You are on page 1of 14

SPE-197822-MS

Application of Machine Learning to Estimate Sonic Data for Seismic Well


Ties, Bongkot Field, Thailand

Nisa Sukkee, Thanapong Ketmalee, and Nattapon Jalernsuk, PTTEP; Renaud Lemaire, Shell and formerly
seconded to PTTEP; Parthasarathi Bandyopadhyay, PTTEP

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference held in Abu Dhabi, UAE, 11-14 November 2019.

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
Seismic well tie is a critical process to verify the time-depth relationship of a well. This process requires
density and sonic transit time data. However, sonic logs are usually not acquired due to cost saving,
unfavorable well path, or other operational issues. Attempts to generate synthetic logs by Gardner equation,
porosity correlation, or depth correlation did not provide the required accuracy. Therefore, the goal of our
project was to generate synthetic sonic logs using machine learning technique for seismic well ties. This
paper will compare the different methods tested, compare the results and lists the advantages of using
Machine Learning.
This approach uses machine learning technique to create synthetic sonic logs. The machine learning
model is trained to predict sonic log from other relevant logs. The model representativeness is confirmed
by blind tests, which consists of two steps. The first step compares the synthetic sonic logs to the actual
sonic logs. In the second step, four synthetic seismograms are generated from actual sonic, machine learning
synthetic sonic, Gardner predicted sonic, and averaged constant sonic. The seismic well ties are compared
between those four synthetic seismograms. Once the machine learning synthetic and actual logs show similar
results, the model is deemed good and can be applied on wells that do not have sonic logs. The synthetic
seismograms are then generated using synthetic sonic logs for all the wells that do not have actual sonic logs.
The use of synthetic sonic logs gives us the ability to
1. Generate synthetic seismogram to tie wells that do not have sonic data
2. Reduce the number sonic data acquisition, saving time and money
3. Reduce the risk of long logging string getting stuck in the hole that would requires fishing
operations and its associated cost.

Geological Background
The Bongkot field is the largest offshore gas and condensate producing field, located in the North Malay
Basin, Gulf of Thailand (Figure 1). The whole Bongkot area covers approximately 3,200 km2, water depth
ranges between 70 and 80 meters. The Oligo-Miocene sandy reservoirs were deposited in fluvial to deltaic
2 SPE-197822-MS

environments with the main reservoir contributors in FM1, 2A and 2B units (Figure 2) (Restrepo-Pace et
al., 2014; Charusiri and Pum-In, 2009). These reservoirs include bars, which are generally thin (<1 to 5 m
thick) and channels (5-20 m thick). The reservoirs are usually disconnected from each other (Leo, 1997).
More than half of them are invisible on seismic, being below seismic resolution. The hydrocarbon trapping
mechanisms are structural trap, combination trap and purely stratigraphic trap (Bin Ali et al., 2018). Due
to the highly compartmentalized and heterogeneous nature of the field, many wells need to be drilled to
drain all the individual potential traps.

Figure 1—Map location of Bongkot field in Gulf of Thailand, Tertiary Basin. The
Bongkot field (GBN & GBS) was developed since 1993 with more than 50 platforms.

Figure 2—Schematic Lithostratigraphy and Tectonic Framework of Bongkot Field


SPE-197822-MS 3

After 26 years of production, more than 600 development wells were drilled from 2 main areas, Greater
Bongkot North (GBN) and Greater Bongkot South (GBS). There are 45 platforms in GBN and 16 platforms
in GBS which are producing. Greater Bongkot North (GBN) is a hydrostatic pressure field, whereas Greater
Bongkot South (GBS) is overpressured, with high CO2 (Figure 1).
The Bongkot field is a mature gas field with its cumulative production above 5 TCF since startup in GBN
(July, 1993) and in GBS (Jan, 2012) to present day.
During the early development phases, the systematic approach used for field development was
conventional: acquiring as much log data as possible, drill the low angle wells (not exceed 60 degree), and
plan the wells into the safest areas. Nowadays, drilling complex wells with alternative methods became
a key part of the Bongkot field development. The harsh economic environment due to the oil price crisis
forced the technical team to find alternative strategies in order to insure additional field development at
lower cost to maximize the benefits (Kaeng et al., 2015). More challenging extended reach drilling wells
(ERD) are proposed. Log data acquisition has been reduced in order to save time and money and minimize
operational difficulty.
The methodology proposed in this paper can quickly and generate sonic logs where they were not acquired
due to technical or cost issues, for a fraction of the cost of the sonic log acquisition.

