SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

DOI: 10.30632/SPWLA-5073

DUAL NEURAL NETWORK ARCHITECTURE FOR DETERMINING

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
PERMEABILITY AND ASSOCIATED UNCERTAINTY
Ravinath Kausik1, Augustin Prado1, Vasileios-Marios Gkortsas1, Lalitha Venkataramanan1, Harish Datir2, Yngve
Bolstad Johansen3

1
Schlumberger-Doll Research Center, 2Schlumberger, 3Aker BP

Copyright 2020, held jointly by the Society of Petrophysicists and Well Log uncertainty, obtained by determining the posterior
Analysts (SPWLA) and the submitting authors.
This paper was prepared for the SPWLA 61st Annual Logging Symposium held weight distributions of the network by using variational
online over 6 sessions, every Wednesday from June 24 to July 29, 2020. inference. This provides us the ability to differentiate in-
and out-of-distribution predictions, thereby identifying
ABSTRACT the suitability of the trained models for application in
new geological formations. The errors in the prediction
The computation of permeability is vital for reservoir of the BNN are fed into a second ANN trained to
characterization because it is a key parameter in the correlate the predicted uncertainty to the error of the first
reservoir models used for estimating and optimizing BNN. Both networks are trained simultaneously and
hydrocarbon production. Permeability is routinely therefore optimized together to estimate permeability
predicted as a correlation from near-wellbore formation and associated uncertainty.
properties measured through wireline logs. Several such
correlations, namely SDR permeability and Timur- The model is trained on a “ground-truth” core database
Coates permeability models using nuclear magnetic representing samples from different geology formations.
resonance (NMR) measurements, K-lambda using The application of the machine learning permeability
mineralogy, and other variants, have often been used, model demonstrates a greater than 50% reduction of the
with moderate success. In addition to permeability, the mean square error in comparison to traditional SDR and
determination of the uncertainties, both epistemic Timur-Coates permeability models (KSDR and KTIM,
(model) and aleatoric (data) are important for
respectively) on wells from the Ivar Aasen Field. We
interpreting variations in the predictions of the reservoir
models. In this paper we demonstrate a novel dual deep also demonstrate how the machine learning workflow
neural network framework encompassing a Bayesian enables us to understand the value of information (VOI)
neural network (BNN) and an artificial neural network of different logging measurements, by replacing the logs
(ANN) for determining accurate permeability values with their median values from nearby wells during model
along with associated uncertainties. inference, and studying the increase of the mean square
error in the permeability predictions.
Deep learning techniques have been shown to be
effective for regression problems, but quantifying the INTRODUCTION
uncertainty of their predictions and separating them into
the epistemic and aleatoric fractions is still considered Permeability is important for appraisal of the reservoir
challenging. This is especially vital for petrophysical because it is an important input in the reservoir models
answer products because these algorithms need the that predict hydrocarbon production. Several methods
ability to flag data from new geological formations that have been developed for the determination of
the model was not trained on as “out of distribution” and permeability using different well logs and calibrated
assign them higher uncertainty. Additionally, the model through core measurements. In this publication we
outputs need sensitivity to heteroscedastic aleatoric noise describe a novel dual neural network pipeline that can
in the feature space arising due to tool and geological provide an improvement over existing models for
origins. Reducing these uncertainties is key to designing permeability prediction. This methodology can also take
intelligent logging tools and applications, such as into account the uncertainties present in heteroscedastic
automated log interpretation. training datasets from the different log and core
measurements, and provides confidence intervals for the
In this paper we train a BNN with NMR and mineralogy predicted permeability values, reflecting whether the
data to determine permeability with associated epistemic data points are in- or out-of-distribution.
1

SPWLA-5073
SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

There have been several models developed over the other than mineralogy based methods such as K-lambda
years through the use of well logs and core data for (Herron et al., 1998), have also been proposed. The
permeability determination (Herron et al., 1987, 1998; application of machine learning based approaches for the

