Professional Documents
Culture Documents
Computers & Geosciences: Guofeng Liu, Xiaohong Meng, Zhaoxi Chen
Computers & Geosciences: Guofeng Liu, Xiaohong Meng, Zhaoxi Chen
a r t i c l e i n f o a b s t r a c t
Article history: There are two types of three-dimensional (3D) magnetic inversion methods based on the classification of the
Received 20 March 2012 inversion result, one is the inversion approach that determines a 3D susceptibility distribution that produces
Received in revised form a given magnetic anomaly, and the other is to inverse the source distribution in a purely probabilistic sense,
21 May 2012
in which the inversion results are equivalent physical parameters between þ 1 and 1. The second method
Accepted 22 May 2012
Available online 31 May 2012
is easier and more stable, but obtaining the susceptibility directly to recognize certain lithology is often more
desirable. Furthermore, it is difficult to add an external geological constraint in the second method for
Keywords: reducing the nonuniqueness of magnetic inversion. Herein, we propose an iterative method to inverse the
Magnetic inversion susceptibility based on the second method. The proposed method obtains the perturbation of susceptibility
Probability tomography
by multiplying some susceptibility with the probability tomography result of misfits in observed data and
GPU
forward data given a certain susceptibility model. We present a graphic processing unit (GPU) scheme to
Optimization
tackle an intensive computing problem. The forward and probability function are computed in parallel on the
GPU. Incorporating reasonable parallel strategies and three key optimization steps like memory optimization,
execution configuration optimization and instruction optimization, the 3D magnetic inversion in this paper
on a Tesla C2050 GPU shows greatly improved efficiency compared to serial code on a 2.5 GHz CPU, with a
60-fold increase in speed especially for the large volumes of data. We design a synthetic model with two
prismatic susceptibility anomalies. The inversion result of this model also proves the effectiveness of the
inversion method introduced in this paper.
& 2012 Elsevier Ltd. All rights reserved.
0098-3004/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cageo.2012.05.025
G. Liu et al. / Computers & Geosciences 48 (2012) 86–92 87
unit (CPU) has been developed for high-performance computing. magnetization of the sphere are I and A0 ,so, we have the following
Compute Unified Device Architecture (CUDA) is a parallel pro- expression of a f:, a¼2UlUL nUN mUM, b¼2UmUM lULnUN,
gramming model and software environment provided by NVIDIA c¼2U(lUL nUL), d¼3U(MUNþnUE), e¼3U(MUlþmUL) and f¼2Un
that is designed to overcome the challenge of using a traditional, UN MUm lUL, in that, L ¼ cosI cosA0 , M ¼ cosIsinA0 , N ¼ sinI,
general-purpose GPU while maintaining a shallow learning curve l ¼ cosI0 cosA00 , m ¼ cosI0 sinA10 , n ¼ sinI0 (Guo et al. 2011a).
for programmers familiar with standard programming languages
such as C. Presently, the T10 series NVIDIA GPU C2050 computa- 2.2. The qth rectangular occurrence probability function
tion ability reaches 1.3 TFlops, but is only the size of a normal
video card and consumes just 300 W of electricity per hour We define the total power L associated with DTa on the
(NVIDIA, 2010, 2011). This GPU has been used successfully in surface as
many geophysical parallel computing fields (Zhang et al., 2009; Z
Wang et al., 2010; Shi et al., 2011), they introduces the main L ¼ DT 2a dS ð4Þ
s
problem in many aspects like floating-point errors, micro-struc-
ture of graphic processing unit and so on. Substituting from (2), (4) becomes
Besides the intensive computation of 3D magnetic inversion, Q
X Z
kq T
the current magnetic survey, especially airborne surveys are L¼ DT a Bq ðx,y,zÞdS ð5Þ
4p s
characterized by extremely large volumes of data, all above q¼1
problems need substantive amounts of parallel computing power. As defined by Mauriello and Patella (2005, 2008), the occur-
In this study, we developed GPU/CPU heterogeneous parallel rence probability function of a generic qth rectangular cell, valid
computing to tackle the computing challenge. for 3D tomography is
R
s DT a Bq dS
ZT a ðqÞ ¼ qRffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R 2 ffi ð6Þ
2. Theory of 3D magnetic inversion based on probability 2
s DT a dS s Bq dS
tomography
According to Schwarz’s inequality property, we can write
2.1. The total magnetic field function Z Z Z
DT 2a dSU B2q dS Z ½ DT a Bq dS2 ð7Þ
s s s
In a reference rectangular coordinate system with the
horizontal (x–y) plane and the z axis positive downwards, the the results of the 3D probability tomography satisfy the condi-
subsurface is represented by a 3D array of rectangular cells, each tion:
of them having a variable susceptibility. To simplify the calcula- 1 r Za ðqÞ r 1
tion, the magnetic effect of each cell is approximated by the effect
of a dipole located at its center. For the qth cell with susceptibility Positive values of Za(q) indicate the influence of concentrated
kq, the magnetic anomaly caused by it at a receiving point p on susceptibility at the qth cell to the surrounding background
the ground surface is (Fig. 1) susceptibility, while negative values indicate a susceptibility
deficit.
