Professional Documents
Culture Documents
Qiuyang Shen, Xuqing Wu, Jiefu Chen, Zhu Han, Yueqin Huang
PII: S0920-4105(17)30921-X
DOI: 10.1016/j.petrol.2017.11.031
Reference: PETROL 4447
Please cite this article as: Shen, Q., Wu, X., Chen, J., Han, Z., Huang, Y., Solving geosteering inverse
problems by stochastic Hybrid Monte Carlo method, Journal of Petroleum Science and Engineering
(2017), doi: 10.1016/j.petrol.2017.11.031.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
PT
RI
U SC
AN
M
D
TE
EP
C
AC
Faster converging speed by HMC Inverse earth model by HMC with quantified uncertainty
ACCEPTED MANUSCRIPT
PT
Qiuyang Shena , Xuqing Wub , Jiefu Chena,∗, Zhu Hana , Yueqin Huangc
a
Department of Electrical and Computer Engineering, University of Houston
RI
b
Department of Information and Logistics Technology, University of Houston
c
Cyentech Consulting LLC
SC
Abstract
U
The inverse problems arise in almost all fields of science where the real-world
parameters are extracted from a set of measured data. The geosteering inver-
AN
sion plays an essential role in the accurate prediction of oncoming strata as
well as a reliable guidance to adjust the borehole position on the fly to reach
one or more geological targets. This mathematical treatment is not easy to
M
solve, which requires finding an optimum solution among a large solution
space, especially when the problem is non-linear and non-convex. Nowa-
days, a new generation of logging-while-drilling (LWD) tools has emerged on
D
the market. The so-called azimuthal resistivity LWD tools have azimuthal
sensitivity and a large depth of investigation. Hence, the associated inverse
TE
problems become much more difficult since the earth model to be inverted
will have more detailed structures. The conventional deterministic methods
are incapable to solve such a complicated inverse problem, where they suffer
from the local minimum trap. Alternatively, stochastic optimizations are in
EP
∗
Corresponding author
Preprint submitted to Journal of Petroleum Science and Engineering November 15, 2017
ACCEPTED MANUSCRIPT
1. Introduction
Obtaining reliable and detailed information about the earth subsurface
is of great challenge. The task to infer the interior structure is fundamen-
PT
tal for the geological exploitation or economic evaluation [1]. Geosteering
5 is a technique to actively adjust the direction of drilling, often in horizontal
wells, based on real-time formation evaluation by using logging-while-drilling
RI
(LWD) measurements [2]. This process enables drillers to efficiently reach
the target zone and actively respond while drilling to geological changes in
the formation so that they can maintain the maximal reservoir contact [3].
SC
10 The real-time adjustment of a geosteering process is determined by the cur-
rent position and depth of the tool to minimize the gas or oil breakthrough
and maximize the economic production [4]. Among several technologies that
U
can be used for geosteering, such as nuclear, acoustic, gamma ray, or elec-
tromagnetic measurement, azimuthal resistivity LWD tools are widely used
15
AN
in geosteering worldwide due to its relatively large depth of detection and
azimuthal directional sensitivity [5]. Compared to other conventional LWD
propagation resistivity tools, the azimuthal directional LWD tool can obtain
M
more information related to the formation anisotropy in addition to the re-
sistivity measurements [6, 7]. Hence, it provides a more sensitive detection
20 of formation boundaries. In the while-drilling environment, this capacity can
D
X
EP
Z
T5 T3 T1 R3 R1 R2 R4 T2 T4 T6
C
AC
Figure 1: The structure and schematic of an azimuthal resistivity LWD tool. T1, T2, T3,
and T4 are the transmitters whose moment are with the tool axis (Z direction), while T5
and T6 are transverse antennas that perpendicular to the tool axis (X direction). Similarly,
R1 and R2 are the receivers directing along the tool axis. R3 and R4 are receiver antennas
with azimuthal sensitivity.
