International Journal of Science and Engineering Applications

Volume 13-Issue 01, 01 - 04, 2024, ISSN:- 2319 - 7560

DOI:10.7753/IJSEA1301.1001

2D Inversion of Magnetic Anomaly data based on Deep

Learning

Haokang Yang Jie Xiong Yicheng Cao

School of Electric Information and School of Electric Information and School of Electric Information and
Electrical Engineering Electrical Engineering Electrical Engineering
Yangtze University Yangtze University Yangtze University
Jingzhou, China Jingzhou, China Jingzhou, China

Abstract: This Magnetic exploration is a geophysical exploration method to study geological structure and mineral resources.
Inversion is an effective method to estimate the horizontal location,depth,and geometry of subsurface geological bodies. To resolve the
problems of traditional inversion methods, such as dependence on initial model and long calculation time, we proposed a 2D inversion
method of magnetic anomaly data based on Deep Learning.With this method, a number of magnetic anomalous body models were
designed to perform forward simulation, which generated sample dateset, firstly; a new convolution neural network(CNN) magnetic
inversion network was designed, secondly; the sample dataset was used to train the network thirdly; and inversion experimental was
performed to evaluate the proposed method lastly. The experimental results show that the proposed method can invert position and
magnetization of magnetic anomaly, with strong learning ability and certain generalization ability, and can solve the magnetic
inversion problem effectively.

Keywords: Magnetic Anomaly; 2D inversion; Deep Learning; convolution neural network(CNN)

component response of an infinite strike and finite depth thick

1. INTRODUCTION plate body is forward modeled, the magnetic anomaly is :
Magnetic exploration is an important geophysical exploration
method, which is based on the magnetic differences of rocks,
minerals, and other media in the crust[1]. It is a geophysical
method that explores geological structures and seeks mineral
resources by observing the changes in magnetic field data.
Magnetic anomaly inversion is an essential method for
quantitatively interpreting magnetic data[2]. Traditional linear
iterative regularization inversion is prone to getting stuck in
local minima, and global optimization inversion methods have
received widespread attention. Nonlinear methods such as
Monte Carlo (MC) [3], simulated annealing (SA) [4], genetic
algorithm (GA) [5], artificial neural network (ANN) [6],
particle swarm optimization (PSO) [7], differential evolution
[8], ant colony optimization (ACO) [9] have been widely used (a) global schematic diagram (b) local schematic diagram
in geophysical inversion.
Figure. 1 2D prism forward modelling schematic diagram
Deep learning has been a research hotspot in the field of
artificial intelligence in recent years, and has been widely F   FeG (1)
applied in fields such as computer vision and natural language
processing [10]. In recent years, deep learning has been
 r r 
introduced into the field of geophysical inversion. However, G  2 cos I sin  ln 2 3  sin I 1  2   3  4  
the applications in the field of gravity and magnetic inversion  r1  r4 ,
are limited [2, 11-12]. Further research is needed on how to
design suitable network structures [12]. Therefore, this article where  is the magnetic susceptibility, Fe is the strength of
proposes a 2D inversion method of magnetic anomaly based
on Deep Learning, in order to achieve better inversion results.
the geomagnetic field, I is magnetic inclination angle, β Is
the magnetic north angle, ri is the distance, i is the angle
2. Methodology
with respect to the horizontal direction.
2.1 Forward and Inversion modelling
(1) Magnetic Forward modelling Forward modelling can be also described as following:
Divide the underground into many regular and uniform cell
sizes (the blue part represents the mesh of the division, and d  Gm (2)
the yellow part represents the anomaly geological body).
When the residual magnetization of a geological body is very
small, residual magnetization may usually not considered. At
this time, the magnetization intensity M is equal to the
induced magnetization intensity Mi. When the vertical

www.ijsea.com 1
 International Journal of Science and Engineering Applications
Volume 13-Issue 01, 01 - 04, 2024, ISSN:- 2319 - 7560
DOI:10.7753/IJSEA1301.1001

Where m is geological body model, represented as

m   m1 , m2 , m3 ,, mL  , d is observed data present-
T

ed as d   d1 , d 2 , d3 ,, d N  ,
T
G is the kernel matrix.