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
Allen et al., 2000; Freedman et al., 2006; Kenyon et al., determination of petrophysical properties including
1988). Some of the popular permeability models include permeability have also been investigated.
SDR permeability (KSDR) and Timur-Coates (KTIM)
models. Based on early research done on sandstones, the METHODS
SDR permeability estimation was empirically derived as
(Kenyon et al., 1988), The application of machine learning, especially neural
networks, has been shown to be extremely effective for
𝑐
KSDR = 𝐴𝜑 𝑏 𝑇2,𝐿𝑀 , (1) various classification and regression problems, but
quantifying the uncertainty of their predictions and
where φ is the porosity of the rock, 𝑇2,𝐿𝑀 is the separating them into the epistemic and aleatoric fractions
logarithmic mean of the 𝑇2 distribution, and A the is still considered challenging. In petrophysical
formation-dependent scalar factor. The parameters b and applications, logging tool measurements are used to
build reservoir models that are subsequently used for
c are empirically determined through calibration with
estimation and optimization of hydrocarbon production.
core measurements. More recently, this relation was
The machine learning algorithms are often used on such
improved through the inclusion of the surface relaxivity
logging tool data to estimate the subsurface rock and
of the rock (Souza et al., 2013), fluid properties. Quantifying the uncertainties of these
𝑐 petrophysical properties is crucial for rock and fluid
𝑘𝜌 = 𝐴𝜑 𝑏 (𝜌2 𝑇2,𝐿𝑀 ) , (2) evaluation and subsequent reservoir optimization and
production decisions. These machine learning
where the relaxivity 𝜌2 is estimated either from algorithms are generally trained on ground-truth of a
diffusion-relaxation (D-T2) maps or from comparison of core database from which carefully measured
nuclear magnetic resonance (NMR) and mercury petrophysical properties are available. During the
injection capillary pressure (MICP) porosimetry data. inference (testing) phase involving application of these
Unfortunately, estimating relaxivity with either method algorithms to field data, it is critical that the machine
is time consuming and challenging. Estimating the learning algorithm flag data as out-of-distribution, data
surface relaxivity from the mineralogy is also from new formation geology or feature properties that
challenging because even though there should be an the model was not trained on. It is also important that the
underlying relationship between relaxivity and the models are sensitive to heteroscedastic aleatoric noise in
the feature space arising from the combination of tool
concentration of paramagnetic elements (e.g., Fe, Mn),
and geological conditions. This would give us a clear
the exact functional form of the relationship is unknown
idea of the region of the input feature space where the
(Foley et al., 1996; Keating and Knight, 2008).
uncertainty in the outputs is high and help us with future
field development plans. Therefore, understanding the
The Timur-Coates model for permeability is given as sources of the uncertainty and reducing them is key to
(Timur et al., 1968) designing intelligent tools and applications such as
automated log interpretation answer products for
FFV 2
KTIM = 𝐵𝜑 4 ( ) , (3) exploration and field development. In this paper, we
BFV
describe a methodology consisting of a system of dual
neural networks comprising of the combination of a
where B is another scalar factor and FFV and BFV
Bayesian neural network (BNN) and an artificial neural
represent the free fluid and bound fluid volumes, network (ANN) addressing the determination of
respectively, obtained from 𝑇2 distributions with permeability and associated uncertainties.
appropriate mineralogy dependent cutoffs. The
parameters in these models such as A and B are BNNs have been traditionally used for epistemic
calculated through laboratory measurements, where the uncertainty estimation in supervised learning settings for
ground-truth rock permeability is measured with fluid both regression and classification problems. Recently,
(e.g., helium or nitrogen gas) flow measurements on core dropout-based neural networks have also been proposed
samples and the results correlated with 𝑇2 distributions. for such purposes and their equivalence to BNNs has
Different studies have also shown a strong correlation been demonstrated. BNNs work by determining the
between the porosity and permeability for different rock posterior weight distributions of the neural networks,
types. Modifications of the KTIM and KSDR methods, providing us with both a mean and variance of the
2

