International Geomechanics Symposium IGS-2022-184

Well Log Prediction Using Deep Sequence Learning

Downloaded from http://onepetro.org/armaigs/proceedings-pdf/IGS22/All-IGS22/ARMA-IGS-2022-184/3055616/arma-igs-2022-184.pdf/1 by Nelly Rubio on 02 March 2024

Weichang Li1* and Lei Fu1
Aramco Americas, Houston Research Center, Houston, Texas, United States
Murtadha J. AlTammar2
Saudi Aramco, Dhahran, Kingdom of Saudi Arabia

Copyright 2022 ARMA, American Rock Mechanics Association

This paper was prepared for presentation at the International Geomechanics Symposium (IGS) on 7 - 10 November 2022 in Abu Dhabi, UAE.
This paper was selected for presentation at the symposium by an IGS Technical Program Committee based on a technical and critical review of
the paper by a minimum of two technical reviewers. The material, as presented, does not necessarily reflect any position of the partner societies
ARMA/DGS/SEG/AAPG/SPWLA/EAGE/SPE, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the partner societies is prohibited. Permission to reproduce in print is restricted to an
abstract of not more than 200 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgement of where and by
whom the paper was presented.

ABSTRACT: Sonic logs including compressional and shear travel time logs (DTC and DTS, respectively) are important
measurements for subsurface elastic and geomechanic property characterization. However, these log types are not always measured
in practice or incomplete in many oil and gas wells for economic and other practical reasons. We propose to accurately predict these
types of log data from the traditional common types of well log measurement data using deep sequence learning methods. After
anomalous data removal and augmenting the ratio DTC/DTS as a new lithology differentiating feature, the preprocessed inputs were
fed into a bidirectional Long Short-Term Memory (LSTM) model for training to minimizing the error in predicting the sonic logs.
Once trained, the model is then applied at the target wells to predict the sonic logs. The predicted sonic logs match the actual values
with good accuracy. This entirely data-driven approach significantly reduces dependency on prior domain knowledge, providing a
generalization advantage that enables automated large-scale well log prediction across fields.