Methodology
Conventional Solutions
In absence of sonic log was availablity, the common methodologies to generate a pseudo-velocity commonly
used are either Gardner's equation, Using velocity model from nearby well or Using constant velocity.
Gardner's equation was published by Gardner et al (Gardner et al., 1974). This is an empirical equation
that relate bulk density and P-wave velocity as illustrate in Equation 1. The empirical coefficients can be
obtained by regression of density and sonic logs results from analogue data. This equation could provide
reasonable accurate results for brine saturated rock, however, it has limitation in low porosity rock in high
pressure condition. (Gardner et al., 1974).

Where,
RHOB is bulk density.
VP is P-wave velocity.
α, β, γ, δ are empirical coefficients.
To use the second method, a velocity curve can be approximated from wells drilled in the same structure
or close-by. The limitation of this alternative is that the thickness and other data may vary from well to well
which leads to matching uncertainty.
Lastly, the velocity can be assumed to be an arbitrary constant velocity for the whole well. This method
is simple and easy. However, the variations among different lithology or fluid are suppressed.

Machine Learning
Machine learning (ML) is a term define group of computer algorithms those perform a specific task without
explicit coding but based on data inference. There are many ML algorithms, for example, artificial neural
network, decision tree, and support vector machine. ML has board applications such as healthcare, financial,
and transportation industries. In E&P business, ML has been implemented for reservoir facies determination
4 SPE-197822-MS

(Bestagini et al., 2017), formation failure determination (Ketmalee and Bandyopadhyay, 2018), drilling
optimization (Pollock, 2018), etc.
Each specific task may favour different ML algorithm. Based on number of experiments, the most
effective algorithm for this project is Adaptive Boosting (AdaBoost). This model is a collective of other
weak ML models (learners). In this project, decision trees were used as weak learners. Each learner will be
optimized in favour for error/misclassified from preceding learner results. Although each learner is weak
but the collective of them could be a strong predictor (Freund, 1999). The AdaBoost algorithm is explained
by Figure 3 (Marsh, 2016).

Figure 3—AdaBoost Algorithms (Marsh, 2016) Each weak learner will be optimized in favor for
error/misclassified from preceding learner results. Collectively, they could be strong predictor.

The actual sonic data, and the generated sonic logs generated from these methods, constant average
sonic data, sonic data derived from Gardner's equation, and from machine learning, are plotted to check
the velocity trend and data quality. The sonic and density logs (from well data) are then used to do the
synthetic seismograms.

Seismic to well tie


Seismic to well tie is the process to generate the synthetic seismogram from sonic and density logs, then
matching that synthetic to the seismic reflection data, thus producing a relationship between the logs
(measured in depth) and the seismic (measured in travel time). Problems arise for all sorts of reasons. For
examples, the quality of the logs, checkshots, and the seismic need to be considered. There are four steps
in our workflow to generate synthetic seismograms for well tie purpose simply illustrated in Figure 4.

Figure 4—Seismic to Well tie Work Flow

The input data for synthetic seismogram generation consists of sonic log, density log, checkshot or VSP
and seismic data. The first step is to QC and edit the log data before conducting the machine learning
prediction as the input training data. Figure 5 shows the example of artifact data needed to be edited. The
log value is wrong due to casing shoe, inconsistent wireline tension, and bad hole conditions.
SPE-197822-MS 5

Figure 5—The Example of Undesirable Data That Needed to be edited.

Sonic calibration. Acoustic-logging tools give high-quality velocity data with a fine vertical resolution
in scale of decimeters. To calibrate surface seismic velocities and reflectors, the interval transit time can
be summed over entire logged interval to get the equivalent of seismic one-way time. Checkshot (velocity)
or vertical seismic profile (VSP) surveys is required to provide more accurate time-depth comparable to
the surface seismic time.
Acoustic log data are commonly calibrated using checkshot or VSP surveys prior to use in geophysical
applications. Data from these surveys, which use downhole receivers and surface acoustic sources, are
used to adjust the log data for drift and borehole conditions and result in improved time-depth correlation.
Acoustic-log data are multiplied by the density-log data, to generate an Acoustic impedance log.
Synthetic generation. From the acoustic impedance, we compute the reflection coefficients for each
reflecting interface. Reflection coefficient obtained from the equation below (Equation 2).

