidelity of seismic data, and it also provides a constraint
on the risk and uncertainty in reservoirs with strong
geologic heterogeneities. It is important to note that the
relationship can be either linear or nonlinear.

Application
The methodology was applied to a reservoir from the
Bohai Basin in eastern China. Poststack seismic data and
six wells with porosity logs were available for analysis.
The seismic data cover an area of 144.11km2 . The sedi-
ments in the area consist of the Kongdian (Ek), Shahejie
(Es), Dongying (Ed), Guantao (Ng), Minghuazhen (Nm),
and Pingyuan (Qp) Formations. The reservoir of interest
is the Guantao Formation with traps that are mostly
lithostratigraphic and were deposited in a fluvial environ-
ment (Yang and Xu, 2004; Tian et al., 2014). The Guantao
member is of Miocene age (1020 Ma). Its thickness
varies between 200 and 900 m, and the data for neural Figure 5. Workflow of CNN with Levenberg-Marquardt opti-
network porosity inversion are taken between 1190 mization.

Interpretation / August 2017 SL37

 The CNN architecture (with three layers) is defined gree of scatter in the velocity-porosity plot (Figure 6a).
with the geophysically meaningful physical model in- This may be due to the effect of higher order factors,
cluded in the neuron equation (Figure 1). As discussed such as clay content, composition, and diagenesis. To
previously, the fidelity of the data has to be tested as a investigate this further, Figure 6b shows the velocity-
constraint on risk and uncertainty. This is done by en- porosity plot color coded by gamma-ray values. We can
suring that the deterministic model embedded in the observe several significant trends from this figure. First,
neurons of the CNN is validated with data at this par- rocks with porosities between 0% and 5% show the high-
ticular reservoir. est scattering and present challenges when predicting
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

To transform amplitude-based attributes to lithologic from seismic data. These rocks generally exhibit mid-
properties, such as porosity, we need to establish a link to high-range gamma-ray values (associated with a high
to acoustic-wave properties, such as velocity. Using clay content). The high clay content is probably the rea-
data from the six wells across the study area, acoustic son for the high level of scattering in the plot. Second,
velocity is plotted versus porosity. There is a high de- there is another major trend for rocks with porosities
from 15% to 38% and relatively low gamma-ray values.
We also observe a good correlation between porosity
and gamma-ray values; that is, generally, high gamma-
ray values correspond to low porosities. Ideally, rock-
physics relations provide the basis for relating seismic
attributes to petrophysical properties but in cases in
which linear relations are difficult to derive in the face
of complex field data, statistical relations such as neural
networks can be used to overcome this challenge. In

Figure 6. (a) Crossplot of P-velocity and porosity for all

available wells. (b) Crossplot of P-velocity and porosity color
coded with the gamma ray for all available wells.

Table 1. First three seismic attributes with the best

correlations to porosity.

Seismic attribute Error

Derivative of amplitude 11.40

Envelope weighted phase 11.27 Figure 7. (a) An example seismic section corresponding to
Root-mean-square energy 10.50 the validating well W1. (b) Porosity section corresponding
to the validating well W1.

SL38 Interpretation / August 2017

 view of this, the flexible expression in equation 5 with sampled in the time domain (between 1190 and 1300 ms)
its nonlinear factor , which adjusts to suit the empirical with the same sample rate (2 ms) as the seismic data.
data depending on lithologic characteristics and strati- The sample window covers the Guantao member (NgS4
graphic constraints, is apt for this data set. The activa- NgS5), the zone of interest. The network learns by using
tion function is used based on this validation, and the the back-propagation algorithm with the Levenberg-Mar-
neural network is then deployed to derive the exact quardt algorithm for error minimization (Figure 2b).
nonlinear relationship.
Input data, seismic attributes, are generated from 3D
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

poststack seismic data. To avoid computing in high di-

mensions, the input space is reduced by looking for only
the relevant attributes with significant relationships, with
the reservoir property of interest. In addition, the con-
straint of ensuring that the seismic attributes generated
have physically meaningful relationships with porosity is
maintained. Attribute selection is achieved with stepwise
regression (see Hampson et al., 2001). The seismic attrib-
utes with the best correlations to porosity are shown in
Table 1. Attributes with the least error correspond to the
best correlation.
To invert for porosity, we define the CNN architecture
with the selected seismic attributes as input and the
porosity curves as output. The porosity curves are

Figure 9. The validation results comparing actual porosity

and predicted porosity within the Guantao member (shown
in black) W1 correlation coefficient 82% and W3 correlation
coefficient 80%. The original porosity is in blue, and the pre-
dicted porosity is in red.

Figure 8. (a) An example seismic section corresponding to

the validating well W3. (b) Porosity section corresponding
to the validating well W3. Figure 10. Validation results with error bars in red.

