See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/343986020

Deep learning seismic full waveform inversion for realistic structural models

Article in Geophysics · August 2020

DOI: 10.1190/geo2019-0435.1

CITATIONS READS
93 1,360

6 authors, including:

Senlin Yang Yuxiao Ren

Shandong University Shandong University
25 PUBLICATIONS 774 CITATIONS 36 PUBLICATIONS 1,073 CITATIONS

SEE PROFILE SEE PROFILE

Peng Jiang Yangkang Chen

Shandong University University of Texas at Austin
67 PUBLICATIONS 1,999 CITATIONS 484 PUBLICATIONS 14,701 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Yangkang Chen on 31 August 2020.

The user has requested enhancement of the downloaded file.

 Deep learning seismic full waveform inversion for realistic

structural models

Bin Liu∗†‡ , Senlin Yang∗ , Yuxiao Ren∗ , Xinji Xu† , Peng Jiang∗ and Yangkang

Chen§

∗ School of Qilu Transportation,

Shandong University,

Jinan, Shandong Province, China, 250061

† Geotechnical and Structural Engineering Research Center,

Shandong University,

Jinan, Shandong Province, China, 250061

‡ Data Science Institute,

Shandong University,

Jinan, Shandong Province, China, 250061

§ School of Earth Sciences,

Zhejiang University, Hangzhou

Zhejiang Province, China, 310027

Corresponding author: [email protected]

(August 30, 2020)

GEO-2020

Running head: DL seismic FWI

1
 ABSTRACT

Velocity model inversion is one of the most important tasks in seismic exploration. Full

waveform inversion (FWI) can obtain the highest resolution in traditional velocity inver-

sion methods, but it heavily depends on initial models and is computationally expensive.

In recent years, a large number of deep learning based velocity model inversion methods

have been proposed. One critical component in those deep learning based methods is a

large training set containing different velocity models. We propose a method to construct

a realistic structural model for deep learning network. Our P-wave velocity model build-

ing method for creating dense-layer/fault/salt body models can automatically construct

a large number of models without much human effort, which is very meaningful for deep

learning networks. Moreover, to improve the inversion result on these realistic structural

models, instead of only using the common-shot gather, we also propose to extract features

from the common-receiver gather as well. Through a large number of realistic structural

models, reasonable data acquisition methods, and appropriate network setups, a more gen-

eralized result can be obtained through our proposed inversion framework, which has been

demonstrated to be effective on the independent testing data set. The results of dense-layer

models, fault models, and salt body models are compared and analyzed, respectively, which

demonstrates the reliability of the proposed method and also provides practical guidelines

for choosing the optimal inversion strategies in realistic situations.

2
 INTRODUCTION

Accurate seismic velocity model inversion from seismic data is an important task in

seismic exploration for a variety of applications. The velocity model plays an essential role

in high-quality and high-resolution seismic exploration, and accurate velocity model can

provide a better research basis for reverse-time migration (Baysal et al., 1983), prestack-

depth migration (Mittet et al., 1995), and other imaging methods (Bunks et al., 1995).

With the increasing complexity of seismic exploration conditions and requirements, the

accuracy of velocity estimation becomes more demanding. To obtain a good prediction of

the true velocity model, many methods have been developed from different perspectives,

such as normal moveout correction (NMO) based velocity analysis (Dunkin and Levin, 1973;

Alkhalifah and Tsvankin, 1995), wave-equation tomography (Woodward, 1992) and full

waveform inversion (FWI) (Lailly and Bednar, 1983; Tarantola, 1984) methods. In recent

years, with the development of machine learning, more and more data-driven methods are

used in velocity model inversion.