SPWLA-5073
 SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

1 1
estimated outputs by using variational inference. 𝐿(𝑥, 𝑦, 𝜃) = ∑𝑁
𝑖=1 ‖𝑓𝜃 (𝑥𝑖 ) − 𝑦𝑖 ‖2 + log[𝜎𝑖2 ], (4)
𝑁 𝜎𝑖2
Dropouts in neural networks were initially introduced as
a technique for regularization during training. The

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
dropout methods involve random dropping or switching- where the first term represents the accuracy of the model
off of neurons from the network, thereby ensuring that and the second term prevents the assignment of very high
the resultant model avoids overfitting. More recently uncertainty values for all data points. Another method
dropout has been used as a Bayesian approximation that has been recently introduced, is the deep direct
through its application in both neural network training estimation method (Adler et al., 2018), which proposes
and testing phases. Both the BNN and dropout neural the application of two coupled ANNs. One of the ANNs
network approaches provide the model (epistemic) is used to approximate the conditional mean and the
uncertainties and therefore can be used to differentiate other ANN is used to estimate the conditional point-wise
in- and out-of-distribution predictions (Gal et al., 2016). variance. However, a separation of the epistemic and
In petrophysical applications it is very important to aleatoric uncertainties cannot be made.
identify the suitability of the trained machine learning
models for application in new geological formations. In this paper, we demonstrate a methodology consisting
of a dual network system of a BNN and an ANN for
determining the total uncertainty and its separation into
the epistemic and aleatoric fractions. The schematic
representation of the dual network workflow is shown in
Figure 1 (Prado et al., 2019). The details of the workflow
are as follows. The first part consists of a neural network
that predicts the epistemic uncertainty using variational
inference. This could be either a BNN that is trained to
estimate the mean and variance of the output using
variational inference, with its weights assumed to be
Gaussian, or an ANN with dropout. The errors in the
prediction of the BNN, namely (𝑦̂𝑖 − 𝑦𝑡𝑟𝑢𝑒 ), are fed into
a second ANN. The goal of the second ANN is to
estimate the total variance using the direct estimation
method (Adler et al., 2018) trained to correlate the
features to the mean square error of the first BNN, as
shown in Eq. (5). Both networks are trained
simultaneously and are therefore optimized together.
The cost function of the second network is given as
Figure 1: Dual network system consisting of a BNN and
ANN in which the BNN is used to estimate a mean and 2 2
variance (epistemic uncertainty) of the output. The error min ∑𝑁 ′
𝑖=1[(ℎ𝜗 (𝑥𝑖 , Σ𝑖 , 𝑦𝑡𝑟𝑢𝑒,𝑖 ) − (𝑦
̂𝑖 − 𝑦𝑡𝑟𝑢𝑒,𝑖 ) ] . (5)
𝜗
from the BNN is an input into the second ANN for the
determination of the total uncertainty (σtot). The Whereas the first neural network based on Bayesian
difference between this uncertainty and that of the BNN inference estimates the mean value of the output (𝑦̂) and
(σ1) is a reflection of the aleatoric uncertainty of the data the epistemic uncertainty 𝜎1 , the second network outputs
(Prado et al., 2019). the total uncertainty. The difference enables us to have
an idea of the aleatoric uncertainty (𝜎2 ) for each data
On the other hand, aleatoric uncertainty is due to the point,
noise inherent in the observations, denoted by Σ ′ , and can
2
be heteroscedastic. The nonstationary nature of noise ℴ𝑡𝑜𝑡 = ℴ12 + ℴ22 . (6)
could be due to properties of the geology or the response
of different sensors to environmental factors. Recently, a The total uncertainty can then be further calibrated to
few different methods have been introduced to assign obtain better results for diverse training datasets
heteroscedastic data dependent standard deviations to (Kuleshov et al., 2018). One of the key benefits of this
outputs of neural networks. Gal et al. (2016) do this by approach is a better understanding of the components of
minimizing a cost function loss L with respect to the the uncertainty, because the epistemic uncertainty is a
model parameters 𝜃 and variance 𝜎𝑖2 corresponding to good metric to differentiate in- and out-of-distribution
output 𝑦𝑖 , datasets (especially valuable for testing formation
evaluation models in different geological formations),