2
ACCEPTED MANUSCRIPT
different polarizations and working frequencies. They play the roles as trans-
mitters and receivers that are oftentimes in pairs. Each pair of antennas
25 can provide a group of electromagnetic signals while the tool is drilling [8].
PT
A schematic model diagram of an azimuthal resistivity LWD tool is shown
in Figure 1. It has transmitters and receivers both along the tool axis (Z
direction) or perpendicular to it (X direction). While drilling, transmitters
RI
are energized and the voltage changes recorded by the receivers. Processing
30 algorithms in the electronic hardware deduce the measured signals from the
receiver antennas to the full mutual inductance tensor, related to the resistiv-
SC
ity tensor of the geological formation occupied by the tool [9]. These differ-
ent inductance tensors, with working frequencies in the range from kilohertz
to megahertz, are used to generate different types of measurement curves.
The curve responses are collected as measurement data and served for the
U
35
evaluate the measurements from the deep-reading tool, and present the re-
constructed earth model in real time. Hence, a robust and efficient inverse
algorithm is always necessary to satisfy the job requirement. In practice,
D
45
3
ACCEPTED MANUSCRIPT
PT
approaches to draw samples from an unknown distribution [14]. The MCMC
65 method guarantees an asymptotically exact solution for recovering the poste-
rior distribution, though the computational cost is inevitably high and most
RI
MCMC algorithms suffer from a low acceptance rate and slow convergence
with long burn-in periods [15]. In fact, solving geophysical inverse problems
by statistical method is not new. The implementation of Bayesian inference
SC
70 on the interpretation of sonic logging measurements has been realized [16].
However, the poor acceptance rate with long burn-in time is unbearable while
applying this scheme on geosteering inversion cases. In this article, Hybrid
Monte Carlo (HMC) is used to solve the geosteering inverse problem. HMC
U
[17] united the MCMC and molecular dynamics and provided a way of sam-
75
AN
pling from the canonical density by simulating the dynamics of a physical
system. Many experiments show that the sampling performance is improved
via the HMC method through avoiding the random walk behavior used by
the traditional MCMC [18].
M
In the following sections, we will first give a brief review on the back-
80 ground of the geosteering problem and its statistical scheme. Then we will
discuss the HMC method and its implementation for the geosteering inverse
D
the HMC method with the synthetic model and demonstrate its advantages
85 on solving statistical geosteering inverse problems.
EP
2. Geosteering Inversions
2.1. Deterministic Inversions
C
error term between the forward model responses and observations. Assume
a collected N measurement data denoted by m ∈ RN . x ∈ RM represents M
earth model parameters of the real world inverted from the measured data.
A computational model function or so-called forward function S : RM → RN
is designed to synthesize N responses from M model parameters. The for-
ward transformation from the model parameters to the responses is calcu-
lated based on the 1D electromagnetic modeling, which is essentially solving
4
ACCEPTED MANUSCRIPT
Maxwell’s equations with magnetic dipole sources in layered media [19, 20].