(2) Magnetic Inversion modelling

Inversion modelling is the process of finding a magnetization
model of a geological body and making its predicted magnetic
anomaly fitting the observed magnetic anomaly as better as
possible. Therefore, the data fitness objective function is
defined as following: Figure. 3 Inversion Network structure

f(𝐦) = 𝐝 − G𝐦 2
→0 (3) 3. Experimental Results
3.1 Training Dataset
where d is observed magnetic anomaly data, m is the When setting up the training set samples, we divided the
predicted model, Gm is the predicted data. observed underground space into 800 (20 rows × 40 columns)
rectangular cells, each cell size is 25m × 25m. In the forward
modeling, 101 ground observation points are set with a
2.2 Inversion Method spacing of 10m between them. The training dataset contains
Our 2D inversion method of magnetic anomaly data gravity anomaly bodies of different sizes, shapes (squares,
based on deep learning can be described as Figure. 2. The rectangles), and positions, such as 3 × 3 (75m × 75m), 4 × 4
inversion include four steps: (1) design various geological (100m × 100m), 5 × 5 (125m × 125m), 3 × 6 (75m × 150), 6 ×
body magnetization intensity models and compute the 3 (150m × 75m), 4 × 8 (100m × 200m), 8 × 4 (200m × 100m).
To prevent the influence of edge data on experimental errors,
observed data by forward modelling; (2) combining the
we discard the data from two units near the edge of the
models and observed data as training dataset to train the CNN; underground grid and only retain the data from the middle
(3) when training finished, a trained CNN is obtained; (4) part. Due to the linear relationship between magnetic
input the new observed data input CNN, obtain the predicted susceptibility and magnetic anomalies, the network can easily
model. learn this linear relationship. Therefore, only two different
values (50A/m and 100 A/m) are set for the magnetic anomaly
bodies. 5542 sets of underground magnetic models and their
corresponding surface observation magnetic anomaly data
were generated through forward modeling of anomalies.
Figure 4 shows six representative forward models in the
dataset.

Figure. 2 Inversion method diagram

Figure. 4 Differential forward models.
2.3 Network Structure
According to the observation points and the dissected 3.2 Training Parameters Setting
underground space, the input layer of the network contains Validate set contains 1109 samples extracted from the
101 neurons, and the output layer contains 800 neurons. There original 5542 data samples, while the training set contains the
are three types in between: Convolutional Layer (CONV), remained 4433 samples. The settings for other network
Maximum Pooling Layer (POOL), and Fully Connected Layer. parameters are shown in Table 1.
There are four convolutional layers, each of which is followed Table. 1 Sample dataset and network parameters setting
by a pooling layer. The last hidden layer is a fully connected
Name Values
layer containing 1600 neurons, and the input of this fully
connected layer is obtained by flattening the output matrix of Sample Training set 4433
the last convolutional pooling layer. Specifically, the four dataset Validate set 1109
convolutional layers have kernel sizes of 1 * 5, 1 * 4, 1 * 3,
and 1 * 3, each with a step size of 1. The number of channels Learning rate η=0.0001
is 10, 20, 40, and 80, respectively; All maximum pooling Activate function ReLU
layers have a pooling kernel size of 1 * 2 and a step size of 1. Network
The inversion convolution neural network(CNN) illustrated as Optimizer Adam
settings
figure 3. L2 regular λ=0.01
Dropout 1

www.ijsea.com 2
 International Journal of Science and Engineering Applications
Volume 13-Issue 01, 01 - 04, 2024, ISSN:- 2319 - 7560
DOI:10.7753/IJSEA1301.1001

Training Iteration 50000 significant difference in the inversion magnetization intensity,

process Batch size 1000 which can reach 100A/m in some places. The magnetization
intensity distribution of vertically separated prisms has a low
3.3 Single anomaly body experiment vertical resolution, shown in Figure 6(f), which may be
the inversion results of the samples in the validate set connected to shallow layers and fail to separate well.
are shown in Figure 5, with different colors representing
different densities, a white box displaying the true position of
the model, a blue dashed line representing the actual gravity
anomaly, and a red solid line representing the predicted
theoretical magnetic susceptibility anomaly obtained from the
forward modeling of the underground magnetic susceptibility
model. It could be found that the predicted magnetic
susceptibility of individual positions of some gravity anomaly
bodies is lower than the actual value, and there is a small
amount of scattered anomaly information around them, but it
does not affect the determination of the center position and
overall shape of the magnetic anomaly bodies.