SPWLA-5073
SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

whereas the aleatoric uncertainty is sensitive to the 𝑀 = 𝐾𝐹 + 𝐸, (8)

heteroscedastic noise inherent to the data within the
training feature space (valuable in understanding the where the matrices K and F are the kernel and the

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
regions of high uncertainty within the feature relaxation distribution, respectively. The dimensionality
distributions). Understanding the aleatoric uncertainty reduction of the kernel can be carried out using singular
originating from the heteroscedastic noise in the feature valued decomposition as
space can also help with logging program design and to
optimize sampling schemes. This deep learning 𝐾 = 𝑈Σ𝑉 𝑇 , (9)
workflow is therefore a powerful method for geophysical
applications. where U and V are the unitary matrices and Σ is a
diagonal matrix with the singular values in decreasing
In this work, we determine the permeability of the order along the diagonal. As the singular values quickly
formation as a function of depth using NMR and nuclear decay to zero, the first few singular values (six in this
spectroscopy log inputs. The approach therefore can be analysis) are sufficient to represent the entire dataset.
seen as an extension over earlier workflows such as Equation 8 can be rewritten as
KSDR, KTIM, or K-lambda, and with significant
improvements. The NMR measurements are considered ̃ = 𝐹̃ + 𝐸,
𝑀 (10)
an important input into permeability predictions because
surface relaxation makes relaxation distributions where 𝑀 ̃ is equal to UTM and 𝐹̃ equals ΣVTF and is the
sensitive to pore geometry. Because the pore geometry reduced dimensional relaxation distribution. This helps
plays a critical role in the flow of fluids through the us reduce the entire relaxation distribution into few (six
porous media, parameters related to it have been used for in this analysis) features that can be used as the input for
correlating it to permeability. KSDR uses the logarithmic the machine learning model. These six features represent
mean of the relaxation times while KTIM uses the free the entire relaxation distribution and therefore contain
to bound fluid ratios, determined using a cutoff value in more complete information of the pore geometry, which
the relaxation distribution, which varies according to is traditionally represented only through a log-mean
lithology. Improvements such as those shown in Eq. (2) relaxation time in KSDR or a free versus bound fluid
have incorporated the surface relaxation into this ratio in KTIM.
workflow for better permeability predictions. In methods
such as K-lambda, a correlation is established between To address the effects of the surface relaxation and the
the permeability and mineralogy of the rock, which can mineralogy, we also input the elemental information
obtained through inelastic capture spectroscopy
reflect the pore geometry in a certain narrow sample
measurements. Nine elements namely silicon,
space such as siliciclastic rocks. In our work, a very
aluminum, magnesium, calcium, sodium, potassium,
general approach toward the estimation of permeability
titanium, manganese, and iron are used as inputs for the
from pore geometry and mineralogy is undertaken. permeability model. In summary, the six features from
the NMR relaxation distributions and the nine dry-
In NMR logging, radio frequency pulse sequences are weight elemental yields form the input features for the
applied and raw spin echo decays reflecting formation determination of the machine learning model for
properties are measured. The amplitudes of these echo permeability. Additionally, features such as the total
decays are inverted to obtain the relaxation distributions. porosity, log mean relaxation times, and free versus
For the case of the most commonly used Carr-Purcell- bound fluid ratios are also included to make the learning
Meiboom-Gill (CPMG) pulse sequence, the echo decay process robust. The permeability model is trained on
can be given as datasets for which the ground truth core permeability
𝑡
measurements are available. A schematic for the training
−
𝑀(𝑡) = ∫ 𝑒 𝑇2 𝑓(𝑇2 )𝑑𝑇2 + 𝐸(𝑡), (7) of the machine learning model is shown in Figure 2. The
model is trained using leave-one-out cross validation to
where E(t) is the additive white Gaussian noise, 𝑓(𝑇2 ) is avoid overfitting and to obtain the best model for the
the T2 relaxation distribution, 𝑀(𝑡) is the spin echo testing. A multilayer neural network with ReLU
decay, and exp(–t/T2) is the exponential kernel relating activation function and an Adam optimizer is used for
the quantities. This equation is a Fredholm integral of the the permeability determination.
first kind and similar formulations can be considered for
the T1 or the 2D T1-T2 experiments. This relationship can
be rewritten in discretized matrix form as
(Venkataramanan et al., 2002)
4