The objective of the inverse problem is to infer the model parameters through
the observed measurement. A good agreement between the response of the
PT
forward model and measured data will be reached if inverse parameters of the
physical model are accurate. The difference between the forward response
and measurements is defined as data misfit F (x), which is written as:
RI
F (x) = S(x) − m (1)
Since both forward responses and measurements are vectors, a cost function
SC
is defined as the square of L2 norm of the misfit function F (x):
N
X
f (x) = Fi2 (x) = kF (x)k22 (2)
U
i=1
AN
where f (x) is the cost function representing the magnitude of the data mis-
fit. Hence, the inverse problem is to find the optimal model parameters,
x, which minimize the cost function given a forward model function and
measurements. Mathematically, this problem is presented as:
M
min f (x) (3)
x∈RM
D
5
ACCEPTED MANUSCRIPT
The assumptions made by the forward model y = f (x) (y is data and x de-
notes the earth model parameter) may not include all factors that affect mea-
surements. Suppose the noise is additive and comes from external sources,
PT
the relationship between observed outputs ye and corresponding model pa-
rameters can be represented as:
ye = f (x) + ε (5)
RI
where ε denotes additive noises. Experiments empirically suggest that ad-
ditive noises usually follow a zero-mean Gaussian random distribution: ε ∼
SC
N (0, σ 2 I). Given model parameters x and observed data ye, the likelihood
can be deduced as
y |x) = p(ε) ∼ N (0, σ 2 I)
p(e (6)
U
Suppose the prior distribution of x is governed by a zero-mean isotropic
Gaussian distribution such that p(x) ∼ N (0, β 2 I). By virtue of the Bayes
AN
equation (4), the posterior distribution of x is given by
y ) ∼ N (0, σ 2 I)N (0, β 2 I)
p(x|e (7)
It suggests that the posterior distribution of model parameters x given ob-
M
servations ye can be obtained by calculating the product of two Gaussian dis-
100 tributions. The solution of x can be sampled and estimated according to the
D
6
ACCEPTED MANUSCRIPT
PT
then moves towards x∗ with the following acceptance probability
p(x∗ )q(x|x∗ )
∗
A(x, x ) = min 1, , (8)
p(x)q(x∗ |x)
RI
otherwise, it remains at x. In the random walk MH algorithm, a zero-mean
normal distribution is a popular choice of q(x∗ |x) as a symmetric candidate-
SC
∗)
generating function, which helps reduce the moving probability to p(xp(x)
. Al-
gorithm 1 presents the MH algorithm for sampling from the posterior distri-
120 bution:
U
Algorithm 1 The Random Walk MCMC Method
AN
Input: initial x(0) , jumping function q(x(i) |x(j) ), max chain length K
Output: x(k) , where k < K
Initialize with arbitrary value x(0)
while k ≤ K do
M
Generate x(k) from q(x (k) (k−1)
n |x (k) ) o
p(x |ey)
A(x(k) , x(k−1) ) = min 1, p(x (k−1) |e
y)
D
keep x(k)
else
x(k) = x(k−1)
EP
end if
save x(k) in the chain
end while
C
distribution requires very few parameter tuning. However, although the ran-
dom walk MCMC method guarantees asymptotically exact recovery of the
posterior distribution as the number of posterior samples grows, it may suffer
125 from an extremely long burn-in period before reaching the equilibrium and
slow convergence. The time cost may be prohibitively high for the inverse
problem when the forward model is repeatedly computed by every sample
drawn from the MH, which may take at least O(N ) operations to draw one
7
ACCEPTED MANUSCRIPT
sample [25]. Even worse, with the increasing dimension of model parameters,
130 the burn-in period will extend exponentially.
PT
3. Hybrid Monte Carlo
3.1. Hamiltonian Dynamics
RI
The HMC or so-called Hamiltonian Monte Carlo replaces the proposal
function of the random walk by the simulation of a dynamic process. The
135 idea initially came from the observation of the motion of molecules and was
SC
well described by the Hamiltonian dynamics in the landmark paper by Duane
et al. [17]. Generalized by Neal in 1993 [26], HMC converges more quickly
to the absolute probability distribution.