Figure. 6 Inversion results of combination anomaly bodies

4. CONCLUSION
This article proposes a deep learning based magnetic
anomaly inversion method. Firstly, a CNN inversion network
is designed for magnetic anomaly inversion. Then, a large
number of magnetic anomaly body models are designed and
labeled datasets are obtained through forward calculation.
Figure. 5 Inversion results of single anomaly body Secondly, the CNN inversion network is trained using this
labeled dataset. Finally, the magnetic anomaly data is input
3.4 Combination anomaly bodies into the trained CNN inversion network to directly obtain the
experiment inversion results. The experimental results show that this
Figures 6 shows the CNN inversion results of six method can invert the position and magnetic susceptibility of
combination anomaly bodies models. The inversion results of anomalous bodies correctly, has good generalization ability
the models show high resolution. For example, single prism and noise resistance, and can effectively solve the problem of
and two parallel rectangular prisms can be restored very well, magnetic anomaly inversion. Deep learning has been
which are shown in Figures 6 (a) (c). Due to the presence of preliminarily applied in magnetic anomaly inversion, and the
false anomalies, the boundaries of stepped prisms become next works will focus on how to design better network
difficult to distinguish, such as one stepped prism and two structures to solve three-dimensional magnetic anomaly
stepped prisms, which are shown in Figure 6(b)(d)(e). inversion problems and joint inversion problems.
Moreover, there is distortion phenomenon, and there is a

www.ijsea.com 3
 International Journal of Science and Engineering Applications
Volume 13-Issue 01, 01 - 04, 2024, ISSN:- 2319 - 7560
DOI:10.7753/IJSEA1301.1001

5. REFERENCES [8] XIONG Jie, MENG Xiaohong, LIU Caiyun, PENG

[1] Adagunodo, T. A., L. A. Sunmonu, and A. A. Adeniji, Miao, The induction logging inversion based on particle
An overview of magnetic method in mineral swarm optimazaiton[J], Geophysical and Geochemical
exploration[J], Journal of Global Ecology and Exploration, 2012, 36(3):448-451.
Environment, 2015, 3(1):13-28. [9] Bouchaoui, Lyes, et al., Vertical electrical sounding data
[2] XUE Ruijie, XIONG Jie, ZHANG Yue, WANG Rong, inversion using continuous ant colony optimization
Magnetic Anomaly Inversion Based on Convolutional algorithm: A case study from Hassi R'Mel, Algeria[J],
Neural Network[J], Geoscience, 2023, 37(1):173-183. Near Surface Geophysics, 2022, 20(42): 419-439.

[3] WEI C, LI X F, ZHEN X D. The group search based [10] Dong, Shi, Ping Wang, and Khushnood Abbas, A survey
parallel algorithm for the serial Monte Carlo inversion on deep learning and its applications[J], Computer
method[J]. Applied Geophysics,2010,7(2), 127−134. Science Review, 2021, 40: 100379.

[4] H Luo, Y Li, H Li, X Cui, Z Chen, Simulated annealing [11] ZHANG L Z, ZHANG G B, LIU Y, et al. Deep Learning
algorithm-based inversion model to interpret flow rate for 3-D Inversion of Gravity Data[J]. IEEE Transactions
profiles and fracture parameters for horizontal wells in on Geoscience and Remote Sensing, 2022, 60,
unconventional gas reservoirs[J], SPE Journal, 2021, 5905919。
26(4):1679-1699. [12] Wang Ｒ， Xiong J， Liu Q， et al． Inversion of gravity
[5] da Conceição Batista, J., & Sampaio, E. E. S., anomalies based on a deep neural network[J],
Magnetotelluric inversion of one-and two-dimensional Geophysical and Geochemical Exploration， 2022， 46(
synthetic data based on hybrid genetic algorithms[J], 2) :451－458．
Acta Geophysica, 2019, 67, 1365-1377. [13] Zhang ZH, LiaoXL, Cao YY, et al. Joint gravity and
[6] Yadav, Apurwa, Kriti Yadav, and Anirbid Sircar, gravity gradient inversion based on deep learning[J],
Feedforward neural network for joint inversion of Chinese J. Geophys. (in Chinese), 2021, 64(4):1435-
geophysical data to identify geothermal sweet spots in 1452.
Gandhar, Gujarat, India[J], Energy Geoscience, 2021,
2(3): 189-200.
[7] XIONG Jie, LIU Caiyun, ZOU Changchun, The .
induction logging inversion based on particle swarm
optimazaiton[J], Geophysical and Geochemical
Exploration, 2013, 37(6):1141-1145.

www.ijsea.com 4