Where,
RC = reflection coefficient, whose values range from −1 to +1
ρ1 = density of medium 1, ρ2 = density of medium 2
 v1 = velocity of medium 1, v2 = velocity of medium 2
Typical values of RC are approximately −1 from water to air, meaning that nearly 100% of the energy is
reflected and none is transmitted; ∼0.5 from water to rock; and ∼0.2 for shale to sand.
The RC is then convolved to the wavelet to generate the synthetic seismogram. We used a simple Ricker
wavelet with a Time Varying filter matching the 3D seismic data frequency spectrum to keep things simple.
The infographic of the synthetic seismogram generation is shown in Figure 6.
6 SPE-197822-MS

Figure 6—The Schematic of Synthetic Seismogram Generation Work Flow

This operation is done four times for each blind well, one for each of the sonic log, from real data, ML,
Gardner or constant velocity.
Seismic well tie and comparison scenarios. The synthetic seismic trace is then compared with the real
seismic trace to perform well-to-seismic tie. The correlation coefficient shows how the synthetic trace
compares to the seismic trace extracted along the wellbore. The well tie is done for each of the four scenarios:
1. actual DT, 2. ML synthetic DT, 3. Gardner synthetic DT, and 4. constant averaged DT. (Figure 7)

Figure 7—The Four Scenarios Comparison of Each Input Sonic Data.

Result and Discussion. The ML model representativeness is confirmed by blind tests which consist of
two steps. The first step compares the ML synthetic sonic logs to the actual sonic logs. In the second step,
the synthetic seismograms are generated from synthetic sonic and actual sonic. The concept was tested in
two platforms, one in the normaly pressured area of Greater Bongkot North (GBN) and the other in the
overpressured area of Greater Bongkot South (GBS).
In the Greater Bongkot North (GBN) area the ML model was trained from one delineation well and
two development wells, and validated with two development wells which were drilled in the two different
compartment as shown in Figure 8.
SPE-197822-MS 7

Figure 8—Location of Training and Validating Wells in Greater Bongkot North (GBN)

In the Greater Bongkot South (GBS) area the ML model was trained from four exploration/delineation
wells and validated with three development wells which were drilled in three various compartment as shown
in Figure 9.

Figure 9—Location of Training and Validating Wells in Greater Bongkot South (GBS)

The synthetic sonic from machine learning and Gardner methodologies were compared on the Validating
wells based on coefficient of determination (R2). In GBN, the R2 values from machine learning and Gardner
method are 0.75 and 0.28, respectively. In GBS, the R2 values from machine learning and Gardner method
8 SPE-197822-MS

are 0.66 and 0.29, respectively. The results are summarized in Table 1. The comparison of machine learning
synthetic well logs and actual well logs is illustrated in Figure 10 - Figure 13, these logs are look resemble.
The ML methodology shows a much better fit to the real data.

Figure 10—Evaluation of Synthetic Sonic by Machine Learning vs Gardner Method in North Area

Figure 11—Evaluation of Synthetic Sonic by Machine Learning vs Gardner Method in South Area
SPE-197822-MS 9

Figure 12—Comparison of Machine Learning Synthetic logs and Actual Well logs in the North Area

Figure 13—Comparison of Machine Learning Synthetic logs and Actual Well logs in the South Area

Next, the seismic well ties are compared between those two synthetic seismograms. Once the synthetic
and actual logs show similar results, the model is deemed good and used on wells that do not have sonic logs.
10 SPE-197822-MS

According to the blind test results, the synthetic seismograms generated by synthetic well logs and actual
well logs are very similar in term of correlation coefficient, the results are shown in Figure 14 and Table 2.