In view of the importance of the velocity model in seismic exploration, researchers have

devoted great efforts on velocity inversion methods (Cohen and Bleistein, 1979). Lailly and

Bednar (1983) and Tarantola (1984) first proposed the idea of seismic full waveform inversion

based on the generalized least-squares criterion, which provides an overall framework for

seismic velocity inversion. Different from traditional tomography methods, FWI requires

high-accuracy modeling of wave propagation so that it can make full use of the kinematics

and dynamics information of prestack seismic wave field. In addition, due to the highly

non-linear characteristics of FWI, it is strongly dependent on initial models and easy to fall

into a local minimum. To solve the local minimum problem, Bunks et al. (1995) propose

3
the dense-scale full-waveform inversion in the time domain, where seismic data information

at different frequencies are used to improve the inversion effect. After that, FWI methods

in the frequency domain (Pratt and Worthington, 1990; Pratt, 1999; Pratt and Shipp, 1999;

Hu et al., 2009b) and the Laplace domain (Shin and Cha, 2008; Shin and Ho Cha, 2009)

have been proposed with satisfactory results achieved. In addition, the directional total

variation (DTV) method has been introduced recently to solve the local minimum problem

(Qu et al., 2019). With the intensive study, FWI methods has been further expanded

in various aspects, including viscoelastic media (Yang et al., 2009), joint inversion (Khan

et al., 2010; Hu et al., 2009a), and gradually been applied in field data (Sirgue et al., 2010).

However, the limitations of FWI methods, such as the impact of initial models, still exist,

and new methods are needed to solve these problems (Chen et al., 2016).

With the continuous development of machine learning, data-driven methods provide

a new idea for velocity model inversion. Among them, deep learning has become one of

the most important research topics, and more and more fields are beginning to introduce

deep learning methods to solve related problems. After decades of development, neural

networks have evolved from the initial form of neurons to the recent architecture of deep

neural networks. The concept of neural networks dates back to the study by McCulloch and

Pitts (1943), which focuses on the neural computing method. Rosenblatt (1958) creatively

proposes the concept of the perceptron, which opened up the research of neural network al-

gorithms. Rumelhart et al. (1988) propose the back-propagating neural network. However,

because of the huge requirements of computational resources, the neural network study

mostly stayed at the theoretical stage. With the great improvement of computing ability

in the 21st century, the neural network algorithms have returned to the research focus via

the publication of a series of papers (Hinton et al., 2006; Lecun et al., 2015; Hinton et al.,

4
2012a). With the development of deep learning and the outbreak of related technologies,

more and more fields begin to use this method to solve problems (Deng et al., 2014), such as

computer vision (Voulodimos et al., 2018), medical diagnosis (Esteva et al., 2017), speech

recognition (Hinton et al., 2012a).

In the seismic exploration community, research began to use deep learning methods and

achieved better results than traditional methods. Röth and Tarantola (1994) first apply the

method of neural network to invert 1D velocity model from seismic data, which confirmed

the applicability of neural network to velocity model inversion. Moseley et al. (2018) apply

the WaveNet to conduct 1D velocity model inversion based on depth-to-time conversion

data. Araya-Polo et al. (2018) achieve velocity model reconstruction through convolutional

neural network (CNN) for a velocity spectrum cube calculated from prestack data. For

layer model and fault model, Wu and Lin (2018) use a CNN with an encoder-decoder, called

InversionNet, to achieve corresponding velocity model building. Yang and Ma (2019) use

fully convolutional neural networks (FCN) to rebuild velocity models, especially salt model,

from prestack data with Gaussian noise. In particular, they further trained SEG data sets by

using a transfer learning method, and the result was better than that from FWI. The above-

mentioned deep neural networks are based on the application of existing network methods

and applied to the seismic data set. Based on in-depth analysis of the characteristics of

seismic data, Li et al. (2020) further design and optimize CNN and fully connected network,

then proposed SeisInvNet, which achieved better results than InversionNet. Some methods

for improving FWI using neural network are proposed, and the computational efficiency and

inversion results are significantly improved(Sun et al., 2020; Ren et al., 2020). Moreover,

in the research of seismic data processing, deep learning has been successfully applied, such

as seismic data denoising (Yu et al., 2018; Chen et al., 2019; Saad and Chen, 2020), fault

5
identification (Wu et al., 2019), lithology prediction (Shi et al., 2019; Zhang et al., 2018),

geological structure classification(Li, 2018), arrival picking (Tsai et al., 2018; Yuan et al.,

2019; Zhang et al., 2020). It is also expected that there will be more deep learning programs

applied in geophysical exploration in the future, such as the mineral exploration (Malehmir

et al., 2012), the forward geological prospecting in tunneling (Li et al., 2017), the four-

dimensional data monitoring (Liu et al., 2020b). In addition to seismic exploration, other

exploration methods also applied deep learning to realize data processing and interpretation

(George and Huerta, 2018; Puzyrev, 2019; Nurindrawati and Sun, 2019; Liu et al., 2020a).

