geosciences

Article
Data-Driven Geothermal Reservoir Modeling: Estimating
Permeability Distributions by Machine Learning
Anna Suzuki 1, * , Ken-ichi Fukui 2 , Shinya Onodera 3 , Junichi Ishizaki 3 and Toshiyuki Hashida 4

1 Institute of Fluid Science, Tohoku University, 2-1-1 Katahira, Aoba-ku, Sendai 980-8577, Japan
2 Department of Architecture for Intelligence, Osaka University, Osaka 567-0047, Japan;
[email protected]
3 Tohoku Electric Power Co., Inc., Sendai 980-8550, Japan; [email protected] (S.O.);
[email protected] (J.I.)
4 Fracture and Reliability Research Institute, Tohoku University, Sendai 980-8579, Japan;
[email protected]
* Correspondence: [email protected]; Tel.: +81-22-217-5284

Abstract: Numerical modeling for geothermal reservoir engineering is a crucial process to evaluate
the performance of the reservoir and to develop strategies for the future development. The governing
equations in the geothermal reservoir models consist of several constitutive parameters, and each
parameter is given to a large number of simulation grids. Thus, the combinations of parameters
we need to estimate are almost limitless. Although several inverse analysis algorithms have been
developed, determining the constitutive parameters in the reservoir model is still a matter of trial-
and-error estimation in actual practice, and is largely based on the experience of the analyst. There are
several parameters which control the hydrothermal processes in the geothermal reservoir modeling.
In this study, as an initial challenge, we focus on permeability, which is one of the most important



parameters for the modeling. We propose a machine-learning-based method to estimate permeability


distributions using measurable data. A large number of learning data were prepared by a geothermal
Citation: Suzuki, A.; Fukui, K.-i.;
Onodera, S.; Ishizaki, J.; Hashida, T.
reservoir simulator capable of calculating pressure and temperature distributions in the natural state
Data-Driven Geothermal Reservoir with different permeability distributions. Several machine learning algorithms (i.e., linear regression,
Modeling: Estimating Permeability ridge regression, Lasso regression, support vector regression (SVR), multilayer perceptron (MLP),
Distributions by Machine Learning. random forest, gradient boosting, and the k-nearest neighbor algorithm) were applied to learn the
Geosciences 2022, 12, 130. https:// relationship between the permeability and the pressure and temperature distributions. By comparing
doi.org/10.3390/geosciences12030130 the feature importance and the scores of estimations, random forest using pressure differences as
Academic Editors: Tobias M. Müller
feature variables provided the best estimation (the training score of 0.979 and the test score of 0.789).
and Jesus Martinez-Frias Since it was learned independently of the grids and locations, this model is expected to be generalized.
It was also found that estimation is possible to some extent, even for different heat source conditions.
Received: 4 February 2022
This study is a successful demonstration of the first step in achieving the goal of new data-driven
Accepted: 7 March 2022
geothermal reservoir engineering, which will be developed and enhanced with the knowledge of
Published: 11 March 2022
information science.
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in Keywords: geothermal reservoir modeling; TOUGH2; inverse analysis; natural state
published maps and institutional affil-
iations.

1. Introduction
Copyright: © 2022 by the authors.
Geothermal reservoir modeling is a crucial process in geothermal developments.
Licensee MDPI, Basel, Switzerland. Reservoir simulation needs correctly constructed governing equations to obtain proper
This article is an open access article numerical solutions of the multiphase fluid and heat flow processes. There have been
distributed under the terms and several numerical simulators developed (e.g., TOUGH2 [1], TETRAD [2], STAR [3], SHE-
conditions of the Creative Commons MAT [4], MODFLOW [5], and COMSOL [6]) to evaluate the performance of reservoirs,
Attribution (CC BY) license (https:// which provides a basis for planning for future developments.
creativecommons.org/licenses/by/ Numerical reservoir modeling can be basically divided into two types. One approach
4.0/). provides a natural-state simulation, and the other can be described as a history simulation [7,8].

Geosciences 2022, 12, 130. https://doi.org/10.3390/geosciences12030130 https://www.mdpi.com/journal/geosciences

Geosciences 2022, 12, 130 2 of 18

Most modelers carry out natural-state simulations at least as the first step in constructing a
numerical model, which simulates preproduction reservoir conditions [9,10]. Natural-state
modeling is usually based on conceptual models, which are based on data obtained from
geological and geophysical surveys [11] and the chosen input parameters (e.g., rock proper-
ties, fluid properties, boundary conditions, initial conditions). By performing long-time
simulations (over several thousand years), the quasi-steady state temperature and pressure
fields can be obtained. The process of substituting the input parameters in the simulation,
solving the governing equations, and generating the conditional variables (i.e., temperature
and pressure fields) is referred to as a forward analysis. Forward modeling requires an
understand of all of the input parameters on each grid. Because insufficient information is
obtained from geological and geophysical surveys for determining the entire structure of
the subsurface, it is necessary to estimate some of the input parameters.
The input parameters need to be adjusted and optimized by fitting between the
simulation results and the observable data in a process known as inverse modeling [12].
The input parameters of geothermal reservoir models consist of the rock properties (e.g.,
permeability, porosity, thermal conductivity), boundary conditions (e.g., the amounts and
locations of heat sources and sinks), and the initial conditions (e.g., temperature and
pressure). The main targets in the inverse analysis in geothermal reservoir modeling are
permeability and the conditions of the heat source/sink because they have significant
impacts on the simulation results. It is possible in some cases to determine the amount and
the location of heat sources and sinks using a conceptual model based on geological and
geophysical exploration [13]. On the other hand, permeability information is somehow
related to resistivity and miscroseismic data for the detection of low-resistive zones or faults.
Since these data are indirect and not directly involved in the flow, it is difficult to obtain
a unique solution using these geophysical data only. Thus, the input parameters of the
conventional approaches need to be optimized by repeating the numerical simulations until
the differences between the observed data of temperature and pressure and the simulation
results become acceptably small.
Several inverse modeling codes have been developed, such as iTOUGH2 [14],
UCODE [15], and PEST [16], to assist with selections of the input parameters and
the evaluations of the sensitivity of the parameters. These techniques contribute by
automating the time-consuming process of optimization and by minimizing the bias
of the modelers in parameter selection. However, their calculations still require the
iteration of the simulation to optimize the parameters for each of the tens of thousands of
grids. Although some progress has been made in developing efficient inverse methods
(e.g., [17]), the methods still tend to require a great deal of time and effort. The only
exception is when the modelers are in a position to perform parallel calculations on a
computer with good specs.
With artificial intelligence (AI) advancing in leaps and bounds, machine-learning-
based approaches have been applied in many fields. Machine learning is an iterative
learning process that uses multiple data to allow the computer to find the underlying
patterns in the data inductively. The processes of machine learning appears to be highly
compatible with the requirements of the inverse analysis in reservoir modeling.
In geothermal fields, machine learning and deep learning algorithms have been
used to achieve a variety of purposes. Assouline et al. [18] estimated a temperature
map at shallow depths by the supervised method. To estimate deep temperature fields,
Spichak et al. [19] and Ishitsuka et al. [20] used neural networks based on resistivity
data, and they showed that the use of machine learning algorithms has led to an improve-
ment in the accuracy of estimates. Rezvanbehbahani et al. [21] estimated geothermal
heat flux using a large collection of relevant geologic features and global measurements.
Siler et al. [22] and Gudmundsdottir and Horne [23] used the unsupervised method to
identify key factors of geothermal production using a geologic dataset and a tracer response
data, respectively. Holtzman et al. [24], Gao et al. [25], and Zheng et al. [26] used unsuper-
vised methods and neural networks to characterize faults and fractures from microseismic
Geosciences 2022, 12, 130 3 of 18