are important in many applications. For instance, seismic-

1. INTRODUCTION
Well-log data analysis and interpretation are commonly
conducted and play a central role in quantitative reservoir
characterization, formation and completion evaluation
(Ellis 2007, Glover 2014). Various borehole
measurements provide information to determine reservoir
rock composition such as solid and fluid volume
fractions, as well as rock types, etc. When integrated with
seismic data, multiple types of well log data can help
reduce geologic interpretation ambiguity and improve
hydrocarbon reservoir models. Certain well logs, like GR,
resistivity, density, and neutron, are considered as “easy-
to-acquire” conventional well logs and are deployed in
most wells. However due to cost consideration or access
limitation, other types of well logs, like nuclear magnetic
resonance (NMR), dielectric dispersion, elemental
spectroscopy, and dipole/shear sonic, are deployed in a
limited number of wells and not as commonly available,
or missing at certain depth intervals in an area of interest.
On the other hand, sonic or acoustic logs, which measures
the travel time of an elastic wave through the formation,
well tie requires sonic and density logs as inputs.
Geomechanical parameters, very challenging to directly
measure, may be derived from sonic logs aided with
additional information such as rock types (Chen and
Zhang 2020). Additionally, sonic logs also provide
information to derive formation porosity, for stratigraphic
correlation, and identification of lithologies, facies,
fractures and compaction (Glover 2014).
Traditionally, various regression methods and algorithms
based on statistical, deterministic or empirical models
such as petrophysics and rock-physics equations, have
been applied to reconstruct missing well log data from
logs measured at nearby wells (Bader et al., 2019). For
example, Gardner’s equation establishes the relationship
between sonic logs and density for brine-saturated rock
types (Gardner et al., 1974). Interval relationships
between resistivity and sonic logs can be traced to the
methods proposed by Faust (1953) and more recently by
Smith (2007). Empirical relationships developed by
Castagna et al. (1985) and Greenberg and Castagna
(1992) allows calculating shear wave (S-wave) velocity
from compressional wave (P-wave) velocity. However,
the results from these approaches can be sensitive to the
particular rock types in the depth intervals and require
laborious and time-consuming calibration with expert
knowledge. Generally speaking, relationship between
conventional well logs (such as Neutron, Gamma Ray,
Resistivity, and Density) and sonic logs can be highly
complex, and sensitive to many factors including the
formation properties as well as how these logs are
measured. Therefore, deterministic physics modeling
effort to establish this type of relationship risks model
mismatch and will inevitably poses significant
uncertainty.
Machine learning approaches, especially deep neural
network model based methods, are capable of
approximating nonlinear highly complex functional
relationships. Especially sequence learning, a special type
of machine learning techniques for applications involving
sequence inputs and/or outputs, have been proven
effective in applications such as automatic speech
recognition (ASR) and genomics. Recently, data driven
approaches, especially deep learning models and
algorithms, have been increasingly applied to predict
missing well log information, as summarized, for
example, by Yu et al. 2021 based on the 2021 SPWLA
Machine Learning Challenge. Missing sonic logs were
predicted from gamma-ray, neutron porosity, and density
logs as given in Saggaf and Nebrija (2003), using
regularized back-propagation neural networks. Fully
connected neural networks (FCNNs) are also used by
Rolon et al. (2009) and Salehi et al. (2017), producing a
point-wise mapping from input to output logs. Recurrent
neural networks (RNNs), capable of taking input logs
both from a previous depth range and other available logs
at the corresponding depth interval, were used by Zhang
et al. (2018) to generate synthetic well logs. More
recently, Feng et. al. 2020 applied Bayesian Neural
Network (BNN) to predict compressional sonic log from
porosity, gamma ray, deep resistivity, photoelectric factor
and bulk density, over certain depth intervals. Pham et. al.
2020 proposed to use Convolutional Long Short-term
Memory (LSTM) model to predict sonic logs from
gamma ray, density and neutron porosity.
2. METHODOLOGY
2.1 Problem set-up
We focus on one typical setting for machine learning
based well log prediction: nearby training wells or well
depth intervals with a complete suite of logs are used to
predict missing logs at a specific well or well intervals
where only certain “easy-to-acquire” conventional well
logs are available, as depicted in Figure 1. This is more of
a post-drill scenario, which is different and should not be
confused with the problem of predicting synthetic logs at
locations where there is no data, such as in pre-drill
scenario or 3D spatial extrapolation of well log data. In
this setting, the types of logs available at the testing well
or the well depth intervals are the input log data, and the
types of missing logs we are interested in predicting as the
output or target log data. The selected machine learning
models are first trained at the training wells or well depth
intervals, where it minimizes the error between the
predicted versus actual target log data, based on the input
log data (lower half of Figure 1). The trained models are
then deployed to predict the target log data missing at the
Downloaded from http://onepetro.org/armaigs/proceedings-pdf/IGS22/All-IGS22/ARMA-IGS-2022-184/3055616/arma-igs-2022-184.pdf/1 by Nelly Rubio on 02 March 2024
testing well or well depth interval, based on the input log
data available at these locations (top half of Figure 1).
For sonic log prediction, we consider predicting
compressional (DTC) and shear (DTS) travel time logs
from the conventional well logs such as neutron, Gamma
ray, density etc.
Fig. 1. Typical setup and workflow for machine learning based
well log prediction.
2.2 Data preprocessing
Data exploration, quality control and extraction of useful
information from the features are important to ensure
successful application of machine-learning models as
well as provide interpretation and understanding if
something goes wrong. Several basic data preprocessing
steps include outlier and non-numerical (NaN) valued
data removal potentially associated with bad hole
intervals and measurement issues for the logs,
transforming resistivity logs into logarithm scale, and
normalizing the input logs. In addition, support vector
machine (SVM) classifier is applied to detect the
anomalous data points, and the ratio DTC/DTS was added
as a new feature to represent different lithology.
2.3 Machine learning models and algorithms
Deep learning methods optimize the network coefficients
to achieve both feature learning to capture the most
relevant representative structures and statistics and
minimizing the prediction error between the predicted and
the true output values (Goodfellow et al., 2016, Zhang et
al. 2021). It has shown to be quite successful in a host of
applications in computer vision, natural language
processing, etc. Deep sequence learning (Lipton et al. on a single training set, and the training model was
2015) addresses tasks involving sequential data where applied to the entire test well or well depth intervals.
sequence dynamics, whether in time or space, are
captured by certain network model structures such as
recurrent networks. The long short-term memory (LSTM)
network is used to take into account long-term
dependencies and was introduced by Hochreiter and
Schmidhuber 1997 to offer a solution to the vanishing
gradients problem.