In this study, we develop a method to build realistic structural models, i.e., dense layer

models, fault models, and salt body models, and a complete framework for P-wave velocity

model inversion using the deep neural network. The research consists of two major aspects.

First, to obtain as many complex models as possible, the model building process is designed.

18,000 dense-layer/fault/salt body models are acquired, which provides sufficient data for

the network training. SeisInvNet is chosen and improved for more complex models. Based

on the generated velocity models and the corresponding dense-shot data, new deep neural

network is trained to approximate the nonlinear mapping from the data to the model.

Through the trained network, the velocity model can be obtained by directly inputting

seismic data. Compared with traditional methods, our method achieves better results in

computing efficiency and accuracy and has a certain degree of generalization. Finally, to

compare with the original SeisInvNet, we analyze the results by four evaluation criteria,

including M AE, M SE, SSIM and M SSIM . From the analysis, the proposed method

demonstrates consistent superiority compared to the original SeisInvNet. Furthermore, the

prediction results of the fault model are carefully studied, and future research plans are

discussed.

6
 METHOD

Problem definition

To obtain the underground geological information, we usually place artificial sources to

excite seismic wavefield and use receivers on the ground to record seismic waves. In this

paper, synthetic data are modeled based on the acoustic wave equation in the time domain:

∂2u ∂2u ∂2u

 
= a2 + 2 + f (x, y, t), (1)
∂t2 ∂x2 ∂y

where a denotes the wave velocity and u denotes pressure, i.e., the acoustic wave field. x

and y denote spatial coordinates, and t is time, f (x, y, t) is the function of source.

For velocity model inversion, the velocity model V is inferred by the corresponding

observation data D, which is an inversion problem and can be described as:

V = L−1 (D). (2)

For FWI methods, the velocity model is iteratively optimized by gradients derived from

an objective function. The velocity model is usually converged to a local minimum. In con-

trast, deep learning (DL) methods learn the nonlinear mapping by optimizing the network

parameters and usually require a lot of data-model pairs to train the network. The perfor-

mance of deep learning methods heavily depends on the training data set and the network

design. According to the latest research (Kawaguchi, 2016; Choromanska et al., 2014), the

convergence point deep learning methods could reach is closer to the global minimum.

Given input data D, the deep neural network can predict a corresponding model V,

so the fit error between the prediction V and the model V can be computed, then the

gradients are derived to update the network parameters. In this way, the trained network

7
is obtained, i.e., the nonlinear mapping F from time-series data to the model is built as:

V = F(D, θ), (3)

where θ denotes the network parameters.

Figure 1 shows the workflow of velocity model inversion based on deep learning, where

the velocity model V of the size [H × W ] serves as the output target (also called label),

and the corresponding seismic data D as the input. The size of input data is [S × R × T ],

where S, R and T represent the number of shotes, receivers and the time step recorded

respectively. During the training process, we can calculate the misfit error between the

network output V and the label V, and update the network parameters by gradient back-

propagation. After multiple epoch iterations, the error will converge to a sufficiently small

value, and the network parameters can be determined. During inference, given the seismic

data, the velocity model can be predicted through the network.

Constructing dense-layer, fault, and salt body models

Generally speaking, most research on DL velocity model inversion follow the supervised

learning paradigm, where a large quantity of data with the target labels are indispensable.

Due to the data-driven characteristics, the performance of DL methods heavily depends on

the data used during training. Thus, for DL velocity model inversion, the testing model

design is crucial for bringing out the nonlinear mapping ability of neural networks. Specif-

ically, the reasonable model design can help the trained neural network more likely to be

applied in realistic situations. Some studies (Wu and Lin, 2018; Yang and Ma, 2019; Li

et al., 2020) have collected their own data sets, including models with faults, salt bodies,

and layered subsurface. However, most of these works design models only following some

8
simple rules, which make the models neither realistic nor complex enough. To obtain a

more realistic structural velocity model, we further propose a new scheme of designing the

dense-layer, fault, and salt body models.