Figure 14—The bar chart shows correlation coefficient of well ties from each DT; actual
DT, ML synthetic DT, Gardner predicted DT and constant DT for comparison in each well

Table 2—Summary of Correlation Coefficient of Seismic Well Tie Comparison between Using Actual DT
and Predicted DT; Machine Learning Synthetic logs, Gardner Synthetic logs and Constant Sonic Logs

The bar chart in Figure 14 shows the correlation coefficients between seismic and synthetic seismogram
from ML synthetic DT are the most similar to the correlation coefficient from the actual DT. On the
other hand, the correlation coefficient from the Gardner predicted DT is low with significant reasons. It
can be explained by looking at the sonic log data for validating well SV2, Figure 15, together with the
pressure profile in this well. There are two different pressure domains which show slower velocity in the
overpressured upper part and the faster velocity in the normaly pressured lower part. The DT prediction
from the Gardner equation can not capture the change of pressure regime which leads to a different DT
prediction in the lower part of this well. This causes a lower correlation coefficient for the Gardner's method
in abnormaly pressured areas.
SPE-197822-MS 11

Figure 15—Shown the DT Log Curve with Different Acoustic Properties. Slower velocity
in upper parts due to overpressure formation and faster velocity in deeper part due to
compaction. The pressure profile of well SV2 is plotted shown the overpressure interval

Well with no Sonic and Density Logs


Occasionally, some wells do not have both sonic and density logs. In these cases, Gardner's equation and
constant velocity approaches cannot be applied and the only option is to use a velocity model from nearby
well. In contrast, ML can handle this problem. The two options are available are:
1. Predict sonic and density, separately. Then, compute the acoustic Impedance from these two synthetic
logs
2. Directly predict AI
Based on trial, these two options yield comparable results. However, the first option is examined in this
exercise because it can easily be integrated into the existing workflow.
The inputs for the model is similar to the previously explained model, except that the bulk density and
neutron porosity logs are excluded. As a result, the model testing yields R2 of 0.52 in sonic log prediction
which is slightly less than 0.66 in model with density and neutron porosity. Regarding density testing, the
model could provide R2 of 0.51. Note that there are some scatter at the top left conner. These are from
organic shales whose properties usually scatter. However, these points contributes to less than 5% of data.
The example of validity check of sonic and density log for the well SV2 is shown in Figure 16 and Figure 17.
12 SPE-197822-MS

Figure 16—Evaluation of Synthetic Sonic by Machine Learning in case of no Density Log Data.

Figure 17—Comparison of Machine Learning Synthetic logs and Actual Well logs in Case of No Density Log.

The seismic well tie for wells without density and sonic logs is demonstrated in well SP1. This well was
originally planned to acquire gamma ray-resistivity-density-neutron logging suit. Unfortunately, the density
and neutron tools failed. The seismic well ties are conducted in 2 cases, 1. using velocity model from nearby
well- ST1 (conventional method) and 2. using synthetic density and sonic log from ML.
According to well tie results, using ML synthetic logs provides mean-absolute-error (MAE) of 16.78 ms
which is lower than using velocity from nearby well that yields MAE of 23.01 ms, the results are depicted in
Figure 18. This can imply that well tie using ML synthetic log is better than using velocity from nearby well.
SPE-197822-MS 13

Figure 18—Seismic Well Ties in SP1 Well (No Density Logs). Comparison
of Between Using Nearby Well Velocity (ST1) and ML Synthetic Logs

Thus, the ML model was deemed representative and synthetic seismograms were generated using
synthetic sonic logs for all wells that do not have actual sonic logs. The ML model shows flexible application
in various geological settings and avaialable data.
This technique has the potential to save money by cancelling sonic log acquisition. Wireline operations
will also be faster, saving some rig time (and money). Removing the long sonic tool from the wireline
string will also reduce the risk of the logging string getting stuck in the hole, which would require fishing
operations and its associated cost.

Conclusions
1. The machine learning technique is practical to derive sonic logs.
2. Machine learning synthetic log can be used to generate accurate seismic well ties.
3. There are high level of uncertainties with Gardner DT prediction in the deeper parts (>2500mSS)
because of the change of acoustic property related to overpressure formation. The method involving
use of ML does not suffer from this limitations and can be used effectively

Benefits
The benefits of using synthetic sonic logs from Machine Learning can be summarized as,
1. To generate accurate synthetic seismogram to tie wells that do not have sonic data.
2. To reduce the number sonic data acquisition, saving time and money.
3. To decrease the risk of logging string being stuck in the hole.
14 SPE-197822-MS

Acknowledgements
Firstly, authors would like to thank the project owner, Awirut Sirimongkolkitti, Krongpol Seusutthiya, and
Krit Ngamkamollert for their continuous support and their kind and immense knowledge. Their guidance
helped us at all the times.
The authors would like to thank Jakkarin Iamboon and Jutaratt Sirisawadwattana for their
encouragements, insightful comments, challenging questions and sharing their knowledge and experiences.
Last but not the least, the authors would like to thank our vice president Puvanat Chumsena, PTTEP and
Total Thailand Company for supporting and encouraging us to participate in this conference.