data. The development of these machine learning models will help to create structural
models of the subsurface. However, it is not yet possible to directly estimate permeability,
which is an important input parameter in geothermal reservoir simulations. In several
studies, machine learning algorithms have been used to predict core permeability from well
log data [27–29] or core samples [30–32]. Similarly, Al-Anazi and Gates [33] predicted core
porosity from well log data. These studies estimated only a limited permeability field along
wells. In order to perform numerical simulations, it is necessary to estimate the spatial
distributions of permeability for an entire reservoir. Efforts are now underway to predict
the flow by using convolutional neural network and deep learning methods in petroleum
fields [34–38]: these studies are based on the input data available in oil development and,
while similar to context of geothermal development, are not simply transferable.
In this study, we demonstrated the first step towards approaching the goal of proposing
a new data-driven geothermal reservoir modeling to estimate permeability distributions for
natural-state simulations. In the natural-state simulation, temperature and pressure in the
quasi-steady state are the simulation output: that is, they are the solutions of mass balance
and energy balance equations. Since it is possible to determine the conditions of heat
sources and sinks from the conceptual model [13], we treated the conditions of heat sources
and sinks as known. In conventional reservoir modeling, the pressure and temperature
in the quasi-steady state are determined by substituting the input parameters. Thus, the
pressure and temperature are expressed as a function of the input parameters in the forward
modeling. To minimize the error between the numerical outputs and the observed data, the
input parameters are adjusted over and over repeatedly. This process is the conventional
inverse modeling. In contrast, the new approach we propose in this study aims to construct
a machine-learning model that captures the permeability by substituting measured data.
Since the permeability is a function of the observed data, it may be possible to derive the
permeability in a single simulation, without the requirement for many iterations.
In this paper, by using the numerical simulator TOUGH2 [1] to generate a large
number of learning datasets, we have developed a machine learning model with several
supervised algorithms. The estimation accuracy was compared to the test dataset, which
had different permeability distributions to that of the training dataset. The applicability of
the learning model to further data-driven developments in geothermal engineering is then
discussed. It should be noted that our analysis is limited to 2D thermohydraulic simulation
as a first step to develop the machine learning approach and that 3D thermohydraulic–
mechanical–chemical simulations are needed to make it available for real field development
in future research.

2. Method
2.1. Preparation of Learning Data
In this study, we applied several popular machine-learning algorithms to estimate
permeability distributions using measurable data in geothermal reservoir modeling. Here,
we assume that two-dimensional temperature and pressure distributions of the area could
be obtained from the temperature and pressure measured in multiple wells by using kriging
or other methods.
First, we prepared large learning datasets from a numerical reservoir simulator
TOUGH2, which simulates fluid and heat flow using the finite volume method [1]. Two-
dimensional synthetic models were prepared with the simulation domain shown in Figure 1.
The simulation area was 2000 m × 2000 m by 30 grids × 30 grids. The grids were discretized
at 100 m in the center (20 grids × 20 grids) and at 50 m in the surrounding area. The top
boundary was open boundary with temperature of 25 ◦ C and pressure of 0.1 MPa. The
bottom and side boundaries were no flow condition except grids with sources and sinks.
The heat source was located at the bottom left, and the sink was located at the right side, as
shown in Figure 1. The simulation domain consisted of three areas: the surrounding rocks,
the reservoir, and the flow channels. The heat source was connected to the reservoir area
by the flow channels. The mass flow rate at the heat source was set to 0.12 kg/s, and the
Geosciences 2022, 12, 130 4 of 18

flowing specific enthalpy was 1085 kJ/kg (250 ◦ C for saturated water). The mass flow rate
at the sink was set at 0.12 kg/s. The other input parameters for the simulation are listed in
Table 1. It should be noted that porosity was set to constant because the effect of porosity is
small on natural state simulation.

(a) constant 0.1MPa

25 oC (b) reservoir permeability [m2]
-11
(i) 10

channel
surrounding rock

flow
-14
10
-17
flow sink 10
channel flow rate: (ii)
2000m

0.12 kg/s
reservoir

(iii)
channel
flow

.
source .
flow rate: 0.12 kg/s .
enthalpy: 1085kJ/kg
2000m

Figure 1. Simulation condition. (a) Simulation area with boundary conditions. (b) Examples of
permeability distribution patterns in the reservoir area. The patterns were generated using a discrete
cosine transformation. (b) (i–iii) depicts three different patterns.

Table 1. Input parameters in TOUGH2.

Methods Parameters SI Unit

Rock density 2250 kg/m3
Porosity 0.1 -
Thermal conductivity 2.5 W/m◦ C
Specific heat 1000 J/kg◦ C

To prepare a large number of training data, different permeability patterns in the

reservoir area were generated. The patterns were generated based on the discrete cosine
transform, which is a basic image generation method used in image analysis [39]. The
two-dimensional discrete cosine transform is given by

N1 −1 N2 −1    !    
π 1 π 1
Xk1 ,k2 = ∑ ∑ 1 2
x n ,n cos n 2 + k 2 cos n 1 + k
n1 =0 n2 =0
N2 2 N1 2 1
N1 −1 N2 −1        
(1)
π 1 π 1
= ∑ ∑ xn1 ,n2 cos n1 + k cos n2 + k2
n =0 n =0
N1 2 1 N2 2
1 2

fork1 = 0, . . . , N1 − 1 andk2 = 0, . . . , N2 − 1.

where X is the image matrix of size N2 by N1 , and Xk1 ,k2 is the matrix element in X. The real
numbers x0,0 , . . . , x N1 −1,N2 −1 are transformed into the real numbers X0,0 , . . . , X N1 −1,N2 −1 .
Examples of generated permeability distributions based on the discrete cosine transform are
shown in Figure 1b. Strip and lattice shapes can be seen. Note that the standard value of the
permeability in the reservoir area was set to 10−15 m2 , the permeability in the surrounding
rocks was set to 10−18 m2 , and the permeability in the flow channels was set to 10−15 m2 .
Here, some of the permeability distributions from the discrete cosine transform appear
Geosciences 2022, 12, 130 5 of 18

to be far from actual geological formations, but the structures close to the actual geology,
such as layer formations, are also included in the training data. In particular, the machine
learning method used in this project, as we explain later, does not grasp the overall trend
of the distribution, but estimates the values based on local information. Therefore, even if
some of the permeability distributions from discrete cosine transform seem geometric and
unrealistic, we think the discrete cosine transform is fine for this project.
Two hundred permeability patterns were generated for each simulation. Each simula-
tion was run for 1014 s to reach the quasi-steady state, which is regarded as the natural state
of geothermal reservoirs. Some of the simulations stopped before reaching the quasi-steady
state. We used 180 simulation results of the simulations which continued until the end as
the learning data.