kq T
DT q ¼ Bq ðx,y,zÞ ð1Þ
4p
2.3. 3D magnetic imaging based on probability tomography
The total magnetic field on p is a sum of all the cell effects
underground: Assuming an initial susceptibility model kq1, the total mag-
XQ netic field at the receiving point p is
kq T
DT a ¼ Bq ðx,y,zÞ ð2Þ Q
4p X kq1 T
q¼1
DT a1 ¼ Bq ðx,y,zÞ ð8Þ
q¼1
4p
Kq is the susceptibility of a rectangular cell, and T is the inducing
field strength of geomagnetic field. The misfit of the observed magnetic field DTa and DT a1 is
1h DT a DT a1 . Their probability tomography results for the generic
Bq ðx,y,zÞ ¼ 5 aðxq xÞ2 þbðyq yÞ2 þ cðxq xÞðzq zÞ qth cell are 1 r ZðDT a DT a Þ ðqÞ r1, indicating the influence or
r i 1
deficit of kq1 compared to the real susceptibility kq at the qth
þdðyq yÞðzq zÞ þ eðxq xÞðyq yÞ þ f ðzq zÞ2 ð3Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rectangular cell. The probability tomography result can be multi-
where r ¼ ðxq xÞ2 þ ðyq yÞ2 þðzq zÞ2 is the distance between plied by a susceptibility value dk to convert the probability value
the center of cell to the receiving point, we define the inclination to susceptibility. The susceptibility can then be added to kq1, to
and declination of the geomagnetic field are I0 and A00 , total make it identical to kq, but it is difficult to initially choose an of
dk, so we include the following iterative procedure:
kqi þ 1 ¼ kqi þ ZðDT a DT a Þ ðqÞ dk ð9Þ
i
The flowchart of the magnetic inversion for the designed Fig. 6. The various memory spaces on a GPU, in that, constant and texture
parallel strategies of the GPU is shown in Fig. 5. memory are read-only memory, others are read-write memory.
90 G. Liu et al. / Computers & Geosciences 48 (2012) 86–92
4. Model test
We designed a 3D susceptibility model to test our method. The Fig. 10. Magnetic field contour of the model. A–A0 is a profile across the magnetic
model length and width is 1200 m 1200 m, the depth is 500 m, field peak.
G. Liu et al. / Computers & Geosciences 48 (2012) 86–92 91
Fig. 11. The quality control profile A–A0 . (a) the observed magnetic data of A–A0 , (b) inversed and observed magnetic data of A–A0 , (c) the model under A–A0 and
(d) inversed model under A–A0 .
Fig. 12. 3D visualization of model (a) and inversion result (b). (b) results using a cutoff value o 0.003.
and the inversion grid is 10 m 10 m 10 m. There are two Tesla C2050 GPU. After low-grain parallel strategies and three key
prismatic susceptibility anomalies in this model; their position optimization steps, the inversion efficiency was greatly improved,
and susceptibility values are shown in Table 1. the contour of the especially for large volumes of data. The effectiveness of the
magnetic field is shown in Fig. 10. The inclination of the magnetic method was confirmed with a model test.
field and the total magnetization of the sphere is 901; their
declination is 01, the inducing field with a strength of 50000 nT.
After 7 iteration, we get the inversion result, the quality Acknowledgments
control of the inversion result across profile A–A0 are shown in
Fig. 11. in this comparison, we can distinguish there are two This work is supported by the National Natural Science
anomalies on the background, for displaying the inversion result Foundation of China under Grant Nos. 41104083 and 41074095,
of the whole lines, the 3D display comparing the original model SinoProb-01 project and the Fundamental Research Funds for the
and inversion results is shown in Fig. 12. After cut the suscept- Central Universities.
ibility under 0.003, the result outlines successfully the magnetic
source distribution at the nearly same location as the designed
model. These results in some way approve that the rational of the References
approach is correct, together with the great improvement of
computing time shown in Figs. 8 and 9, I believe the method of Boulanger, O, Chouteau, M., 2001. Constraints in 3D gravity inversion. Geophysical
this paper can provide a new choice for 3D magnetic inversion. Prospecting 49, 265–280.