SPWLA-5073
 SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

RESULTS AND DISCUSSION based permeability estimates involving input features

from NMR and spectroscopy data fare better than the
The training and testing of the machine learning traditional KSDR and KTIM permeability, obtained

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
permeability model was done on three wells from the using only the NMR data. The results of this analysis are
Ivar Aasen oil field. Figure 3 shows the cross section of also shown in Table 1, in which the mean square error
the Ivar Aasen oil field, located on the Gudrun Terrace (MSE) of the permeability prediction for the different
on the Eastern flank of the Viking Graben in the techniques is shown for each of the wells. The machine
Norwegian North Sea. The reservoir is located within a learning based permeability from the dual neural
sedimentary sequence of Mid-Jurassic to Late Triassic network has more than 50% lower error in comparison
age, which consists of shallow marine to fluvial, alluvial, with the traditional techniques as summarized in Table
floodplain, and lacustrine deposits overlying a regionally 1. The traditional KTIM model was fit with A=1, b=2,
extensive fractured calcrete interval. The stratigraphic c=4 as shown in equations 2. The results with the
sequence exhibits a complex mineral composition and is comparisons of the machine learning permeability and
heterogeneous at a subsensor scale. Shale layers, KTIM versus that of core samples are shown in figures
redeposited shale, and calcrete fragments are present in 4-6 for the three geopilot wells. It can be seen that the
various forms throughout the sequence. As a result, core machine learning permeability is a better match in-
data is essential, but analyzing the reservoir interval at a comparison to core permeability, and this is especially
high sample rate is costly and, in some cases, true in the Alluvial Fan section of the basin.
impractical. This implies that inference from logs such
as NMR are essential because they provide a continuous Table 1: Comparison of the MSE of the permeability
depth profile of total porosity, bound fluid, and empirical predictions from the machine learning model (KDNN)
permeability (KTIM, KSDR), as discussed earlier in the versus those from traditional techniques such as KSDR
text. and KTIM.

The three ‘geopilot’ wells from this oilfield used in the 16/1-22S 16/1-21A 16/1-21S
machine learning workflow have data from a
comprehensive list of advanced wireline tools, in Mean square error Mean square error Mean square error
(log mD) (log mD) (log mD)
addition to the core data that is used to improve the
understanding of the Ivar Aasen sedimentology and its KTIM 1.84 1.59 2.9
impact on the reservoir performance. The specialized
wireline logs are used to characterize mineralogy (via
KSDR 1.05 0.98 1.17
neutron-induced GR spectroscopy), saturation in
laminated shale/sand sections (triaxial induction), fluid
KDNN 0.45 0.52 0.43
identification and pore volumes (well formation
testing/sampling, NMR), and finally near-wellbore
saturation using nonresistivity-based methods (NMR,
dielectric dispersion). The improved understanding of
the reservoir properties based on the interpretation of the The machine learning workflow also enables us to
advanced wireline data has contributed to a significant understand the value of information (VOI) or the role of
increase in the in-place volumes in the Triassic the different logging measurements in estimating
Skagerrak 2 reservoir zone (Johansen, Y. B., et al., permeability. During the testing or inference phase of the
2018). These results have led to substantial workflow, the value of the different features can be
modifications of the drainage strategy and the replaced by the median of the values obtained from the
development campaign of the Ivar Aasen field. One of training data from nearby wells. The impact this has on
the critical factors was the estimates of the net reservoir, the MSE of the permeability provides us with an
which were done through the application of the KTIM indication of the value or the importance of the feature
(Johansen, Y. B., et al., 2018) showing the value of NMR input (and therefore the logging measurement) in
permeability determination. estimating permeability. The results for such an analysis
for all three wells are shown in Table 2. The machine
The machine learning permeability model was trained learning based dual neural network permeability model
using two of the wells and tested on the third, repeated (KDNN) MSE values include all input features from the
across all three wells. The permeability results with NMR and nuclear spectroscopy measurements during
respect to the core permeability data along with the the inference stage. The second row constitutes
uncertainty for the three wells are shown in Figures 4-6. permeability measurements for which the elemental
It can be seen from the figure that the machine learning composition from the GR spectroscopy is not used, and
instead their median values from the training wells are
5