2.2. Development of Machine Learning Model

The combination of permeability, temperature, and pressure distributions for 180 sim-
ulations was used for developing learning models. The variable being predicted is referred
to as the “output” or “target”, while the input variables are referred to as “features”. In
this study, the values of permeability in the reservoir domain were set as the target vari-
ables. The feature variables were given based on the values of the simulation outputs (i.e.,
temperature and pressure).
A regression analysis is a statistical method for modeling the relationship between
targets and feature variables. Among the various types of regression algorithms for super-
vised machine learning, we applied the following in this study: a linear regression, a ridge
regression, a Lasso regression, a support vector regression (SVR), a multilayer perceptron
(MLP), a random forest, a gradient boosting, and a k-nearest neighbor algorithm. We used
Python packages sklearn (v1.0.2), numpy (v1.22.1), optuna (v2.10.0), lightgbm (v3.3.2), and
matplotlib (v3.0.3) in Python (v3.7.12).
The ridge regression and the Lasso regression, which are among the most robust
versions of linear regression, introduce regularization techniques to reduce the complexity
of the model [40,41]. The support vector machine (SVM) is a popular supervised machine
learning algorithm and is representative of nonparametric machine learning methods [42].
We implemented the support vector regression (SVR) with the kernel functions of the linear,
polynomial, and radial basis function (rbf). The multilayer perceptron (MLP) refers to a
neural network of multiple formal neurons connected in multiple layers [43]. In the case of
the random forest, a large number of decision trees are created by random sampling that
allows for duplication, and the final prediction is determined by taking a majority vote of
the prediction results obtained for each tree [44]. Gradient boosting continuously modifies
and adds predictors to the ensemble, and also modifies the predictor to fit the residual
error [45]. The k-nearest neighbor algorithm is a method based on the nearest training
examples in the feature space [46]. We imported their modules from scikit-learn [47].
Two different approaches to building the models were applied. The first was to build a
model with dependence on the grid. Each learning model was developed on each grid, and
there were 400 learning models for each grid in the reservoir domain (20 grids × 20 grids).
This type of model is referred as grid-dependent. The second was to build a model with no
dependence on the grid. A single learning model was developed using data from all grids.
This type of model is referred as grid-independent.
For both the grid-dependent model and grid-independent model, we first arranged
the simulation results in a random order to separate the simulation results into a learning
dataset and test dataset. The dataset of the first 70% of simulations was used as the training
data, and the remaining 30% was used as the test data. The results of 180 simulations were
divided into 126 simulations for use as the training data and 54 simulations for use as the
test data.
The training dataset was normalized to adjust a wide numeric range of input variables
to the range of [0, 1] using the minimum and maximum values of each feature variable. The
normalized data was used to construct the learning model. Except for the linear regression,
Geosciences 2022, 12, 130 6 of 18

it was necessary to tune the hyperparameters. We used an automatic hyperparameter

optimization software framework, Optuna [48], which is designed to automatically and
efficiently search for optimal hyperparameters in large spaces. We performed a three-fold
cross validation search for the hyperparameters. The hyperparameters and the range tuned
in this study are listed in Table 2. The performance of the prediction model was scored by
the coefficient of determination (R2 ).

Table 2. Hyperparameters in each model.

Methods Parameters Ranges

Linear - -
Ridge α 0.00001–100
Lasso α 0.00001–100
max iteration 100,000
SVR (linear) C 0.01–10,000
SVR (polynomial) C 0.01–10,000
degree 2–4
SVR (rbf) C 0.01–10,000
γ 0.0001–100
e 0.0001–0.01
MLP solver sgd, adam, lbfgs
activation identity, logistic,
relu, tanh
max layer size 50–300
α 0.001–1000
Random forest number of trees in the forest 100–1000
Gradient boosting number of boosting stages to perform 100–1000
maximum depth 3
k-nearest neighbors number of neighbors 3–7

After developing a learning model based on the training set, the learning model was
evaluated with the test dataset. The test dataset was normalized with the scaling equation
according to the attribute range of the training dataset with the best hyperparameters tuned
with the training dataset.

3. Results
3.1. Model Selection
Selecting the feature variables is one of the core concepts in machine learning and
has a large impact on the performance of the learning model [49]. We prepared 18 feature
variables based on the temperature and pressure data given in Figure 2 to estimate the
permeability of a grid point (Ki,j ) (Figure 2a). These features were the temperature of
the point to be estimated (Ti,j ) (Figure 2b), the pressure of the point to be estimated (Pi,j )
(Figure 2c), and the temperature and pressure of the points adjacent to the point to be
estimated in the x- and y-directions (Ti−1,j , Ti+1,j , Ti,j−1 , Ti,j+1 , Pi−1,j , Pi+1,j , Pi,j−1 , Pi,j+1 ). In
addition, we used the spatial differences in the temperature and pressure between the
point to be estimated and the points adjacent to the point to be estimated in the x- and
y-directions, which are denoted as ∆Ti−1,j , ∆Ti+1,j , ∆Ti,j−1 , ∆Ti,j+1 , ∆Pi−1,j , ∆Pi+1,j , ∆Pi,j−1 ,
∆Pi,j+1 , as shown in Figure 2b,c. The grid numbers were assigned starting from the lower
left of the computational domain, as shown in Figure 1a. Since the heat source was set to
the lower left, the smaller grid number is considered to be upstream, and the larger grid
number is considered to be downstream.
Geosciences 2022, 12, 130 7 of 18

(a) y (b) y (c) y

Ti, j+1 Pi, j+1

Ki, j Ti-1, j T Ti+1, j Pi-1, j P Pi+1, j

i, j i, j

Ti, j-1 Pi, j-1

x x x
ΔTi+1, j =Ti+1, j-Ti, j ΔPi+1, j =Pi+1, j-Pi, j
ΔTi-1, j =Ti, j-Ti-1, j ΔPi-1, j =Pi, j-Pi-1, j
ΔTi, j+1 =Ti, j+1-Ti, j ΔPi, j+1 =Pi, j+1-Pi, j
ΔTi, j-1 =Ti, j-Ti, j-1 ΔPi, j-1 =Pi, j-Pi, j-1

(d)
0.6
Feature importances

0.4

0.2

Ti,j Pi,j Ti-1,j Ti+1,j Ti,j-1 Ti,j+1 Pi-1,j Pi+1,j Pi,j-1 Pi,j+1 ΔTi+1,j ΔTi,j+1 ΔPi+1,j ΔPi,j+1
ΔTi-1,j ΔTi,j-1 ΔPi-1,j ΔPi,j-1
Features
(e)
Feature importances

0.3

0.2

0.1

0
Ti,j Pi,j Ti-1,j Ti+1,j Ti,j-1 Ti,j+1 Pi-1,j Pi+1,j Pi,j-1 Pi,j+1 ΔTi+1,j ΔTi,j+1 ΔPi+1,j ΔPi,j+1
ΔTi-1,j ΔTi,j-1 ΔPi-1,j ΔPi,j-1
Features

Figure 2. (a–c) Setting of feature variables: (a) permeability Ki,j , (b) temperature Ti,j , and (c) pressure
Pi,j for grid x = i and y = j. (d,e) Results of relative importance of feature variables based on training
data with random forest for (d) grid-dependent models and (e) grid-independent model.