U
To interpret the algorithm of HMC, first, we introduce the Hamiltonian
dynamics with two sets of real variables q and p. For a physical interpre-
AN
tation of these variables, the vector q = {q1 , q2 , . . . , qn } stands for the posi-
tion of n molecules on a frictionless surface of varying height. The vector
p = {p1 , p2 , . . . , pn } represents their current momentums. The corresponding
energy functions, U (q) and K(p), also share the physical meaning as the po-
M
tential energy and the kinetic energy, respectively. Suppose we are going to
sample from the canonical distribution for the states of molecules, it becomes
our interest to find the probability distribution of the molecules’ position. In
D
1
P (q) = exp(−U (q)) (9)
Ze
EP
where Ze is the normalizing constant. The equation (9) bridges the view
of energy and statistical probability given a state of the object. In the real
world, vector q is a set of unknown parameters that we want to find its
C
8
ACCEPTED MANUSCRIPT
PT
It describes the dynamical system through the perspective of the total energy,
to which the partial derivatives determine how q and p change over continuous
time t according to the Hamiltons equation:
RI
dqi ∂H
=
dt ∂pi
SC
(12)
dpi ∂H
=−
dt ∂qi
The stochastic gradient equations above construct the foundation of HMC
U
process. According to the definition of the Hamiltonian function, some prop-
AN
erties are able to be concluded and promise the feasibility applying on the
MCMC process. Starting from the first property, the Hamiltonian dynam-
ics is reversible, which means from the current state at time t, (q(t), p(t)),
to the state at time t + s, (q(t + s), p(t + s)), is one-to-one, and hence has
M
an inverse. The reversibility of dynamic guarantees the desired distribution
invariant during sampling. The second property is conservation, which de-
notes that the Hamiltonian function keeps invariant along time, which can
D
be easily shown as
TE
n
dH X dqi ∂H dpi ∂H
= + =0 (13)
dt i=1
dt ∂q i dt ∂p i
EP
villes theorem. One brief explanation is that if we apply the time mapping T
145 to the points in some region R of the (q, p) space with volume V , the image
of R under T will also have volume V [27]. This property guarantees the
acceptance probability for the Metropolis updates without any influence by
the change in its volume.
In order to sample the state via variables qi and draw a distribution to
reflect the statistical properties, the continuous-time Hamiltons equations
9
ACCEPTED MANUSCRIPT
PT
minimizes the error introduced by the discretization. A single step to upgrade
the position and momentum, p and q is shown as,
ε ∂
p(k) (t + ε/2) = p(k) (t) − · (k) U (q (k) (t))
RI
2 ∂q
ε
q (k) (t + ε) = q (k) (t) + ε · p(k) (t + ) (14)
2
SC
ε ∂
p(k) (t + ε) = p(k) (t + ε/2) − · (k) U (q (k) (t + ε))
2 ∂q
It starts with a half step update for the momentum variable p, followed by
U
150 a full step update for the target variable q, and finally another half step for
p. With dT /ε times updating, the system will move to a new state. The
AN
discretized leapfrog process realizes a full update of a sample q, governed by
the Hamiltonian dynamic. The update of the auxiliary variable p is always
ignored since p is drawn randomly each time at the beginning of leapfrog
M
155 process.
are two main processes at each iteration in the HMC algorithm. In the first
160 part, new values of the momentum variables p are drawn randomly from their
Gaussian distribution. Starting from the current state (q, p), an update is
performed using the simulation of Hamiltonian dynamics and moves (q, p) in
EP
equation (8). If the proposed state is rejected, the current q will be kept for
the next iteration.