SPWLA-5073
SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

used. In the third row a similar approach is applied to the inelastic & capture GR spectroscopy in addition to total
NMR data as only the median values from the training porosity, NMR logarithmic mean relaxation times, and
wells are used. In the fourth row only the total porosity free to bound fluid ratio are also used as inputs. It is

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
values are used for the determination of the permeability shown that the machine learning model provides
and the NMR and nuclear spectroscopy data are replaced superior permeability values in comparison to traditional
by the training dataset median values. It can be clearly methods such as KSDR and KTIM based on comparison
seen that as more of the logging measurements are to core permeability measurements. The value of
replaced by the just median values, the MSE values for information of each of these input measurements is also
the permeability predictions increase, illustrating the demonstrated.
impact of each of the missing measurements. The NMR
measurements have the highest value of information ACKNOWLEDGMENTS
followed by the elemental information from nuclear
spectroscopy logs. This is a logical outcome as NMR The authors would like to thank Schlumberger, AkerBP
measurements are sensitive to the pore geometry and and the Ivar Aasen licensees for allowing the publication
have been traditionally used for the determination of of this work.
permeability through the KSDR and KTIM models.
More importantly, it can be seen that the combination of REFERENCES
the NMR and nuclear spectroscopy measurements
together provides better permeability predictions than Herron, MH., Johnson, DL., Schwartz,LM., “A robust
either one of them by itself, showing the value of permeability estimator for siliciclastics”. In: SPE Annual
additional information for answer product determination.
Technical Conference and Exhibition. Society of
Petroleum Engineers. 1998.
Table 2: The VOI of different measurements shown by
the increase in the MSE values for permeability
Herron MH.,. “Estimating the intrinsic permeability of
prediction in their absence during model inference.
clastic sediments from geochemical data”. In: SPWLA
28th Annual Logging Symposium. Society of
16/1-22S 16/1-21A 16/1-21S
Petrophysicists and Well-Log Analysts. 1987.
Mean square Mean square Mean square
error (log mD) error (log mD) error (log mD)
Kenyon, WE.,. “A three-part study of NMR longitudinal
relaxation properties of water-saturated sandstones”. In:
KDNN 0.44 0.52 0.43
SPE formation evaluation 3.03 (1988), pp. 622–636.
No 0.45 0.81 0.44
Timure, A., “An investigation of permeability, porosity,
Spectroscopy
and residual water saturation relationships”. In: SPWLA
inputs
9th annual logging symposium.Society of Petrophysicists
and Well-Log Analysts. 1968.
No NMR 0.89 1.62 1.46
inputs
Allen. D., Flaum, C., Ramakrishnan, TS., Fairhurst, D.,
Pritchart, T., Bedford, J., Gebelin, G., Heaton, N., minh,
No NMR/ 2.0 2.10 1.25
CC., Ramamoorthy, R., Castelijns, K., Norville, M.,
Spectroscopy
inputs Seim, M., “Trends in NMR logging”. In: Oilfield Review
12.3 (2000), pp. 2–19.

Robert Freedman, R., “Advances in NMR logging”. In:

CONCLUSION Journal of Petroleum Technology 58.01 (2006), pp. 60–
66.
A novel dual neural network machine learning
architecture that can provide estimates of the Souza, A., Carneiro, G., Zielinski, L., Polinski, R.,
permeability and associated uncertainty with NMR and Schwartz, L., Hürlimann, M. D., de Vasconcellos
nuclear spectroscopy measurements as inputs is Azeredo, R. B. . “Permeability Prediction Improvement
described. The NMR relaxation distribution is reduced to Using 2D NWR Diffusion-T2 Maps”. In: SPWLA 54th
six dimensions using singular valued decomposition and Annual Logging Symposium. Society of Petrophysicists
used as inputs for the model. The elemental outputs from and Well-Log Analysts. 2013.