The importance of features is a measure of the extent to which feature partitioning

contributes to the regression of the target. We calculated the impurity-based feature im-
portances from scikit-learn modules with random forest [44]. The importances of feature
variables for the grid-dependent models and for the grid-independent model are plotted in
Figure 2d,e, respectively. For grid-dependent models, we plot their mean values and the
standard deviation with error bars (Figure 2d). As shown in Figure 2d, the feature impor-
tance of the pressure difference with downwind in the x- and y-directions (∆Pi+1,j , ∆Pi,j+1 )
was higher in both the grid-dependent models and the grid-independent model (Figure 2e).
It is important to note that rather than reflect the intrinsic predictive value of a partic-
ular feature by itself, the impurity-based feature importance indicates the importance of
Geosciences 2022, 12, 130 8 of 18

this feature for a particular model. In other words, the results obtained with random forest
may not be applicable to other machine learning models. Nevertheless, it was clear that the
pressure differences were more important than the other feature variables, as can be seen
in Figure 2. In addition, when we consider the physical meaning of the feature variables,
since permeability was used in the TOUGH2 with the Darcy law along with the pressure
gradients [1], it is understandable that the pressure differences affect the flow conditions
and can be more important than the other feature variables in estimating permeability.
The statistic values of pressure differences for each grid are shown in Figure 3.
Figure 3a,b show the mean and standard deviation of pressure differences between adja-
cent grids in the x-direction, while Figure 3c,d show the mean and standard deviation of
pressure differences between adjacent grids in the y-direction. As we can see, the values in
the lower left and upper right of the region are larger or smaller than the others. The flow
channels connecting to the heat source and sink were located near the bottom left and top
right of the region. Because the flow channels were near the heat source and sink, the flow
movement was rapid in these areas and the pressure difference was larger. Note also that
trends in the results of the pressure differences upwind are similar to those downwind.

(a) Mean ΔPi+1,j (b) std ΔPi+1,j

[MPa]
0.2
(c) Mean ΔPi,j+1 (d) std ΔPi,j+1 0

-0.2

Figure 3. Spatial distributions of mean and standard deviation of pressure differences with downwind
for each grid. (a,c) The mean and (b,d) standard deviation (std) of ∆Pi+1,j and ∆Pi,j+1 .

For both the grid-dependent and grid-independent models, the pressure differences
downwind were more important than those upwind. Here, we observed that the accuracy
of the model estimation was improved by using both the upwind and downwind pressure
difference rather than only using the downwind pressure difference. When we prepare
the pressure data, it is always possible to obtain the values of the downwind and upwind
pressure differences. Thus, in the following results, we used the four pressure differences
(∆Pi−1,j , ∆Pi+1,j , ∆Pi,j−1 , ∆Pi,j+1 ) as feature variables to build the learning model.
The different machine learning algorithms for the grid-dependent models and the grid-
independent model were compared. We calculated the estimation scores as the coefficient
of determination (R2 ). The results of the scores are plotted in Figure 4a. Since there were
400 results for the grid-dependent models, the means and the standard deviation with the
error bars were plotted for the grid-dependent models, as shown in Figure 4a.
Geosciences 2022, 12, 130 9 of 18

(a)

1.0
training test
grid
-dependent
0.8 -independent

Score 0.6

0.4

0.2

-0.2

rs
g
st
LP

in
f)
ge

o
ar

bo
r)

l)

(rb
ss

re
ea

ia

st
ne

gh
id

M
om
La

fo

oo
(lin
R

R
Li

ei
SV

m
lyn

tb

tn
R

do
SV

es
(p

en
an

ar
R

di
SV

ne
R

ra

k-
G
(b) Training (c) Test

Lasso

R2
1
0
((d)) Training
g (e) Test
-1

MLP

Figure 4. (a) Comparison of scores (coefficient of determination: R2 ) for different methods. (b) Train-
ing and (c) test scores of Lasso model and (d) Training and (e) test scores of MLP model.

The linear, ridge, and Lasso models are based on a linear model. The obtained scores
were lower than others (Figure 4a). The accuracy of the grid-dependent models based on
their training data is better than that of the grid-independent model. The more diverse
relationships between the feature and the target at different locations with different flow
patterns in the independent model are more difficult to characterize in the linear models.
This may explain the lower accuracy of the grid-independent model than that of the
dependent model. The scores of training and test dataset using grid-independent model
with Lasso, as an example of linear model, are shown in Figure 4b,c. The accuracy of the
training data near the upper right was poor. In the upper right corner, water flowed out
to the top and right boundaries due to the boundary setting in this study. This resulted in
larger or smaller pressure differences, as shown in Figure 3a,c. It is expected that the flow
movement was steeper than in other areas and that the linear model could not capture such
different flow behaviors. In addition, the accuracy of the test data was poorer in the lower
right areas. Since the pattern of the permeability distribution given by the discrete cosine
Geosciences 2022, 12, 130 10 of 18

transform appears to be reflected in the obtained scores, it is likely that the permeability
distribution also affects the estimation accuracy. These results confirm that the simple linear
model was not capable of dealing with the differences in flow patterns and differences in
permeability distribution given by the discrete cosine transform.
Because SVR and MLP are nonlinear models, they are expected to be able to represent
more complex relationships than the linear models. As expected, the training data for
the grid-dependent model provided better scores. However, the test data for the grid-
dependent SVR model with polynomial function showed very poor scores. When we
observed the results of the scores in the spatial distribution, grids with poor prediction
accuracy appeared randomly, which suggests that the heterogeneous permeability distribu-
tions given by the discrete cosine transform in the training data affected the accuracy of the
learning model. There is a possibility that the accuracy can be improved by increasing the
number of training data, but for the sake of comparison with other models, we used the
same amount of training data to obtain the results in this study.
Although the scores of the nonlinear grid-independent models (SVR and MLP) were
better than those of the linear models, the values of the test data stopped at about 0.5. The
scores of the training and test dataset using the grid-independent model with MLP, as
an example of a nonlinear model, are shown in Figure 4d,e. The accuracy of the training
data near the upper right was poor, and was similar to the score obtained by the Lasso
model. Nonlinear models do not characterize the features well with discontinuities in the
mapping function. The large difference in the pressure differences between the center area
and the upper right area suggests that the mapping function with the feature variables
may be discontinuous. This may explain the poor results obtained by the test data in the
grid-dependent model for the nonlinear models.
The scores of random forest and gradient boosting were almost 1.0 for the training
dataset and around 0.8 for the test dataset. Random forest and gradient boosting are based
on ensembles (approximation with multiple functions), and may be suitable even in cases
where the mapping function is discontinuous.
In random forest and gradient boosting, although the performance of the grid-dependent
models was better than the grid-independent model, the difference was small. Because it is
easier to prepare the training data in the grid-independent model, it may be able to be used in
different model settings, such as different positions of the source and sink in the boundary
conditions. We adopted the grid-independent model with a random forest algorithm in
this study.