AC
10
ACCEPTED MANUSCRIPT
PT
Initialize with arbitrary value q (0) , stepsize ε, and leapfrog steps l
while k ≤ K do
Randomly generate p(k) from N (0, 1)
RI
(k+1)
q0 = q (k)
(k+1)
p0 = p(k) − 2ε · ∂q∂(k) U (q (k) )
for i = 1 to l do
SC
(k+1) (k+1) (k+1)
qi = qi−1 + ε · pi−1
(k+1) (k+1) ∂ (k+1)
pi = pi−1 − ε · (k+1) U (qi )
∂qi
end for
U
A(q (k+1) , q (k) ) = min 1, exp[ U (q (k) ) − U (q (k+1) ) + K(p(k) ) − K(p(k+1) )]
Generate A0 from uniform distribution U(0, 1)
if A0 < A(q (k+1) , q (k) ) then
keep q (k+1)
AN
else
M
q (k+1) = q (k)
end if
save q (k+1) in the chain
D
end while
TE
11
ACCEPTED MANUSCRIPT
PT
rate than the random walk exploration. Also, the average acceptance rate of
HMC is 0.65 while the random walk is around 0.23 [27]. The higher accep-
tance rate leads to a faster exploration of the canonical distribution and its
RI
180 statistical properties. In the meantime, the random mechanism inside the
Hamiltonian dynamics differentiates HMC from deterministic inversions and
avoids local optimums. As a class of MCMC methods, choosing a suitable
SC
stepsize and leapfrog steps is crucial when implementing HMC. Together, the
stepsize and leapfrog steps determine the length of the trajectory in fictitious
185 time, εL. A large stepsize may result in unstable Hamiltonian dynamics due
to errors introduced into the system by the leapfrog discretization. Too small
U
a stepsize will produce little change of the state, waste computation time,
AN
and cause a slow exploration of the target distribution [27]. The selection
of leapfrog steps is also a subtle work and the distance to move within a
190 simulation of the Hamiltonian may differ from one state to another state. In
practice, a preliminary run of HMC is often required to determine a suit-
M
able choice of the stepsize and trajectory length. An effective strategy is
to select a range of these two parameters according to the preliminary run,
and randomly choose a pair of these parameters within this range for each
D
195 simulation of the Hamiltonian dynamics. The randomness helps increase the
overall possibility of fast exploration of the target distribution.
TE
200 the increasing dimension of the sampling space, the multiple-chain sampling
strategy has proven to be helpful on weakening the correlation of each sample
and thus improving the possibility of convergence [29]. Multiple-chain HMC
C
was implemented in this paper to solve the geosteering inverse problem. The
target distribution and model parameters are obtained by taking between-
AC
12
ACCEPTED MANUSCRIPT
PT
215 4. Convergence and Model Inversions
RI
In this section, we demonstrate the application of the HMC method on
solving the geosteering inverse problem and present the advantages of the
HMC compared to the traditional inversion methods. The examples are con-
SC
ducted based on the synthetic dataset generated by the simulated azimuthal
220 resistivity LWD tool. To its configuration, multiple working frequencies are
used, which is from the lowest 1 kHz to the highest 2 MHz. The correspond-
ing transmitter-receiver spacing ranges from 800 inches to 30 inches. The
U
assumption to the depth of investigation (DOI) implies it can achieve up to
100 ft from the wellbore by using curves with working frequency of 1 kHz.
AN
225 Actually, in these couple of years, many new generation LWD tools with
such deep reading ability have emerged on the oil and gas market. This sim-
ulated tool is presented and verified according to the most recent industrial
M
development [10]. A synthetic earth structure is constructed by models with
three or more layers. The performance of HMC method is evaluated through
230 several aspects including the convergence plot, uncertainty analysis, and full
D
misfit. Of course, in the practical geosteering work, the true solution of earth
model parameters are never available while the comparison of data misfit is
the only approach to evaluate the inversion performance. However, the non-
C
convexity caused local minimum denotes a very small data misfit even the
240 inverse model is tremendously different from a real one. Hence in our ex-
AC
13
ACCEPTED MANUSCRIPT
150
HMC sampling chain
MCMC sampling chain
PT
model misfit
100
RI
50 MCMC chain
SC
HMC chain
0
U
100 101 102 103 104
ANsamples
Figure 2: Convergence of model misfit with the increasing number of samples.The upper
MCMC chain converges after 2000 samples while the lower HMC chain converges faster
M
within 30 samples with higher acceptance rate.
between the true solution and the inverse result, as an indicator to show the
250 performance and chain convergence.