SPWLA-5073
 SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

Foley, I., Farooqui, SA., Kleinberg, RL., 1996 “Effect of development of next-generation NMR diffusion and
paramagnetic ions on NMR relaxation of fluids at solid relaxation measurements. He has served thrice as a
surfaces”. In: Journal of Magnetic Resonance, Series A distinguished speaker for SPWLA, and on the

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
123.1 (1996), pp. 95–104. international advisory committee of the Magnetic
Resonance in Porous Media (MRPM). He obtained a
Keating, K., Knight, R., “A laboratory study of the effect MSc from IIT Madras, India, and PhD degree in physics
of magnetite on NMR relaxation rates”. In: Journal of from the Universität Ulm, Germany. He worked as a
Applied Geophysics 66.3-4 (2008), pp. 188–196. postdoctoral fellow at the University of California, Santa
Barbara before joining SDR in 2009. He has co-authored
Gal, Y., and Ghahramani. Z., “Dropout as a Bayesian more than 30 peer-reviewed publications and several
approximation: Representing model uncertainty in deep patent applications and is a scientific reviewer for more
learning.” international conference on machine than 10 international journals.
learning. 2016.
Augustin Prado is a PhD candidate at the Imperial
Jonas, A, Ozan, K., ”Deep bayesian inversion.” arXiv College in London. He has obtained a Bachelor and
preprint arXiv:1811.05910 (2018). Master’s degree in mathematics and data science
respectively from the Ecole polytechnique federale de
Kuleshov, Volodymyr, Nathan Fenner, and Stefano Lausanne and was a machine learning intern at
Ermon. ”Accurate uncertainties for deep learning using Schlumberger-Doll research center from June 2018-June
calibrated regression.” arXiv 2018). 2019.

Prado, A., Kausik, R., Venkataramanan, L., Dual Neural Vasileios-Marios Gkortsas is a
Network Architecture for determining Epistemic and Senior Research Scientist at
Aleatoric Uncertainties, arXiv: 1910.06153 (2019) Schlumberger-Doll Research
Center (SDR) at Cambridge, MA,
Venkataramanan, L., Song, Y.-Q., Hürlimann, M.D., USA. He received a Diploma in
2002, Solving Fredholm integrals of the first kind with electrical and computer
tensor product structure in 2 and 2.5 dimensions, IEEE engineering from the National
Transactions on Signal Processing, 50(5). 1017–1026. Technical University of Athens
(Greece) in 2006, an SM in applied
Johansen, Y. B., Christoffersen, K., Chatterjee, A., physics from Harvard University in 2008, and a PhD in
Elfenbein, C., Olsborg, L. H., Kvilaas, G. F., Hurlimann, electrical engineering and computer science from the
M. “Use of Advanced Wireline Logs To Reduce Massachusetts Institute of Technology in 2013. He
Uncertainties in a Complex Reservoir: A Case Study joined SDR in 2013 and has worked on the development
From the Ivar Aasen Oilfield in the Norwegian Central of models that describe the effects of wettability on NMR
North Sea” Society of Petrophysicists and Well-Log and dielectric measurements, the development of
Analysts (2018, June 2). algorithms that improve the precision of answer
products, and on application of machine learning
techniques for cement evaluation. His current interests
include the development and application of machine
ABOUT THE AUTHORS learning and deep learning techniques for the automation
of log interpretation. He has co-authored 25 peer-
Ravinath Kausik K.V is a reviewed publications and several patent applications.
Senior Research Scientist in the
Applied Math and Data Lalitha Venkataramanan is a
Analytics program at the Scientific Advisor and Program
Schlumberger-Doll Research Manager of the Automated Log
Center (SDR) in Cambridge, Interpretation program in the
Massachusetts, USA. His work Applied Math and Data Analytics
has focused on the development department at Schlumberger-Doll
of novel petrophysical Research Center, Cambridge. She is
techniques such as TGIP-NMR and RPI for the Associate Editor for NMR-
unconventional shale gas and tight oil plays and the Petrophysics and a regional
7

SPWLA-5073
SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

distinguished speaker for SPWLA for 2018-19. She is on

the board of SIAM and NSERC as well as the Business-
Industry-Government Math network. Her current

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
interests include machine learning, mathematical
modeling and inversion, optimization, probability, and
stochastic processes. Trained as an electrical engineer,
she obtained her MS and PhD degrees from Yale
University in 1998. She has co-authored more than 40+
peer-reviewed publications and has over 24 granted US
patents and 18 pending patent applications.