3.2. Estimation of Permeability Distributions

The results of the permeability estimation for the training dataset using the the grid-
independent model with the random forest algorithm are shown in Figure 5. Among the
results obtained for 126 simulations using the training dataset, we show two examples
of permeability patterns (Figure 5a,c) and their estimation results (Figure 5b,d), which
were selected randomly from among all the simulation results. Figure 5e plots the spatial
distribution of the mean of the squared error (MSE) between the expected values and the
estimated values for each grid for the training dataset. In Figure 5f, the spatial distribution
of the coefficient of determination (R2 ) is shown for the expected and estimated values for
each grid. There was almost no error, and the expected and estimated values are highly
correlated. The score was calculated by using the coefficient of determination (R2 ). The
score for the training dataset was 0.979, as listed in Table 3.
Geosciences 2022, 12, 130 11 of 18

(a) Expected (b) Estimated

training
case 1

Permeability [m2]
-11
10
-14

(c) Expected (d) Estimated 10

-17
10

training
case 2

(e) MSE (f) R2

2 1

1 0.5

0 0

Figure 5. Results of permeability estimation for training dataset. Examples of (a,c) expected distribu-
tions and (b,d) estimated results. Spatial error distributions with (e) MSE and (f) R2 .

Table 3. Scores for training and test datasets. * Represents the same conditions as the training datasets.

Conditions
Mass Flow Rate (kg/s) Position Score (R2 )
Training 0.12 left 0.979
0.12 * left * 0.789
0.04 left * 0.715
Test 0.4 left * 0.768
0.12 * center 0.576
0.12 * right 0.450

The results of the estimation for the test dataset are shown in Figure 6 for two examples
of permeability patterns (Figure 6a,c) and their estimation results (Figure 6b,d). Although
some parts were not perfectly estimated, the estimation captured most of the trends in the
permeability distribution. This indicates that this learning model can be applied to the
test dataset.
Figure 6e,f are plots of the spatial distributions of the mean of the squared error and
2
R between the expected values and the estimated values for each grid for test dataset. As
shown in Figure 6e,f, near the center of the domain, the error was small and the expected
and estimated values were highly correlated. On the other hand, the estimation accuracy at
the top and left edges appeared to be degraded. Since the heat source was placed in the
bottom left corner, the estimation of the area distant from the heat source may have been
compromised. The score was 0.789.
Geosciences 2022, 12, 130 12 of 18

(a) Expected (b) Estimated

test
case 1

Permeability [m2]
-11
10
-14
10
(c) Expected (d) Estimated
-17
10
test
case 2

(e) MSE (f) R2

2 1

1 0.5

0 0

Figure 6. Results of permeability estimation for test dataset. Examples of (a,c) expected distribu-
tions and (b,d) estimated results and spatial error distributions by (e) MSE and (f) coefficient of
determination R2 .

3.3. Estimation for Different Heat Source Conditions

In the previous subsection, it was shown that the grid-independent model with
random forest algorithm can estimate permeability distributions for the test dataset. The
above test data were obtained from a reservoir model in which only the permeability
distribution was changed from the training data. When preparing a reservoir model, in
addition to the permeability distribution, settings of the conditions of the heat source and
sink may also have large impacts on the simulation results. Therefore, in this subsection,
we examine whether the training model can be applied to a test dataset generated under
different source and sink conditions.
First, the effects of the magnitude of the mass flow rate at the source were investigated.
The simulations were performed with different patterns of permeability and varied mass
flow rates of 0.04, 0.12, and 0.4 kg/s, respectively. The simulation results at the mass flow
rate of 0.12 kg/s were used as the training data. These results are the same as those shown
in Figure 6. In order to compare the results with the test data at the mass flow rate of
0.12 kg/s, 30% of the total data was also used as the test data for the mass flow rates of 0.04
and 0.4 kg/s.
Figure 7 shows the estimation results for test dataset prepared using different source
mass flow rates. Figure 7a,b show plots of an example of permeability distribution and
their estimation results for mass flow rate of 0.04 kg/s, and Figure 7c,d are plots of the
spatial distributions of the mean of the squared error and R2 between the expected values
and the estimated values for each grid. Figure 7e–h show the results for the test case for
mass flow rate of 0.4 kg/s. In the case of the mass flow rate of 0.04 kg/s, the errors at the
bottom left corner were relatively high, and the errors at the mass flow rate of 0.4 kg/s
were relatively high at the bottom right corner. Nevertheless, we can see that the middle
of the reservoir region was estimated with good accuracy. The scores for the test dataset
with mass flow rate of 0.04 kg/s and 0.4 kg/s were 0.715 and 0.768, respectively, as listed in
Table 3. There was little difference between those scores and the score for the test dataset
with the same conditions as the training dataset (R2 = 0.789). This suggests that the learning
model can be applied to reservoir models with different mass flow rates to the training data.
Geosciences 2022, 12, 130 13 of 18

(a) Expected (b) Estimated

Permeability [m2]
-11
10
test 10
-14

(0.04 kg/s)
-17
10

(c) MSE (d) R2

2 1

1 0.5

0 0

(e) Expected (f) Estimated

Permeability [m2]
-11

Permeability [m2]
10
-14
test 10
(0.4 kg/s)
-17
10

(g) MSE (h) R2

2 1

1 0.5

0 0

Figure 7. Estimation accuracy in the case of different mass flow rates at the source from the training
data. (a) Expected vs. (b) estimated values, and spatial distributions of (c) MSE and (d) R2 for the
mass flow rate of 0.04 kg/s. (e) Expected vs. (f) estimated values, and spatial distributions of (g) MSE
and (h) R2 for the mass flow rate of 0.4 kg/s.