Downloaded from http://onepetro.org/armaigs/proceedings-pdf/IGS22/All-IGS22/ARMA-IGS-2022-184/3055616/arma-igs-2022-184.pdf/1 by Nelly Rubio on 02 March 2024

Fig. 3. LSTM network structure consisting of repeating
modules each with four interacting layers (left to right): 1.
forget gate layer, 2. new memory network, 3. the input gate
layer, and 4. the output gate layer.

2.4 Workflow
The process taking raw well log data from training wells
to predict sonic logs at testing wells using machine
learning models consists of four stages, as shown in
Figure 4. It is worth noting that before training and
applying the trained models to predict sonic logs for new
wells, there are several important steps for data
preprocessing (bad data removal), exploration (additional
features and clustering etc.), and sample preparation (data
scaling and normalization, forming sequence samples and
Fig. 2. Bidirectional long short-term memory (LSTM) based subset partition for training, validation and testing). These
well log prediction. steps ensure the quality of the data samples subsequently
fed into the machine learning models to achieve good
The basic architecture of bidirectional LSTM consists of quality prediction.
two hidden layers of opposite order directions with the
same output, as shown in Figure 2, before they are
combined and fed through an activation layer and a dense
layer to generate sonic log predictions. Structures such as
this aim to learn bidirectional order dependence in
sequence prediction problems. As shown in Figure 3, an
LSTM cell can learn to recognize an important input with
an input gate, store it in the long-term state, and preserve
it for as long as needed, ensure that it is maintained in the
forget gate, and learn to extract it when needed. This
greatly helps the model capture long-term sequence
patterns and has been quite successful in learning time
series, text sequences, and audio clips. It is well suited to
tasks like prediction with sequence data. Because the
properties measured by various well-log techniques are
influenced by properties from the adjacent depth of both
upside and downside, this form of sequence deep-learning Fig. 4. Workflow for deep sequence learning based sonic log
model can be a good fit for the well-log prediction prediction.
problem. Before feeding the log data into the LSTM
model, significant effort was made to remove anomalies The well log data were then formed into sequences of
and choose a good combination of features, as well as samples of the same short length sliding down the depth
experiment with different hyperparameters during model direction, with padding at the boundaries. For data from
training, including input sequence length, number of training wells, we split them into 75% for training and
hidden layers, number of nodes, batch size, dropout 25% for validation, with random partition.
ration, regularization weighting, learning rate, and early
stop (number of epochs). The LSTM model was trained
3. RESULTS AND CONCLUSIONS 𝑅𝑀𝑆𝐸 =

3.1 Field data set ∑ (𝐷𝑇𝐶 , − 𝐷𝑇𝐶 , ) + (𝐷𝑇𝑆 , − 𝐷𝑇𝑆 , ) (1)

The dataset we use comes from the Equinor Volve field 3.3 Results
data, we applied the machine-learning algorithm and used
Following the workflow in Figure 4, we removed samples
the “easy-to-acquire” conventional logs to predict the
with log data values that are either NaN, out of range or
DTC and DTS logs. A total number of 20,525 data points
detected as anomalous, then applied normalization to
(corresponding to distinct depths) collected from three
scale the data values within the same dynamic range.
wells were used to train the models using machine-
After that we clustered samples using minibatch-K-
learning techniques. Each of the data points has seven
Means, based only on the log data that will be used as
features, which are the conventional “easy-to-acquire”
inputs for prediction, excluding the log data to be