TE
In Figure 2, the result shows that the chain governed by HMC enters an
equilibrium state after a burn-in period within 30 iterations, and the model
misfit is near to zero, which means the samples are drawn from the tar-
EP
eters has very low efficiency to explore the parameter space given a limited
AC
chain length. Our experiment indicates that the model misfit of random walk
260 chain converges to around 0 after 2000 samples. This test agrees with that
the statistical inversion by the HMC sampling method is more efficient than
the MCMC method. HMC is able to draw samples precisely on the canon-
ical distributions of model parameters. The other prominent improvement
of HMC is that the acceptance rate is much higher, which guarantees the
265 effectiveness in sampling the canonical distributions. MCMC suffers from
14
ACCEPTED MANUSCRIPT
PT
RI
SC
(a) The synthetic three-layer earth model, where the logging tool is drilling horizontally in
the middle high-resistive layer.
U
AN
M
D
TE
(b) The inverse model with uncertainty towards two boundaries. The inverse results of
further distance-to-boundary have higher uncertainty.
EP
the low acceptance rate as shown in Figure 2 where the upper line of chain
C
maintains the same sample value for many iterations. This behavior indi-
cates that the randomness is the only force to make the parameters change,
AC
15
ACCEPTED MANUSCRIPT
275 a 2000 ft horizontal well. The synthetic earth model, shown in Figure 3(a),
is a three-layer model, where the resistivity is 10 ohm-m, 50 ohm-m and 1
ohm-m from the top to bottom, respectively. The central dash line indicates
PT
the tools navigation trajectory. In this case, we assume the tool relative dip
angle is fixed at 90 degrees and the drilling trajectory is horizontal. The
280 depth to either the upper or lower boundary is varying on different positions,
RI
where the furthest boundary is up to 70 ft and closest one is 3 ft.
In Figure 3(b), we show the inversion result by the HMC method with
the uncertainty of depth of layer boundaries. It is obvious that the inverted
SC
model by HMC agrees with the true earth model with satisfactory perfor-
285 mance. Beyond this, the error bar given by the HMC method on the two
boundaries denotes the uncertainty of inversion results. It presents the vari-
ance of collected samples from a drawn distribution. In this case, the uncer-
U
tainty becomes larger when the tool goes far from the boundary, and this is
290
AN
consistent with drilling engineers’ experiences in the practice. Overall, the
statistical HMC inversion is able to reconstruct the earth model and gives the
relevant information about inverse uncertainty, which is meaningful to the
real-time geosteering operation and helps keep the drilling tool from breaking
M
out.
To the last example, we use a complicated five-layer model to verify the
capability of HMC inversion, comparing with deterministic inversion. The
D
295
Figure 4(a) presents the real model, where the tool is drilling across from
the top layer to the bottom along the dash line. The formation resistivity
TE
changes from low to high and to low alternatively. Figure 4(b) presents in-
version results by a deterministic method based on the Levenberg-Marquardt
300 algorithm. Generally, the deterministic inversions rely heavily on the initial
EP
guess. The outcome is unsatisfactory, and many results suffer from incorrect
recovery because of the non-convexity problem. We yield a hybrid result as
shown in Figure 4(c). The HMC method is launched on a multiple-point
C
inverse model while we collect the drawn samples after 150 iterations. We
305 calculate the mean value of 100 samples after this burn-in period as the
AC
16
ACCEPTED MANUSCRIPT
PT
RI
(a) The true five-layer model.
U SC
AN
(b) The inverse model by Levenberg-Marquardt algorithm.
M
D
TE
EP
315 5. Conclusion
Experiment results with our synthetic earth model draw a promising
scheme that the statistical inversion by the HMC method is capable to in-
vert the model parameters and reconstruct the earth structure accurately
17
ACCEPTED MANUSCRIPT
PT
fective and accurate solution for solving these inverse problems. Meanwhile,
a more complicated earth model with five, seven or more layer-structures
poses great challenges to the conventional deterministic inversion scheme.
RI
325 The stochastic strategies undoubtedly will become promising players on the
stage of inversions. The proposed HMC method possesses an innovation and
a higher accuracy compared to traditional MCMC sampling methods, which
SC
can meet requirements for the fast-growing industry.
6. Acknowledgement
U
330 This material is based upon work supported by the U.S. Department
of Energy, Office of Science, and Advanced Scientific Computing Research,
AN
under Award Numbers DE-SC0017033.