Next, the effect of positions of the heat source was investigated. The simulations
were performed with different patterns of permeability with different positions of heat
sources. The learning dataset was generated by setting the heat source on the bottom left
of the reservoir, as shown in Figure 1a. Here, we prepared the test dataset with the heat
source positioned at the bottom center and the bottom right, as shown in Figure 8a,b. The
simulation results with the source locating at the bottom left were used as the training data,
while the simulation results with source locating at the bottom left, bottom center, and
bottom right were used as the test data.
(a) (b)

surrounding rock surrounding rock

sink sink

reservoir reservoir

source source

Figure 8. Setting of locations of heat sources for test data.The positions of heat sources are set to
(a) bottom center and (b) bottom right.
Geosciences 2022, 12, 130 14 of 18

Figure 9 show the estimation results for the test dataset prepared using different
locations of heat source. Figure 9a,b show plots of an example of permeability distribution
and their estimation results when the heat source was located on the center bottom, and
Figure 9c,d are plots of the spatial distributions of the mean of the squared error and R2
between the expected values and the estimated values for each grid. Figure 9e–h show the
results for the test case when the heat source was located on the right bottom. It can be
observed that as the heat source shifted to the right, the estimation accuracy of the left side
of the reservoir region decreased. The scores for the test dataset where the source position
was center and right were 0.576 and 0.450, respectively, as listed in Table 3. The more the
source position deviated from the training data, the worse the test data score became. Since
the sink was placed at the upper right of the reservoir, the main water flow was from the
bottom left (the heat source) to the top right (the sink). When the heat source was located
at the bottom right, the left side of the reservoir region is considered to have had less water
movement. In this case, even if the permeability was high, the lack of water movement
would lead to underestimations of permeability. The low accuracy of the estimation at the
left side of the reservoir area may be attributed to this lack of water movement.
We have shown that the learning model developed in this study can be used to estimate
most of the trends of permeability, indicating that the estimation works well even when
for different positions of sources. In order to further improve the accuracy, we intend to
develop a learning model that can estimate the source and sink conditions in future research.
For application to objects with different conditions, the application of data augmentation
and domain adaptation need to be considered as future tasks. Both of these have been
shown to be effective in deep learning [50,51].

(a) Expected (b) Estimated

Permeability [m2]

-11
10
Permeability [m2]

-14
test 10
(center) -17
10

(c) MSE (d) R2

2 1

1 0.5

0 0

(e) Expected (f) Estimated

Permeability [m2]

-11
10
-14
test 10
(right)
-17
10

(g) MSE (h) R2

2 1

1 0.5

0 0

Figure 9. Estimation accuracy in the case of different source locations from the training data. (a) Ex-
pected vs. (b) estimated values, and spatial distributions of (c) MSE and (d) R2 when the heat source
was located at the bottom center. (e) Expected vs. (f) estimated values, and spatial distributions of
(g) MSE and (h) R2 when the heat source was located at the bottom right.
Geosciences 2022, 12, 130 15 of 18

4. Discussion
The machine learning approach proposed in this study directly estimates input pa-
rameters in the reservoir model from measurable data. This eliminates the need for a
trial-and-error search for input parameters in the reservoir modeling, which is a major
challenge in conventional modeling approaches. Of course, inverse analysis methods
developed in the past (e.g., iTOUGH2 [14]) could also provide good estimates. Although
we have not compared the accuracy of the estimation between inverse analysis methods
and our approach, the estimation accuracy may be better in both cases, depending on the
optimization. On the other hand, the good point of our approach is that once the learning
model is created, it does not need to be computed over and over again, and it can be used
even on computers with low specifications. This would help to expand the spread of small
fields, which cannot be developed with large amounts of equipment. It also provides
reliable reservoir modeling which does not rely on the experience and intuition of analysts.
This allows non-experienced developers to work on reservoir modeling and helps to create
an environment where new people can take part in geothermal development. Reservoir
modeling is one of the most important challenges in making strategies for geothermal
reservoir development. Improvements in the reliability of reservoir modeling are expected
to greatly accelerate the geothermal development.
Geothermal developments cannot drill a large number of wells due to the drilling
cost. Thus, the pressure data in the natural state is only available for the discrete data
from the limited wells. The learning model proposed in this study, however, requires two-
dimensional pressure distributions. If we apply our method to field data, it is necessary to
measure the pressure from at least three points surrounding the target surface for the 2D
estimation and to interpolate the discrete data by interpolation techniques, such as kriging
(e.g., [52]). Future research will examine the estimation errors when interpolations are
performed.
In addition, when not many data points are available, as in the case of geothermal
development, it is important to evaluate carefully how uncertain the measurements based
on the data are. To evaluate the impacts of the uncertainties on the estimation, uncer-
tainty quantification methods, such as Bayesian approximation and ensemble learning
techniques (e.g., [53]), play some pivotal roles. Combining uncertainty quantification with
our approach is also desirable for future study.
This study is limited to numerical experiments and has not been applied to actual field
data. At present, a model for two-dimensional data has been developed, but a model for
three-dimensional data is expected to be developed, assuming the actual field. In addition,
our simulation was limited to thermohydraulic simulation as a first step to develop the
machine learning approach. To make our analysis available for real field development, 3D
thermohydraulic–mechanical–chemical simulations are needed in the next research.
This study applied several regression models in machine learning to estimate the
permeability based on the two-dimensional distributed features. Two-dimensional image
recognition can be powerfully trained by convolutional neural networks (CNNs) [54], and
the application of CNNs should be considered in future research.
Our approach is similar to data-driven physics-informed machine learning to search
for new governing equations of physical phenomena, which has been a hot topic in the
field of machine learning (e.g., [55–57]). Although such approaches are still only applied to
basic science, they are expected to be developed in the fields of Earth science and energy
resource engineering. In this study, we successfully demonstrated the first step toward
the goal of achieving a new data-driven geothermal reservoir engineering, which will be
developed and enhanced by incorporating information science.

5. Conclusions
We developed a machine learning model to uniquely estimate input parameters based
on measured data in order to improve the conventional reservoir modeling approach, which
determines the input parameters by trial and error. In this study, the relationship between
Geosciences 2022, 12, 130 16 of 18

the measurement data and permeability, one of the most important input parameters in
reservoir modeling, was successfully determined by machine learning. First, we generated
a large number of permeability distributions based on the discrete cosine transform and
conducted natural state simulations using each permeability distribution. The simulation
results of pressure and temperature distributions were used for feature variables. We used
several popular supervised machine learning approaches: the linear regression, the ridge
regression, the Lasso regression, the support vector regression (SVR), multilayer perceptron
(MLP), random forest, gradient boosting, and the k-nearest neighbor algorithm. Learning
models were developed both with and without the dependence of the grids. The results
showed that the grid-independent model by random forest with the pressure differences
as feature variables provided good estimations of permeability. It was also found that
the model could be applied to the test dataset with different mass flow rates of the heat
sources. The estimation became more difficult when the position of the source was different.
However, the estimation was successful for the region with a flow field even when the
position of the source was different. We successfully demonstrated the first step toward
the goal of achieving new data-driven geothermal reservoir engineering, which will be
developed and enhanced by incorporating information science.