References
Achalabhuti, C. 1975. Petroleum Geology of the Gulf of Thailand. Bangkok: Department of Mineral Resources of
Thailand.
Bestagini, P., Lipari, V., Tubaro, S. 2017. A machine learning approach to facies classification using well logs. 2017 SEG
International Exposition and Annual Meeting, Houston, Texas: Society of Exploration Geophysicists.
Bin Ali M. et al 2019. New Plays and New Reservoirs - Reviving Untapped Potential in the Northern Malay Basin,
Southern Gulf of Thailand. 2018 AAPG International Conference and Exhibition (2018), Cape Town.
Boonsri, K. and Natepracha, J. 2014. Potentiality of pre-well planning in development field - the next step to reducing
time and enhancing well planning by collaboration drilling's and geoscience's workflow - case study: development
field in Gulf of Thailand. IADC/SPE Asia Pacific Drilling Technology Conference. Bangkok: SPE.
Charusiri, P. and Pum-Im, S. 2009. Cenozoic tectonic evolution of major sedimentary basins in central, northern, and the
GOT. Bull of Earth Science Tech 2(1 and 2), 40–61.
Duval, B. and Gouadain, J. 1994. Exploration and development in the Bongkot field area (GOT): an integrated approach.
Proceeding of the 14th WPC. John Wiley & Sons.
Freund, Y., and Schapire, R. 1999. A Short Introduction to Boosting. Journal of Japanes Society for Artificial Intelligent,
September 1999.
Kaeng G. et al 2015. Structural and Stratigraphic Controls of Complex Well Planning in the Gulf of Thailand: Case Study
from Arthit Development Field. APGCE (2015), Kuala Lumpur.
Ketmalee, T. and Bandyopadhyay P. 2018. Application of Neural Network in Formation Failure Model to Predict Sand
Production. Offshore Technology Conference Asia, Kuala Lumpur, Malaysia.
Leo, C.T. 1997. Exploration in the Gulf of Thailand in deltaic reservoirs, related to the Bongkot Field. In A.J. Fraser, S.J.
Matthews, and R.W. Murphy (Eds.), Petroleum Geology of Southeast Asia. 126, 77–87. Geol. Soc. London Special
Publication.
Malinur, O. and Cyril U. E. [2010] Application of Novel Machine Learning Algorithms for Facies Classification. AAPG
Africa Region Annual Conference, Abuja, Nigeria.
Maniar, H., Ryali, S., Kulkarni, M. S., Abubakar, A. 2018. Machine-learning methods in geoscience. 2018 SEG
International Exposition and Annual Meeting. Anaheim, California, USA: Society of Exploration Geophysicists.
14-19.
Marsh, B. 2016. Multivariate Analysis of the Vector Boson Fusion Higgs Boson. University of Missouri, 8 August 2016.
Morley, C. 2015. Five anomalous structural aspects of rift basins in Thailand and their impact on petroleum systems.
Journal of Geological Society of London.
Ng, Patrick. 2017. Using Little Big Data and Machine Learning to Manage Risk. AAPG Geosciences Technology
Workshop, Houston, Texas.
Pollock, J., Stoecker-Sylvia, Z., Veedu, V., Panchal, N., Elshahawi, H. [2018] Machine Learning for Improved Directional
Drilling. Offshore Technology Conference. Houston, Texas, USA: Society of Exploration Geophysicists.
Restrepo-Pace P., Dalrymple M., Morley R. 2014. Finding New Exploration Targets in the ‘Mature’ Petroleum Basins
Offshore Thailand: All About Nuance. AAPG International Conference & Exhibition 2015. Istanbul, Search and
Discovery Article #10696 (abstract).
Saifuddin F. 2018. Combination Traps In Jasmine And Ban Yen Fields, Gulf Of Thailand: Identification, Recognition And
Development Strategy. AAPG Asia Pacific Region GTW, Back to the Future – The Past and Future of Oil and Gas
Production in the Asia Pacific Region (2018), Bangkok.

You might also like