Chianese, D, Lapenna., V., 2007. Magnetic probability tomography for environ-
mental purposes test measurements and field applications. Journal of Geophysics
and Engineering 4, 63–74.
5. Conclusion Fullagar, P.K., Pears, G.A., McMonnies, B., 2008. Constrained inversion of geologic
surfaces—pushing the boundaries. The Leading Edge 27, 98–105.
Guo, L,H., Shi, L., Meng, X.H., 2011a. 3D correlation imaging of magnetic total field
We develop a new 3D magnetic inversion method to inverse anomaly and its vertical gradient. Journal of Geophysics and Engineering 8,
the susceptibility distribution directly based on the probability 287–293.
tomography function. This method is easier and more stable than Guo, L,H., Meng, X.H., Shi, L., 2011b. 3D correlation imaging of the vertical gradient
of gravity anomaly. Journal of Geophysics and Engineering 8, 6–12.
those general methods associated with large matrix computation. Li, Y., Oldenburg, D.W., 1996. 3-D inversion of magnetic data. Geophysics 61,
We also implemented this new inversion method on a NVIDIA 394–408.
92 G. Liu et al. / Computers & Geosciences 48 (2012) 86–92
Li, Y., Oldenburg, D.W., 2003. Fast inversion of large-scale magnetic data using Portniaguine, O., Zhdanov, M.S., 2002. 3-D magnetic inversion with data compres-
wavelet transforms and a logarithmic barrier method. Geophysical Journal sion and image focusing. Geophysics 67, 1532–1541.
International 152, 251–265. Patella, D., 1997a. Introduction to ground surface self-potential tomography.
Mauriello, P., Monna, D., Patella, D., 1998. 3D geoelectric tomography and Geophysical Prospecting 45, 653–681.
archeological application. Geophysical Prospecting 46, 543–570. Patella, D., 1997b. Self-potential global tomography including topographic effects.
Mauriello, P., Patella, D, 1999a. Resistivity anomaly imaging by probability Geophysical Prospecting 45, 843–863.
tomography. Geophysical Prospecting 47, 411–429. Pilkington, M., 1997. 3-D magnetic imaging using conjugate gradients. Geophysics
Maurillo, P., Patella, D., 1999b. Principles of probability tomography for natural- 62, 1132–1142.
source electromagnetic induction fields. Geophysics 64, 1403–1417. Pilkington, M., 2009. 3D magnetic data-space inversion with sparseness con-
Maurillo, P., Patella, D., 2001. Gravity probability tomography: a new tool for straints. Geophysics 74, L7–L15.
buried mass distribution imaging. Geophysical Prospecting 49, 1–12. Shi, X,H., Li, C., Wang, S.H., et al., 2011. Computing prestack Kirchhoff time
Maurillo, P., Patella. D., 2005. Localization of magnetic sources underground by a migration on general purpose GPU. Computers and Geosciences 37,
data adaptive tomographic scanner. arxiv:physics/0511192v2, 1–5.
1702–1710.
Mauriello, P., Patella, D., 2008. Localization of magnetic sources underground by a
Vogel, C.R., 2002. Computational methods for inverse problems. Society of
probability tomography approach. Progress in Electromagnetic Research M 3,
Industrial and Applied Mathematics, 37–140p, p.
27–56.
Wang, X., Gao, X., Yao, 2010. Accelerating POCS interpolation of 3D irregular
NVIDIA Corporation, 2011. NVIDIA GPU Computing Developer Homepage, NVidia Inc.
seismic data with graphics processing units. Computers and Geosciences 36,
/http://developer.nvidia.com/object/gpucomputing.html (accessed12.01.10)S .
NVIDIA Corporation, 2010. CUDA C Programming Best Practices Guides 3.2. 1292–1300.
/http://nvidia.com/object/cuda_develop.htmlS. Zhdanov, M.S., 2002. Geophysical Inverse Theory and Regularization Problems.
Oldenburg, D.W., McGillivray, P.R, Ellis, R.G., 1993. Generalized subspace methods Elsevier Science, pp. 76–98.
for large-scale inverse problems. Geophysical Journal International 114, Zhang, J.H., Wang, S.Q., Yao, Z.X., 2009. Accelerating 3D fourier migration with
12–20. graphics processing units. Geophysics 74, WCA129–WCA139.