Author Contributions: Conceptualization, A.S. and T.H.; methodology, K.-i.F. and A.S.; software,
validation, formal analysis, investigation, data curation, writing—original draft preparation, visu-
alization, A.S.; writing—review and editing, K.-i.F., S.O., J.I. and T.H.; project administration, T.H.;
funding acquisition, A.S. and T.H. All authors have read and agreed to the published version of
the manuscript.
Funding: This research was funded by JSPS KAKENHI Grant Numbers JP20H02676 (Japan) and JST
ACT-X Grant Number JPMJAX190H (Japan).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The datasets generated in this study and the program code are available
from the authors on a reasonable request.
Acknowledgments: We We would like to thank the members in Joint Research program 2021-B-01
and 2018-B-01 for useful discussions.
Conflicts of Interest: The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in this paper. This
study was supported by Tohoku Electric Power Co., Inc., of whom Shinya Onodera and Junichi
Ishizaki are employees. The company and the funders had no role in the design of the study; in the
collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to
publish the results.

References
1. Pruess, K.; Oldenburg, C.M.; Moridis, G.J. TOUGH2 User’s Guide, version 2; LBNL-43134; Lawrence Berkeley National Lab.:
Berkeley, CA, USA, 1999.
2. Vinsome, K.; Shook, M. Multi-purpose simulation. J. Pet. Sci. Eng. 1993, 9, 29–38. [CrossRef]
3. Pritchett, J.W. STAR: A geothermal reservoir simulation system. In Proceedings of the World geothermal Congress, Florence,
Italy, 18–31 May 1995; pp. 2959–2963.
4. Keller, J.; Rath, V.; Bruckmann, J.; Mottaghy, D.; Clauser, C.; Wolf, A.; Seidler, R.; Bücker, H.M.; Klitzsch, N. SHEMAT-Suite: An
open-source code for simulating flow, heat and species transport in porous media. SoftwareX 2020, 12, 100533. [CrossRef]
5. Hughes, J.; Langevin, C.; Banta, E. Documentation for the MODFLOW 6 framework. In USGS: Techniques and Methods 6-A57; U.S.
Geological Survey: Reston, VA, USA, 2017; p. 42.
6. Mahmoodpour, S.; Singh, M.; Turan, A.; Bär, K.; Sass, I. Hydro-Thermal Modeling for Geothermal Energy Extraction from
Soultz-sous-Forêts, France. Geosciences 2021, 11, 464. [CrossRef]
7. Ganguly, S.; Kumar, M.S. Geothermal reservoirs—A brief review. J. Geol. Soc. India 2012, 79, 589–602. [CrossRef]
8. Pratama, H.B.; Saptadji, N.M. Numerical simulation for natural state of two-phase liquid dominated geothermal reservoir with
steam cap underlying brine reservoir. IOP Conf. Ser. Earth Environ. Sci. 2016, 42, 012006. [CrossRef]
Geosciences 2022, 12, 130 17 of 18

9. Manggala Putra, R.P.; Sutopo, S.; Pratama, H.B. Improved natural state simulation of Arjuno-Welirang Geothermal field, East
Java, Indonesia. IOP Conf. Ser. Earth Environ. Sci. 2019, 254, 012022. [CrossRef]
10. Jalilinasrabady, S.; Tanaka, T.; Itoi, R.; Goto, H. Numerical simulation and production prediction assessment of Takigami
geothermal reservoir. Energy 2021, 236, 121503. [CrossRef]
11. Grant, M.A.; Bixley, P.F. Geothermal Reservoir Engineering, 2nd ed.; Academic Press: Oxford, UK, 2011; p. 378.
12. Finsterle, S.; Pruess, K. Development of Inverse Modeling Techniques for Geothermal Applications; LBNL-40039; Lawrence Berkeley
Lab.: Berkeley, USA, 1997; p. 8.
13. O’Sullivan, M.J.; Pruess, K.; Lippmann, M.J. State of the art of geothermal reservoir simulation. Geothermics 2001, 30, 395–429.
[CrossRef]
14. Finsterle, S. iTOUGH2 User’s Guide; LBNL-40040; Lawrence Berkeley Lab.: Berkeley, USA, 2007; p. 137.
15. Poeter, E.P.; Hill, M.C. UCODE, a computer code for universal inverse modeling. Comput. Geosci. 1999, 25, 457–462. [CrossRef]
16. Doherty, J. Calibration and uncertainty analysis for complex environmental models. Groundwater 2015, 53, 673–674. [CrossRef]
17. Bjarkason, E.K.; O’Sullivan, J.P.; Yeh, A.; O’Sullivan, M.J. Inverse modeling of the natural state of geothermal reservoirs using
adjoint and direct methods. Geothermics 2019, 78, 85–100. [CrossRef]
18. Assouline, D.; Mohajeri, N.; Gudmundsson, A.; Scartezzini, J.L. A machine learning approach for mapping the very shallow
theoretical geothermal potential. Geotherm. Energy 2019, 7, 19. [CrossRef]
19. Spichak, V.; Geiermann, J.; Zakharova, O.; Calcagno, P.; Genter, A.; Schill, E. Estimating deep temperatures in the Soultz-sous-
Forêts geothermal area (France) from magnetotelluric data. Near Surf. Geophys. 2015, 13, 397–408. [CrossRef]
20. Ishitsuka, K.; Kobayashi, Y.; Watanabe, N.; Yamaya, Y.; Bjarkason, E.; Suzuki, A.; Mogi, T.; Asanuma, H.; Kajiwara, T.; Sugimoto, T.;
et al. Bayesian and neural network approaches to estimate deep temperature distribution for assessing a supercritical geothermal
system: Evaluation using a numerical model. Nat. Resour. Res. 2021, 30, 3289–3314. [CrossRef]
21. Rezvanbehbahani, S.; Stearns, L.A.; Kadivar, A.; Walker, J.D.; van der Veen, C.J. Predicting the geothermal heat flux in Greenland:
A machine learning approach. Geophys. Res. Lett. 2017, 44, 12271–12279. [CrossRef]
22. Siler, D.L.; Pepin, J.D.; Vesselinov, V.V.; Mudunuru, M.K.; Ahmmed, B. Machine learning to identify geologic factors associated
with production in geothermal fields: A case-study using 3D geologic data, Brady geothermal field, Nevada. Geotherm. Energy
2021, 9, 17. [CrossRef]
23. Gudmundsdottir, H.; Horne, R.N. Prediction modeling for geothermal reservoirs using deep learning. In Proceedings of the 45th
Workshop on Geothermal Reservoir Engineering, Stanford University, Stanford, CA, USA, 10–12 February 2020; p. 12.
24. Holtzman, B.K.; Paté, A.; Paisley, J.; Waldhauser, F.; Repetto, D. Machine learning reveals cyclic changes in seismic source spectra
in Geysers geothermal field. Sci. Adv. 2018, 4, eaao2929. [CrossRef] [PubMed]
25. Gao, K.; Huang, L.; Lin, R.; Hu, H.; Zheng, Y.; Cladohous, T. Delineating faults at the soda lake geothermal field using machine
learning. In Proceedings of the 46th Workshop on Geothermal Reservoir Engineering, Stanford University, Stanford, CA, USA,
16–18 February 2021; p. 8.
26. Zheng, Y.; Li, J.; Lin, R.; Hu, H.; Gao, K.; Huang, L.; Sciences, A.; Alamos, L. Physics-Guided Machine Learning Approach to
Characterizing Small-Scale Fractures in Geothermal Fields. In Proceedings of the 46th Workshop on Geothermal Reservoir
Engineering, Stanford University, Stanford, CA, USA, 16–18 February 2021; p. 9.
27. Ali, S.S.; Nizamuddin, S.; Abdulraheem, A.; Hassan, M.R.; Hossain, M.E. Hydraulic unit prediction using support vector machine.
J. Pet. Sci. Eng. 2013, 110, 243–252. [CrossRef]
28. Al-Mudhafar, W.J. Integrating well log interpretations for lithofacies classification and permeability modeling through advanced
machine learning algorithms. J. Pet. Explor. Prod. Technol. 2017, 7, 1023–1033. [CrossRef]
29. Anifowose, F.; Abdulraheem, A.; Al-Shuhail, A. A parametric study of machine learning techniques in petroleum reservoir
permeability prediction by integrating seismic attributes and wireline data. J. Pet. Sci. Eng. 2019, 176, 762–774. [CrossRef]
30. Erofeev, A.; Orlov, D.; Ryzhov, A.; Koroteev, D. Prediction of porosity and permeability alteration based on machine learning
algorithms. Transp. Porous Media 2019, 128, 677–700. [CrossRef]
31. Kaydani, H.; Mohebbi, A.; Eftekhari, M. Permeability estimation in heterogeneous oil reservoirs by multi-gene genetic program-
ming algorithm. J. Pet. Sci. Eng. 2014, 123, 201–206. [CrossRef]
32. Sudakov, O.; Burnaev, E.; Koroteev, D. Driving digital rock towards machine learning: Predicting permeability with gradient
boosting and deep neural networks. Comput. Geosci. 2019, 127, 91–98. [CrossRef]
33. Al-Anazi, A.F.; Gates, I.D. Support vector regression for porosity prediction in a heterogeneous reservoir: A comparative study.
Comput. Geosci. 2010, 36, 1494–1503. [CrossRef]
34. Mo, S.; Zabaras, N.; Shi, X.; Wu, J. Deep autoregressive neural networks for high-dimensional inverse problems in groundwater
contaminant source identification. Water Resour. Res. 2019, 55, 3856–3881. [CrossRef]
35. Wen, G.; Tang, M.; Benson, S.M. Multiphase flow prediction with deep neural networks. arXiv 2019, arXiv:1910.09657. [CrossRef]
36. Tang, M.; Liu, Y.; Durlofsky, L.J. A deep-learning-based surrogate model for data assimilation in dynamic subsurface flow
problems. J. Comput. Phys. 2020, 413, 109456. [CrossRef]
37. Jin, Z.L.; Liu, Y.; Durlofsky, L.J. Deep-learning-based surrogate model for reservoir simulation with time-varying well controls. J.
Pet. Sci. Eng. 2020, 192, 107273. [CrossRef]
38. Liu, Y.; Durlofsky, L.J. 3D CNN-PCA: A deep-learning-based parameterization for complex geomodels. Comput. Geosci. 2021,
148, 104676. [CrossRef]
Geosciences 2022, 12, 130 18 of 18