To design a dense-layer model with one or two faults or salt bodies, we first generate a

dense-layer structural model, and then randomly add faulting structures on it. For dense-

layer models, the difficulty is how to ensure the continuity and variability of each layer and

increase the amount of underground media in the limited depth and exploration resolution.

We generate the dense-layer models following the steps as follows:

1. Randomly generate a curve as the first interface of the model.

2. Iteratively generate curves by making some adjustments according to the upper inter-

face to ensure no drastic change between two adjacent interfaces and keep realistic.

3. Fill P-wave velocity value to the media between two adjacent interfaces, following the

criterion that the deeper media corresponds to a larger velocity value.

The target inversion model size of this paper is set to be 100 × 100. Around the model,

there is an extra 20-grid absorption boundary at the left, right, and bottom sides.

We define a function formed by multiple trigonometric and linear equations to generate

continuous, fluctuant, and complex curves. The function and separate equations are shown

as follows:
y = y1 + y2 + y3 ,
   
x x
y1 = a1 sin + θ1 + a2 (sin + θ2 ) 2 ,
2πT1 2πT2
  (4)
x
y2 = a3 (sin + θ3 )i , i = 1, 2, 3,
2πT3

y3 = r × x + c0 ,

9
where aj , Tj , and θj , j = 1, 2, 3, are parameters for different trigonometric equations, r and

c are constants to control tilt and depth of interface, respectively. We let a1 > a2 > a3 and

T1 < T2 < T3 , while the θs are given randomly, so y1 is the dominant part of the curve. To

make the trends between the two adjacent layers similar, the y1 for each interface should

be adjusted, compared with the previous one. y2 is assigned in the way that its period

and magnitude are re-given at each interface individually. Finally, as for y3 , c and r are

assigned to control the depth and tilt of the current interface, respectively. Through the

above definitions, we could obtain multiple interfaces with the same trend, as shown in

Figure 2.

To further weaken the smoothness of the curve caused by the trigonometric term in the

function 4 and make the model more realistic, we select some discrete points from each

curve and form the new curve by reconnecting these points. In this way, the randomness of

the interface is also increased. This process is illustrated in Figure 3.

After all the interfaces are determined, we set the P-wave velocity value for each layer

from top to bottom, and the value range is [1500, 4000] m/s. Moreover, according to the

consolidation effect of the real earth media, the velocity value of each layer is set to be

increasing with layer depth. In our implementation, we set the velocity difference between

adjacent layers greater than 200 m/s. According to the required maximum wave velocity

vmax , the wave velocity of the previous layer of media vupper , the number of next media ln,

and the random term r, the wave velocity range of each layer of media is selected. The

range of velocity is as follows:

[200, r × (vmax − vupper )/ln]. (5)

In Figure 4, we carry out a statistical analysis on the velocity value changes for different

10
kinds of layer models (five-to-nine layers). The trend is the same for both kinds of layer

models that with the increase of depth, the velocity value increases. Through the above

curve design, interface setting, and velocity assignment processes, the realistic structural

models could be constructed. We finally generate dense-layer models of five-to-nine layers,

each with 1000 models in this way.

Based on the dense-layer model design, the fault setting is further studied. We use the

dense-layer model and the line of fault l(x) as input, and generate a fault model in a random

position. l(x) can be defined as a diagonal line, and can also be set as a curve, such as a

parabola. The procedure is as follows:

1. Randomly generate the starting position P of the fault and the amount of movement

in the vertical direction ∆h.

2. Randomly select the moving part, i.e., the hanging wall or the footwall and the moving

distance.

The randomness ensures the possibility of a fault appearing at any position, which

improves the model complexity. However, in some cases, the reflected wave information

from the fault is difficult to receive, which affects the prediction result. The impact of faults

in models to the inversion results is discussed in the Discussion section. Consequently, for

each point [h0 , w0 ] of the moving part, the updated position [h∗ , w∗ ] can be calculated by

the equation:

[h∗ , w∗ ] = [h0 + ∆h, w0 ± ((h0 + ∆h) − l(h0 ))] . (6)

It is worth noting that sometimes, the velocity value at the boundary positions will be

missing, and needs to be filled. The workflow of the fault setting is shown in Figure 5.