Harish B. Datir is a Petrophysics

Domain Champion, based in
Stavanger, Norway. He joined
Schlumberger in 2007 as a Wireline
Field Engineer in Saudi Arabia, and
since then he has taken different roles
in different locations prior to assuming his current role
as Petrophysics Domain Champion. He worked in the
Middle East for 4 years, and for the past 9 years he has
been working in Scandinavia. He has a BS degree in
applied geology and a master’s degree in applied
geophysics. His work focuses on improving and
developing integrated petrophysical answer products.

Yngve Bolstad Johansen is Chief

Petrophysicist for AkerBP. He received
a Cand.Scient. degree in physics from
NTNU, Norway, in 2000. His thesis was
delivered within the field of
measurement physics (NMR). He joined
Schlumberger in 2001 as a Field
Engineer and worked with data
acquisition in the Middle East and North Sea before
starting as a Logging Tool Physicist at the Houston
Formation Evaluation Center in 2005. After 2009 he
started as Principal Petrophysicist for Statoil, and two
years later he moved to Statoil R&D center in
Trondheim. From 2013 on, Yngve has been leading
Petrophysics within Aker BP with the ambition to build
a world-leading team of petrophysicists. He is currently
a member of SPWLA, SPE, and SCA.

SPWLA-5073
 SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
Figure 2: Workflow for the determination of permeability using input features from the NMR relaxation distribution
and nuclear spectroscopy measurements. The NMR T2 distributions are reduced to six features using singular valued
decomposition, and nine elements from GR spectroscopy are also used. Additionally porosity, logarithmic T2, and free
to bound fluid ratios are also input into the model to predict the permeability.

Fig 3. Cross section of the sedimentary sequence that comprises the Ivar Aasen (IA) reservoir. The wells names
highlighted in blue are geopilot wells where the machine learning permeability model were tested. 16/1-21S and 21A
were cored and logged with advanced wireline logs in the east of IA and 16/1-22S in the west.

SPWLA-5073
SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
Figure 4: The interpretation for the geopilot well 16/1-21A from the Ivar Aasen basin. The first three tracks show the
density-neutron logs, the fourth track the mineralogy, and the fifth and sixth tracks are the NMR porosity and T2
distribution, respectively. The final tracks compare the traditional permeability interpretations of KTIM and the
machine learning based permeability with the ground-truth core permeability.

10

SPWLA-5073
 SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
Figure 5: The interpretation for the geopilot well 16/1-21S from the Ivar Aasen basin. The first three tracks show the
density-neutron logs, the fourth track the mineralogy, and the fifth and sixth tracks are the NMR porosity and T2
distribution, respectively. The final tracks compare the traditional permeability interpretations of KTIM and the
machine learning based permeability with the ground-truth core permeability.

11

SPWLA-5073
SPWLA 61st Annual Logging Symposium, June 24 to July 29, 2020

Downloaded from http://onepetro.org/SPWLAALS/proceedings-pdf/SPWLA20/36-SPWLA20/D363S022R001/2246043/spwla-5073.pdf by Khalifa University of Science and Technology user on 12 October 2021
Figure 6: The interpretation for the geopilot well 16/1-22S from the Ivar Aasen basin. The first three tracks show the
density-neutron logs, the fourth track the mineralogy, and the fifth and sixth tracks are the NMR porosity and T2
distribution, respectively. The final tracks compare the traditional permeability interpretations of KTIM and the
machine learning based permeability with the ground-truth core permeability.

12

SPWLA-5073