Interpretation / August 2017 SL39

 Finally, the learned neural network is applied to porosity the results are shown in Figure 11. From visual com-
inversion using the seismic attributes as input. Figures 7 parison, our methodology shows much better results
and 8 show seismic sections from field data and the re- in terms of resolution.
sults from the porosity volume.
Comparing inverted sections and seismic sections
shows that the resolution vertically and laterally has Conclusion
been increased significantly. In addition, porosity curves Combining deterministic and statistical methods for
showed excellent correlation with the porosity volume porosity is an optimal approach for mapping the distri-
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

when cross-validation was applied (Figure 9). This indi- bution of subsurface properties for reservoir characteri-
cates the ability of the method to estimate porosity zation. The CNN is an effective tool to merge these two
beyond the well locations used for training. Figure 10 methods with an added advantage of being able to out-
shows the porosity logs with error bars to give an idea put results as vectors. Extending the methodology into
of the uncertainty in the prediction. The horizontal width the multiattribute domain improves the inversion re-
of the error bars signifies the uncertainty bounds of a sults and increases the resolution by making added
particular predicted porosity value. use of the responses in the seismic data to specific res-
Significantly, inversion results will help in character- ervoir properties. Furthermore, to avoid the many prob-
izing the reservoir as the porosity distribution across lems of the steepest-descent algorithm, a more robust
the reservoir has improved. Improved resolution verti- Levenberg-Marquardt algorithm is deployed in this
cally and laterally will be beneficial in identifying the study to minimize the errors and improve convergence.
zones of interest. To test the robustness of the method- A multiattribute domain extension of the CNN with
ology outlined in this paper, we apply a commercially the Levenberg-Marquardt technique, an improvement
available software to the same inversion problem and over the steepest-descent algorithm, was applied to
seismic data from the Bohai Basin of eastern China.
This methodology successfully inverted a seismic vol-
ume into a porosity volume in an oilfield with six wells
cutting across the zone of interest. Cross-validation re-
sults show excellent correlation, and it was achieved
between the porosity curves at well locations and the
porosity volume. Our formulation is an extension of
current methods and tools, and it can bring significant
improvement to reservoir property inversion methodol-
ogy especially in heterogeneous reservoirs with few
wells available for inversion.

Acknowledgments
We are grateful to the Chinese Academy of Sciences
and The World Academy of Sciences for sponsoring this
research through the CAS-TWAS Presidents Fellowship
program. Further appreciation goes to H. Bui (the assis-
tant editor), F. Forouhideh, and two anonymous re-
viewers whose suggestions improved this paper. We also
appreciate dGB Earth Sciences for donating Opendtect
seismic interpretation software, which was very useful
for visualization and seismic attribute generation.

References
Ahmed, O., R. Abdel-Aal, and H. AlMustafa, 2010, Reservoir
property prediction using abductive networks: Geo-
physics, 75, no. 1, P1P9, doi: 10.1190/1.3298443.
Allred, L. G., and G. E. Kelly, 1990, Supervised learning
techniques for back-propagation networks: IJCNN
International Joint Conference on Neural Networks,
1, 721728, doi: 10.1109/IJCNN.1990.137654.
Arenas, E., C. van Kruijsdijk, and T. Oldenziel, 2001, Semi-
automatic history matching using the pilot point method
including time-lapse seismic data: Annual Technical
Figure 11. Porosity inversion from a popular commercial Conference and Exhibition, SPE, doi: 10.2118/71634-
software. MS.