11
To further enrich our model and improve the complexity of the model, we further designed

multiple faults following the similar steps.

We further design the salt body model based on the dense-layer model. The salt body

can be considered as a dense stratum rising upward from the bottom, while salt body

would fluctuate as it flows into the upper stratum. We use Gaussian function to simulate

the gradual fluctuation:

2 /2c2
f (x) = ae−(x−b) , (7)

where a, b, and c are amplitude, center, and standard deviation, respectively. To simulate

the influence of the salt body on the dense layers, where the deeper layer will have a larger

deformation, a set of Gaussian curves is generated randomly, as illustrated in Figure 6.

The original layered model fluctuates according to this set of Gaussian curves, simulating

the invading influence of the salt body. Finally, a new rock mass is designed through a

parabola with random parameters and given a wave velocity in the range [4400, 4500] m/s.

The construction process of the salt body is shown in Figure 6.

According to the proposed design method for dense-layer, fault, and salt body models,

we have generated 18,000 sets of models in batches (each type contains 6000 models) for

the next step of network training. Some of them are shown in the Figure 7.

Deep Learning Network framework

Another research focus in this work is the network design for seismic inversion. For the

more complex velocity models generated in this work, a stable and high-resolution deep

neural network is needed. Here, we select SeisInvNet (Li et al., 2020) as our backbone, and

we further modify and adapt it to meet our needs. We first introduce the four parts of the

12
network and explain how to implement the end-to-end inversion process that maps data to

the corresponding model. Then, our improvements are carefully presented and discussed.

The network architecture and modifications are as follows:

1) Encoder: For seismic data, considering the weak correspondence between seismic

data and velocity model, there are some troubles when using convolutional neural networks

or fully connected networks. One of the main advantages of SeisInvNet is to analyze and

extract the characteristics of seismic data, especially in the Encoder. The convolution

method is used to extract global information and domain information from the common-

shot gathers, and the one-hot vector (with a unit equals one while others leave zero) is used

to label the trace, which increases the single trace information and provides the possibility

of introducing a fully connected network. For seismic data of size [S × R × T ], through two

sets of convolutions of S common-shot profiles [R × T ], the global information and neighbor

information are obtained, whose sizes are [R × C] and [R × T ], respectively. Besides, the

source and receiver location information are also encoded by a one-hot vector with a length

of R + S. In this way, each trace can be encoded to obtain data of size T + R + S + C,

which contains more information than the previous single trace.

Improvement: To extract the effective features of data as much as possible, in this work,

R common-receiver gathers [S × T ] are further considered. That is, we obtain data informa-

tion from two aspects, common shot points, and common receiver points. Correspondingly,

the global information will have sources from two profiles. Specifically, the convolution op-

eration is performed on two kinds of seismic data gathers (i.e., the common-shot gather and

the common-receiver gather), respectively, and two vectors with a length of C are obtained.

In the same way, the neighboring information also comes from the neighboring traces of the

two profiles and then is linearly weighted as the final neighbor information with a length of

13
R.

Through Encoder, the size of each seismic trace is increased to T + S + R + C + C,

which is slightly larger than that in the original SeisInvNet. In this way, each trace contains

more valid information, which provides more advantages for the fully connected network.

Figures 8a and 8b show the difference between SeisInvNet and the proposed method on

the Encoder part, respectively.

2) Generator: Given the encoded S ×R traces, the main purpose of this part is to obtain

S × R feature maps through a fully connected network. For problems of mapping a time

series to a spatial sequence, using a fully connected network is considered the most direct

and efficient way. We extract the feature of each encoded trace, whose size is [w × d] for

model construction.

3) Decoder: In this part, the generated S × R features are decoded to get the velocity

model V similar to the decoder part of many networks (Simonyan and Zisserman, 2014).