39. Ahmed, N.; Natarajan, T.; Rao, K.R. Discrete cosine transform. IEEE Trans. Comput. 1974, C-23, 90–93. [CrossRef]
40. Hoerl, A.E.; Kennard, R.W. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics 1970, 12, 55–67.
[CrossRef]
41. Tibshirani, R. Regression shrinkage and selection via the lasso. J. R. Stat. Soc. Ser. B (Methodol.) 1996, 58, 267–288. [CrossRef]
42. Vapnik, V. The Nature of Statistical Learning Theory, 2nd ed.; Springer Science & Business Media: Berlin/Heidelberg, Germany,
Springer: New York, NY, USA, 2000. [CrossRef]
43. Orr, G.B.; Müller, K.R. Neural Networks: Tricks of the Trade, 2nd ed.; Springer: Berlin/Heidelberg, Germany, 2012. [CrossRef]
44. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
45. Mason, L.; Baxter, J.; Bartlett, P.; Frean, M. Boosting algorithms as gradient descent in function space. Proc. NIPS 1999, 12, 512–518.
[CrossRef]
46. Peterson, L.E. K-nearest neighbor. Scholarpedia 2009, 4, 1883. [CrossRef]
47. Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.;
et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830. [CrossRef]
48. Akiba, T.; Sano, S.; Yanase, T.; Ohta, T.; Koyama, M. Optuna: A next-generation hyperparameter optimization framework. In
Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’19), Association
for Computing Machinery, Anchorage, AK, USA, 4–8 August 2019; pp. 2623–2631. [CrossRef]
49. Guyon, I.; Elisseeff, A. An introduction to variable and feature selection. J. Mach. Learn. Res. 2003, 3, 1157–1182. [CrossRef]
50. Shorten, C.; Khoshgoftaar, T.M. A survey on image data augmentation for deep learning. J. Big Data 2019, 6, 60. [CrossRef]
51. Zhao, S.; Yue, X.; Zhang, S.; Li, B.; Zhao, H.; Wu, B.; Krishna, R.; Gonzalez, J.E.; Sangiovanni-Vincentelli, A.L.; Seshia, S.A.; et al.
A review of single-source deep unsupervised visual domain adaptation. IEEE Trans. Neural Netw. Learn. Syst. 2022, 33, 473-493.
[CrossRef]
52. Teng, Y.; Koike, K. Three-dimensional imaging of a geothermal system using temperature and geological models derived from a
well-log dataset. Geothermics 2007, 36, 518–538. [CrossRef]
53. Jiang, Z.; Zhang, S.; Turnadge, C.; Xu, T. Combining autoencoder neural network and Bayesian inversion to estimate hetero-
geneous permeability distributions in enhanced geothermal reservoir: Model development and verification. Geothermics 2021,
97, 102262. [CrossRef]
54. Lawrence, S.; Giles, C.L.; Tsoi, A.C.; Back, A.D. Face recognition: A convolutional neural-network approach. IEEE Trans. Neural
Netw. 1997, 8, 98–113. [CrossRef] [PubMed]
55. Champion, K.; Lusch, B.; Nathan Kutz, J.; Brunton, S.L. Data-driven discovery of coordinates and governing equations. Proc.
Natl. Acad. Sci. USA 2019, 116, 22445–22451. [CrossRef] [PubMed]
56. Raissi, M.; Yazdani, A.; Karniadakis, G.E. Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations.
Science 2020, 367, 1026–1030. [CrossRef] [PubMed]
57. Harp, D.R.; O’Malley, D.; Yan, B.; Pawar, R. On the feasibility of using physics-informed machine learning for underground
reservoir pressure management. Expert Syst. Appl. 2021, 178, 115006. [CrossRef]