SL40 Interpretation / August 2017

 Barnes, A., 2007, Redundant and useless seismic attributes: offshore oil field: A case study: Journal of Petroleum
Geophysics, 72, no. 3, P33P38, doi: 10.1190/1.2716717. Science and Engineering, 78, 740747, doi: 10.1016/j
Boadu, F. K., 1997, Rock properties and seismic attenua- .petrol.2011.08.016.
tion neural network analysis: Pure and Applied Geo- Kumar, R., B. Das, R. Chatterjee, and K. Sain, 2016, A meth-
physics, 149, 507524, doi: 10.1007/s000240050038. odology of porosity estimation from inversion of post-
Caianiello, E. R., 1961, Outline of a theory of thought-proc- stack seismic data: Journal of Natural Gas Science and
esses and thinking machines: Journal of Theoretical Bi- Engineering, 28, 356364, doi: 10.1016/j.jngse.2015.12
ology, 1, 204235, doi: 10.1016/0022-5193(61)90046-7. .028.
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Calderon, J., and J. Castagna, 2007, Porosity and lithologic Leite, E. P., and A. C. Vidal, 2011, 3D porosity prediction
estimation using rock physics and multi-attribute trans- from seismic inversion and neural networks: Com-
forms in Balcon Field, Colombia: The Leading Edge, 26, puters & Geosciences, 37, 11741180, doi: 10.1016/j
142150, doi: 10.1190/1.2542439. .cageo.2010.08.001.
de Lautour, O. R., and P. Omenzetter, 2009, Prediction of Levenberg, K., 1944, A method for the solution of certain
seismic-induced structural damage using artificial neu- non-linear problems in least squares: Quarterly of Applied
ral networks: Engineering Structures, 31, 600606, doi: Mathematics, 2, 164168, doi: 10.1090/qam/1944-02-02.
10.1016/j.engstruct.2008.11.010. Lines, L. R., and S. Treitel, 1984, A review of least-squares
Draper, N. R., and H. Smith, 1966, Applied regression inversion and its application to geophysical problems:
analysis: Wiley & Sons. Geophysical Prospecting, 32, 159186, doi: 10.1111/j
Erickson, S. N., and R. D. Jarrad, 1998, Velocity-porosity .1365-2478.1984.tb00726.x.
relationships for water-saturated siliciclastic sediments: Liu, Z., and J. Liu, 1998, Seismic-controlled nonlinear
Journal of Geophysical Research, 103, 3038530406, extrapolation of well parameters using neural networks:
doi: 10.1029/98JB02128. Geophysics, 63, 20352041, doi: 10.1190/1.1444496.
Fu, L., 2002, Joint lithologic inversion, in P. Wong, F. Amin- Marquardt, D., 1963, An algorithm for least-squares estima-
zadeh, and M. Nikravesh, eds., Soft computing for reser- tion of nonlinear parameters: Journal of the Society for
voir characterization and modeling: Studies in Fuzziness Industrial and Applied Mathematics, 11, 431441, doi:
and Soft Computing, 511530. 10.1137/0111030.
Fu, L., 2003, An information integrated approach for reservoir Martinez, R., 1985, Deterministic estimation of porosity
characterization, in W. A. Sandham, and M. Leggett, eds., and formation pressure from seismic data: 55th Annual
Modern approaches in geophysics: Springer, 157178. International Meeting, SEG, Expanded Abstracts, 461
Fu, L., 2004, Joint inversion of seismic data for acoustic 464.
impedance: Geophysics, 69, 9941004, doi: 10.1190/1 Mavko, G., T. Mukerji, and J. Dvorkin, 2009, The rock phys-
.1778242. ics handbook, 2nd ed.: Cambridge University Press.
Gardner, G. H. F., L. W. Gardner, and A. Gregory, 1974, For- McCulloch, S. W., and W. Pitts, 1943, A logical calculus of
mation velocity and density: The diagnostic basics for the ideas immanent in nervous activity: The Bulletin of
stratigraphic traps: Geophysics, 39, 770780, doi: 10 Mathematical Biophysics, 5, 115133, doi: 10.1007/
.1190/1.1440465. BF02478259.
Gelfand, V., and K. Larner, 1984, Seismic lithologic model- Menke, W., 2012, Geophysical data analysis: Discrete in-
ing: The Leading Edge, 3, 3034, doi: 10.1190/1.1439033. verse theory, 3rd ed.: Elsevier.
Hagan, M. T., and M. B. Menhaj, 1994, Training feedforward Naeem, M., H. M. El-Araby, M. K. Khalil, M. K. Jafri, and F.
networks with the Marquardt algorithm: IEEE Transac- Khan, 2015, Integrated study of seismic and well data
tions on Neural Networks, 5, 989993, doi: 10.1109/72 for porosity estimation using multi-attribute transforms:
.329697. A case study of Boonsville Field, Fort Worth Basin,
Hagan, M. T., T. Martin, H. B. Demuth, and M. Beale, 1996, Texas, USA: Arabian Journal of Geosciences, 8,
Neural network design: PWS Publishing Co. 87778793, doi: 10.1007/s12517-015-1806-7.
Hampson, D., J. Schuelke, and J. Quirein, 2001, Use of multi- Nur, A., G. Mavko, J. Dvorkin, and D. Galmudi, 1998, Criti-
attribute transforms to predict log properties from seismic cal porosity: A key to relating physical properties to
data: Geophysics, 66, 220236, doi: 10.1190/1.1444899. porosity in rocks: The Leading Edge, 17, 357362, doi:
He, V., and A. Reynolds, 1995, Estimation of porosity in 10.1190/1.1437977.
thin layered reservoirs by seismic inversion: 65th An- Oldenburg, D., T. Scheuer, and S. Levy, 1983, Recovery of
nual International Meeting, SEG, Expanded Abstracts, the acoustic impedance from reflection seismograms:
10221024. Geophysics, 48, 13181337, doi: 10.1190/1.1441413.
Herrera, V., B. Russell, and A. Flores, 2006, Neural net- Poulton, M. M., 2001, Computational neural networks for
works in reservoir characterization: The Leading Edge, geophysical data processing: Elsevier.
25, 402411, doi: 10.1190/1.2193208. Poulton, M. M., 2002, Neural networks as an intelligence
Khoshdel, H., and M. A. Riahi, 2011, Multiattribute trans- amplification tool: A review of applications: Geophys-
form and neural network in porosity estimation of an ics, 67, 979993, doi: 10.1190/1.1484539.