The dropout (Hinton et al., 2012b) is used to ensure that the network does not depend on a

particular feature. By randomly dropping some features from a certain channel, the results

will not heavily depends on certain feature. In this way, it is also possible to achieve good

inversion results even after missing certain data, which will be discussed later.

4) Loss function: The loss function determines the direction of approximating to the

velocity model. Mean Squared Error (MSE) is widely used in velocity inversion literature.

In the network, the misfit between the predicted model and the actual model is calculated

to derive gradients and update the network parameters. The MSE loss function is defined

as follows:
H×W

i
 1 X  2
LM SE V , Vi = Vki − Vki . (8)
H ×W
k=1

14
 Among them, Vi and Vi are the prediction of the i-th model and the corresponding

ground truth.

Multiscale Structural Similarity (M SSIM ): Structural similarity SSIM , and M SSIM

could better represent the structural similarity between two figures than M SE, and have

been widely used in many computer vision tasks. Here, to better optimize the structural

characteristics of the prediction model, M SSIM is also used as the loss function. In previ-

ous work (Li et al., 2020), the advantage of this kind of loss functions has been demonstrated.

The equation of SSIM is expressed as follows:

(2µx µy + C1 ) (2σxy + C2 )
SSIM(x, y) =

, (9)
µ2x+ µ2y + C1 σx2 + σy2 + C2

where x and y represent two corresponding windows in two images respectively. µx and µy

are averages of x and y, and σx and σy are variances of x and y, while σxy is the covariance

of x and y. This metric measures the similarity of two image windows. The value ranges

from 0 to 1, and the closer the value is to zero, the lower the similarity. With SSIM , the

loss term LM SSIM in this paper is defined as:


i
 1 X H×W
X 
i

LM SSIM V , Vi = 1 − i
λr . SSIM Vx(k,r) , Vy(k,r) , (10)
H ×W
r∈R k=1

where V is the model (image) and Vy(k,r) is a window of V that centers in k with width of

r, the definition is similarly for V. Thus, this equation computes the similarity by summing

the SSIM over windows at all the position and width. λr is the weight for SSIM of window

with width r. Specific M SSIM parameter details are referred to in Wang et al. (2003).

Figure 9 shows the architecture of the network.

Through the calculation of the above two loss functions, we sum them and get the final

loss function:
     
i i i
LSU M V , Vi = LM SE V , Vi + LM SSIM V , Vi . (11)

15
 EXPERIMENT AND RESULTS

Dataset Setting

Through the above model construction method, we obtain 18,000 dense-layer fault and

salt body models (of size 140×140) in total. Each model has five to nine layers. In the

numerical simulation of the acoustic wavefield, the size of each grid is defined as 10m×10m.

That is, the inversion region is 1km×1km for a model size of 100×100 after removing the

absorbing boundary. As for the observation system, 20 seismic sources and 32 receivers

are placed uniformly and symmetrically on the surface, respectively. The interval between

each source is 50m (five gridpoints), and the interval between each receiver is 30m (three

gridpoints).

A pseudospectral method (Kosloff and Baysal, 1982; Furumura et al., 1998; Virieux

et al., 2011) is used to simulate the seismic wave propagation. The Ricker wavelet with the

dominant frequency of 20Hz is excited as seismic source. The sampling interval is 1 ms, and

the data of the first 1000 time steps are recorded. We present some representative seismic

data and the corresponding velocity model in Figure 10. Finally, 18,000 model-data pairs

are divided into the training set, validation set, and test set in a ratio of 10:1:1. They are

used to train the network, validate performance to save the optimal parameters, and test

the network accuracy, respectively.

Network parameter setting

The setting of network hyperparameters also plays a key role in achieving a good inver-

sion effect. Especially for the mapping between the seismic data and the velocity model,

due to the large number of parameters and the weak mapping relationship, an advantageous

16
method is needed to prevent the gradient from disappearing and overfitting. Based on our

previous research and experience, we have selected the appropriate network parameter to

take advantage of the network. We select Adam optimizer (Kingma and Ba, 2014) with

batchsize of 36 and the initial learning rate of 5×10−5 to optimize our network and SeisIn-

vNet. For each layer of the network, including the convolutional layer and fully connected

layer, the Rectified Linear Unit (ReLU) activation and batch normalization (BN) are ap-

plied, who are as one of the most commonly used techniques in deep learning. The ReLU

activation function formula is as follows:

f (x) = max(x, 0). (12)

This activation function is simple and fast to calculate. Similarly, this derivative is very

simple, when the value x is greater than zero, it is 1, which can effectively avoid the problem

of gradient disappearance and gradient explosion. BN is demonstrated to be helpful in the

network optimization process, and improve the predictability and stability, to achieve the

purpose of making the network training process faster (Santurkar et al., 2018).