Downloaded from http://onepetro.org/armaigs/proceedings-pdf/IGS22/All-IGS22/ARMA-IGS-2022-184/3055616/arma-igs-2022-184.pdf/1 by Nelly Rubio on 02 March 2024

logs, namely caliper, neutron porosity, gamma ray, deep
predicted. The clustering results of all the training
resistivity, medium resistivity, photoelectric factor, and
samples are given in Figure 6, which shows a clear
bulk density, respectively, and two targets, which are the
distribution of several different group patterns.
sonic traveltime logs, DTC and DTS, respectively. We
use neutron porosity, gamma ray and bulk density as
inputs to predict the sonic logs. Figure 5 shows the three
input logs and the measured sonic logs at the training
wells.
The separate data set of 11,089 samples from a fourth well
was then used as the blind test data set.

Fig. 6. clustering of training data based on gamma ray (GR),

density (ZDEN) and neutron porosity (CNC).

After forming short sequence samples from adjacent log

samples and partition the data from training wells into
training and validation subsets, we trained the the bi-
directional LSTM based deep sequence learning network
model given in Figure 2.

Fig. 5. Well logs from the training set (left to right): Gamma
ray, density, porosity and sonic logs (DTC and DTS)
3.2 Performance metrics
Fig. 7. Training results: predicting sonic logs DTC, DTS from
The prediction performance of the model was evaluated gamma ray, bulk density and neutron porosity.
using root mean squared error (RMSE) as the metric,
shown in the equation below:

Fig. 8. Testing results: predicting sonic logs DTC, DTS from
gamma ray, bulk density, and neutron porosity.
Figures 7 and 8 provide the prediction results for the
training and testing well samples, respectively. In each
figure, the top row gives the comparison between the
measured versus the predicted DTC (left) and DTS (right)
data. The bottom row provides the scatter plots for DTC
and DTS prediction. As shown, the testing performance
of the Deep sequence learning network model produced a
robust and accurate sonic log prediction, especially for
DTC. For DTS, a small subset of samples yield prediction
with relatively large variance. In Table 1, we have listed
the RMSE values evaluated for DTC, DTS as well as the
total RMSE error, for both training and testing.
Table 1. Prediction Root Mean Squared Error (RMSE)
Training Testing
DTC 4.2 4.7
DTS 16.4 20.6
Total 12.0 14.9
3.4 Conclusions
In this paper we have developed a deep sequence learning
framework for sonic log prediction. Specifically, the
model predicts the travel times for compressional (DTC)
and shear waves (DTS) from gamma ray, bulk density and
neutron porosity. The bidirectional LSTM based network
model has shown to be capable of capturing the spatial
dependency among the log data and producing accurate
and robust prediction of sonic logs. The proposed data-
driven approach significantly reduces dependency on
prior domain knowledge, provides a generalization
advantage that enables automated large-scale application
of this approach for well log prediction across fields.
REFERENCES
1. Bader, S., X. Wu, and S. Fomel, 2019, Missing log data
interpolation and semiautomatic seismic well ties using
data matching techniques, Interpretation, 7: T347–T361
2. Castagna, J. P., M. L. Batzle, and R. L. Eastwood, 1985,
Relationships between compressional-wave and shear-
wave velocities in clastic silicate rocks, Geophysics, 50:
571–581
3. Chen Y and D. Zhang. 2020, Physics-Constrained Deep
Downloaded from http://onepetro.org/armaigs/proceedings-pdf/IGS22/All-IGS22/ARMA-IGS-2022-184/3055616/arma-igs-2022-184.pdf/1 by Nelly Rubio on 02 March 2024
Learning of Geomechanical Logs, IEEE Transactions on
Geoscience and Remote Sensing, 58:(8), 5932-5943.
4. Ellis D.V. 2007, Well logging for earth scientists,
Dordrecht, The Netherlands Springer.
5. Faust, L. Y., 1953, A velocity function including
lithologic variation, Geophysics, 18, 271–288
6. Feng R, D. Grana and N. Balling. 2021, Imputation of
missing well log data by random forest and its uncertainty
analysis, Computers & Geosciences., 152: 104763.
7. Feng R, D. Grana and N. Balling. 2021, Variational
inference in Bayesian neural network for well-log
prediction, Geophysics,86:(3) M91-M99.
8. Gardner, G., L. Gardner, and A. Gregory, 1974, Formation
velocity and density - The diagnostic basics for
stratigraphic traps, Geophysics, 39: 770-780
9. Glover PWJ. 2014, Petrophysique Course Notes, May,
2014.
10. Greenberg, M. L., and J. P. Castagna, 1992, Shear-wave
velocity estimation in porous rocks: Theoretical
formulation, preliminary verification and applications:
Geophysical Prospecting, 40: 195–209
11. Goodfellow, I., Y. Bengio, and A. Courville, 2016, Deep
learning: MIT Press.
12. He J, H. Li and S. Misra. 2019, Data-Driven In-Situ Sonic-
Log Synthesis in Shale Reservoirs for Geomechanical
Characterization, SPE Reservoir Evaluation &
Engineering., November, 22: 04, 1225-1239.
13. Joshi D, AK Patidar, A. Mishra, et. al. 2021, Prediction of
sonic log and correlation of lithology by comparing
geophysical well log data using machine learning
principles, GeoJournal., September, 2021.
14. Lipton, Z. C., J. Berkowitz, and C. Elkan. 2015. A critical
review of recurrent neural networks for sequence learning.
arXiv preprint arXiv:1506.00019
15. McNally G. H. 1990, The Prediction of Geotechnical
Rock Properties from Sonic and Neutron Logs,
Exploration Geophysics., 21: 1-2, 65-71.
16. Onalo D, S. Adedigba, F. Khan, LA. James and S. Butt.
2018. Data driven model for sonic well log prediction,
Journal of Petroleum Science and Engineering, 170:
1022-1037.
17. Pham N, X. Wu and NE. Zabihi. 2020, Missing well log
prediction using convolutional long short-term memory
network, Geophysics, 85:(4), WA159-WA171.
18. Rolon, L., S. D. Mohaghegh, S. Ameri, R. Gaskari, and B.
McDaniel, 2009, Using artificial neural networks to
generate synthetic well logs, Journal of Natural Gas
Science and Engineering, 1: 118–133
19. Saggaf, M., and L. Nebrija, 2003, Estimation of missing
logs by regularized neural networks, The American
Association of Petroleum Geologists Bulletin 87: 1377–
1389