References
M
[1] M. S. Zhdanov, Inverse theory and applications in geophysics, Vol. 36,
335 Elsevier, 2015.
D
1996.
345 [5] M. S. Bittar, Electromagnetic wave resistivity tool having a tilted an-
tenna for determining the horizontal and vertical resistivities and rel-
ative dip angle in anisotropic earth formations, Google Patents, u.S.
Patent 6,163,155 (Dec. 19 2000).
18
ACCEPTED MANUSCRIPT
PT
tion, SPE Reservoir Evaluation & Engineering 12 (02) (2009) 270–279.
[7] S. Li, J. Chen, T. L. Binford Jr, Using new lwd measurements to evalu-
ate formation resistivity anisotropy at any dip angle, in: SPWLA 55th
RI
355 Annual Logging Symposium, Society of Petrophysicists and Well-Log
Analysts, 2014.
SC
[8] J. Chen, Y. Yu, An improved complex image theory for fast 3d resistiv-
ity modeling and its application to geosteering in unparallel layers, in:
SPE Annual Technical Conference and Exhibition, Society of Petroleum
U
360 Engineers, 2014.
AN
[9] J. Chen, Y. Huang, T. L. Binford Jr, X. Wu, Managing uncertainty
in large-scale inversions for the oil and gas industry with big data, in:
Guide to Big Data Applications, Springer, 2018, pp. 149–173.
M
[10] Ø. Bø, J.-M. Denichou, U. Ezioba, E. Mirto, J. Donley, J. Telford,
365 C. Dupuis, L. Pontarelli, G. Skinner, M. Viandante, Reservoir mapping
while drilling, Oilfield Review 27 (1) (2015) 38–47.
D
19
ACCEPTED MANUSCRIPT
PT
[16] S. Huang, Q. Yang, P. J. Matuszyk, C. Torres-Verdı́n, High-resolution
interpretation of sonic logging measurements using stochastic inversion
with spatial slowness sensitivity functions, in: SEG Technical Program
Expanded Abstracts 2013, Society of Exploration Geophysicists, 2013,
RI
385
pp. 524–528.
SC
carlo, Physics letters B 195 (2) (1987) 216–222.
U
390 geosteering inversion by hamiltonian dynamics monte carlo method, in:
SEG Technical Program Expanded Abstracts 2017, Society of Explo-
AN
ration Geophysicists, 2017, pp. 900–904.
1148–1163.
400 [21] K. Levenberg, A method for the solution of certain non-linear problems
in least squares, Quarterly of applied mathematics 2 (2) (1944) 164–168.
EP
405
20
ACCEPTED MANUSCRIPT
PT
[26] R. M. Neal, Probabilistic inference using markov chain monte carlo
methods, Tech. rep., Department of Computer Science, University of
Toronto Toronto, Ontario, Canada (1993).
RI
415
SC
[28] A. Gelman, D. B. Rubin, Inference from iterative simulation using mul-
tiple sequences, Statistical science (1992) 457–472.
U
420 [29] Q. Shen, X. Wu, J. Chen, Z. Han, Distributed markov chain monte
carlo method on big-data platform for large-scale geosteering inversion
AN
using directional electromagnetic well logging measurements., Applied
Computational Electromagnetics Society Journal 32 (5) (2017) 405–412.
M
[30] L. Murray, Distributed markov chain monte carlo, in: Proceedings of
425 Neural Information Processing Systems workshop on learning on cores,
clusters and clouds, Vol. 11, 2010.
D
TE
C EP
AC
21
ACCEPTED MANUSCRIPT
Highlights
• A stochastic hybrid Monte Carlo method is proposed for geosteering inverse problems.
• The stochastic HMC increases the sampling efficiency compared to other statistical methods.
• The HMC method is capable of searching for the global solution of earth model parameters.
PT
RI
U SC
AN
M
D
TE
C EP
AC