The dropout ratio is 0.2 used in the Decoder. All the hyperparameters of the network

are shown in Table 1. We train the network for 200 epochs in total. After each epoch of

training, one validating process is carried out on the validation set. The parameters perform

best on the validation set are saved for inference.

Result analysis

In this subsection, we evaluate and compare the results of the two networks under differ-

ent configurations, including visual comparison and quantitative comparison with different

metrics, such as MAE, MSE, SSIM, and MSSIM. All the scores are calculated after velocity

17
model normalized.

The loss curves on the training set and validation set with respect to the epoch iteration

are shown in Figure 11. It can be seen that the loss curves for LM SE and LM SSIM decrease

monotonously and converge to a very small value on both training and validation sets.

However, our network is more stable and reliable on the validation set than the original

SeisInvNet. The minimum values of these curves are shown in Table 2, which demonstrates

the superiority of our network. Considering that the realistic structural model includes

various models such as salt body, this may be the main reason for the large fluctuation of

loss function curve on the validation set.

Some ground truth and prediction results are shown in Figures 12- 14. We observe that

for relatively simple dense-layer models, the prediction results are very accurate, i.e., almost

identical to truth, while the results get slightly worse for the fault models, sometimes the

fault location is not clear. Especially for the new models with multiple faults, the higher

degree of model complexity affects the amount of effective information collected in the data.

Lack of data leads to poor agreement in the multi-fault models. Furthermore, we provide

the statistics for dense-layer models, fault models, and salt body models on the test set in

Table 3. We choose four metrics to obtain these statistics, which are: M AE, M SE, SSIM ,

and M SSIM (the better network will have a smaller M AE and M SE and a larger SSIM

and M SSIM ). From these statistics, we observe a similar phenomenon as before, i.e., our

network performs better on the dense-layer model. Thus, we conclude that the fault model

has more complex features, which pose more challenges to the network.

Three groups of the dense-layer model, fault model, and salt body model are selected

to make the visual comparison in Figures 12- 14. We plot the velocity curves in these

18
figures to help compare the difference. It can be seen that for simple layer models, the

velocity is almost fitted, though there are still some errors, especially for the model with

large fluctuation. Besides, since the seismic data only have a limited recording time, less

reflection information can be obtained for the deeper layers, which leads to a poor inversion

effect in those layers, as shown in Figure 12c. For the fault model, the change of the velocity

at the fault site is more complex and does not increase monotonously with the depth, so

the network’s predicting ability on the fault models degenerated significantly, though it can

still reflect the trend of fault. Especially for the multi-fault model, as shown in Figure 13b,

it is difficult to obtain the information of multiple faults. For the salt body model, the

exact position, wave velocity and shape of salt body can be obtained by both methods,

while the inversion results of other layers is still accurate. However, our method is better

than SeisInvNet in terms of the position and shape of salt body, as shown in Figure 14.

The same conclusion can be easily drawn from Table 3. For the three types of models, the

dense layer model is the best, while the fault model is the worst.

To further demonstrate the advantage of our method compared with SeisInvNet, in

terms of robustness, we evaluate the prediction results when some shot points are lost, as

shown in the Figure 15. Here we show the inversion results of of data with zero missing shot,

two missing shots, four missing shots and eight missing shots, respectively. Figure 15a is the

ground truth, Figures 15b-15e are the results of our proposed method, and Figures 15f-15i

are the prediction results of SeisInvNet. It can be seen that with several missing shots,

both methods can still work well. With the increase of missing data, the prediction results

of SeisInvNet, especially the salt body, become more and more inaccurate. However, the

inversion quality of our improved SeisInvNet could still be preserved to some extent. The

salt body and interface are acquired accurately, and the absence of shots will not affect the

19
prediction of the position and shape of salt body. The significance of this test lies in that

in the process of acquiring seismic data on the earth surface, considering the severe terrain

conditions, it is often difficult to realize the shot excitation at all of the planned positions.