Interpretation / August 2017 SL41

 Pramanik, A., V. Singh, R. Vig, A. Srivastava, and D. Tiwary, A data-driven interpretation methodology): The Leading
2004, Estimation of effective porosity using geostatis- Edge, 13, 305310, doi: 10.1190/1.1437020.
tics and multiattribute transforms: A case study: Geo- Sheriff, R., 2002, Encyclopedic dictionary of applied geo-
physics, 69, 352372, doi: 10.1190/1.1707054. physics: Geophysical references series: SEG.
Pujol, J., 2007, The solution of nonlinear inverse problems Tarantola, A., 1984, Inversion of seismic reflection data
and the Levenberg-Marquardt method: Geophysics, 72, in the acoustic approximation: Geophysics, 49, 1259
no. 4, W1W16, doi: 10.1190/1.2732552. 1266, doi: 10.1190/1.1441754.
Qin, N., and L. Fu, 2013, Reliability measurement of joint Tarantola, A., 2004, Inverse problem theory and methods
Downloaded 10/02/17 to 14.139.38.9. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

seismic inversion based on seismic-to-well correlation: for model parameter estimation, 1st ed.: SIAM.
Exploration Geophysics, 44, 87103, doi: 10.1071/ Tian, J., F. Hao, X. Zhou, H. Zou, and L. Lan, 2014, Charging
EG12043. of the Penglai 9-1 oil field, Bohai Bay Basin, China func-
Raymer, D., E. Hunt, and J. Gardner, 1980, An improved sonic tions of the delta on accumulating petroleum: Marine
transit time-to-porosity transform: Presented at the Trans- and Petroleum Geology, 57, 603618, doi: 10.1016/j
action of the SPWLA 21st Annual Logging Symposium. .marpetgeo.2014.07.007.
Robinson, E., 1967, Predictive decomposition of time van der Baan, M., and C. Jutten, 2000, Neural networks in
series with application to seismic exploration: Geophys- geophysical applications: Geophysics, 65, 10321047,
ics, 32, 418484, doi: 10.1190/1.1439873. doi: 10.1190/1.1444797.
Rumelhart, D. E., G. E. Hinton, and R. J. Williams, 1986, Wyllie, M. R. J., A. R. Gregory, and G. H. F. Gardner, 1958,
Learning representations by back-propagating errors: An experimental investigation of factors affecting elas-
Nature, 323, 533536, doi: 10.1038/323533a0. tic wave velocities in porous media: Geophysics, 23,
Russell, B., 1988, Introduction to seismic inversion meth- 459493, doi: 10.1190/1.1438493.
ods: Course notes series: SEG. Yang, Y., and T. Xu, 2004, Hydrocarbon habitat of the off-
Russell, B., D. Hampson, J. Schuelke, and J. Quirein, 1997, shore Bohai Basin, China: Marine and Petroleum Geol-
Multiattribute seismic analysis: The Leading Edge, 16, ogy, 21, 691708, doi: 10.1016/j.marpetgeo.2004.03.008.
14391444, doi: 10.1190/1.1437486. Yu, H., and B. Wilamowski, 2011, Levenberg-Marquardt
Ryder, R. T., J. Qiang, P. J. McCabe, V. F. Nuccio, and F. training industrial, in J. D. Irwin, ed., Levenberg-Mar-
Persits, 2012, Shahejie-Shahejie/Guantao/Wumishan and quardt training industrial electronics handbook: CRC
Carboniferous/Permian Coal-Paleozoic total petroleum Press, 121 to 1216.
systems in the Bohaiwan basin, China: US Geological Zhao, H., B. Ursin, and L. Amundsen, 1994, Frequency
Survey Scientific Investigations Report, 20115010. wavenumber elastic inversion of marine seismic data:
Sandham, W., and M. Leggett, 2013, Geophysical applica- Geophysics, 59, 18681881, doi: 10.1190/1.1443574.
tions of artificial neural networks and fuzzy logic:
Springer Science & Business Media.
Schultz, P. S., S. Ronen, M. Hattori, and C. Corbett, 1994, Biographies and photographs of the authors are not
Seismic-guided estimation of log properties (Part 1: available.

SL42 Interpretation / August 2017