Downloaded from http://onepetro.org/armaigs/proceedings-pdf/IGS22/All-IGS22/ARMA-IGS-2022-184/3055616/arma-igs-2022-184.pdf/1 by Nelly Rubio on 02 March 2024

20. Salehi, M. M., M. Rahmati, M. Karimnezhad, and P.
Omidvar, 2017, Estimation of the non records logs from
existing logs using artificial neural networks, Egyptian
Journal of Petroleum, 26: 957–968
21. Smith, J. H., 2007, A method for calculating pseudo sonics
from e-logs in a clastic geologic setting: Gulf Coast
Association of Geological Societies Transactions, 57:
675–678
22. Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. 2014,
Sequence to sequence learning with neural networks.
Advances in neural information processing systems 27
23. Wu L, Z. Dong, W. Li, C. Jing and B. Qu. 2021. Well-
Logging Prediction Based on Hybrid Neural Network
Model, Energies, 14:(24), 8583.
24. Yu Y, C. Xu, S. Misra, W. Li and M. Ashby. et al. 2021,
Synthetic Sonic Log Generation With Machine Learning:
A Contest Summary From Five Methods, Petrophysics –
The SPWLA Journal of Formation Evaluation and
Reservoir Description, 62:(4), 393-406.
25. Zhang, D., Y. Chen, and J. Meng, 2018, Synthetic well
logs generation via recurrent neural networks, Petroleum
Exploration and Development, 45:629–639
26. Zhang, A. et al. 2021, Dive into deep learning. arXiv
preprint arXiv:2106.11342