Our method can still get acceptable prediction results in the presence of missing traces.

DISCUSSION

In this study, a novel method in designing dense-layer, fault, salt body models is de-

veloped, and the applicability and generalizability of SeisInvNet are also demonstrated.

Unlike previous models for DL based inversion methods, the proposed velocity model build-

ing workflow can automatically generate various models containing dense-layers, faults and

salt bodies, thus provide realistic model-data pairs to train the DL network. We consider

the variation of stratum fluctuations, layer thickness, and the change of velocity, and pro-

pose an architecture based on the design workflow of the dense-layer model. Similarly,

for the fault model, the range of fault location, tilt angle and moving direction is selected

and the fault setting function is completed, which could add the faulting structure to the

layer model. Finally, a total of 18,000 models with up to fifteen classes of velocities were

constructed, and good data-model pair were provided for deep learning. For SeisInvNet,

due to its generalizability and applicability, the effective information of the fault velocity

model data can also be extracted by the encoder and generator, and the velocity model is

reconstructed by the decoder.

While analyzing the results, we found that for the fault model with steep-dip or at the

edge of the observation system, the cause for poor predictions is that the reflected wave data

is difficult to record. As can be seen from Figure 16, the velocity values and the interface

locations are well predicted, but the fault is not obtained. This shows that the fault can

20
hardly be predicted because the reflected wave information of the fault is not recorded by

the receivers. Although the full waveform information is used, the network relies more on

the strong first reflected wave for prediction. Information such as diffractions is hard to be

effectively used by the network. In future research, we will further consider the problem

of inverting fault models, in which the long-offset/diffraction data could be extracted to

improve the inversion performance.

To verify the generalization ability of our training network, we test the ten and twelve-

layer models (not included in the training set). As shown in Figure 17a and 17b, for the

ten-layer model, the trained network can still carry out accurate prediction and get obvious

hierarchical results. For the twelve-layer model, due to the small layer thickness, part of the

interface prediction is not accurate, but the overall speed prediction is achieved, as shown

in Figure 17c and 17d.

In addition, we make a simple comparison with FWI (as shown in Figure 18). Given a

better initial model, the resolution of the full waveform inversion in the shallow layer is still

better. However, building the initial model and the tendency into a local minimum, are still

problems for full waveform inversion. Our method does not rely on the initial model, and

can retrieve the exact velocity of each layer, with a high calculation efficiency. Therefore,

we think that deep learning inversion and full waveform inversion may not be completely

opposite, but they may be complementary, which will also be our next research focus.

Furthermore, we adjust the network slightly and test it on a small elastic wave dataset.

As shown in Figure 19, P- and S-wave velocity models are predicted by network. Although

the elastic wave data are more complex than the acoustic wave data, and the training data

is less, the improved SeisInvNet has also achieved considerable results. Considering the

21
lack of training data and that the network structure has not been optimized for elastic wave

data, more in-depth research on elastic wave data will be carried out in the future.

CONCLUSION

In this study, we developed a model design method and proposed an applicable seismic

inversion network based on SeisInvNet. In our model design scheme, we have considered a

number of controlling factors including the variation of stratum fluctuations, layer thickness,

the change of velocity for dense-layer, the range of fault location, moving direction for fault,

and invading salt body. Finally, a total of 18,000 realistic structural and complex models

were constructed. With the simulated data from these models, we collect a large data-model

dataset for training, validation and testing of deep neural networks. We improve SeisInvNet

by introducing more information from the common-receiver gathers and further enhance the

embedding vector. Correspondingly, a more accurate velocity model can be reconstructed.

For the prediction results, the dense-layer, fault models, and salt body models are analyzed

separately. We find that the proposed method can work successfully for the dense-layer and

salt body models and reasonably for the fault models. The inversion results of the fault

models have improvement space by optimizing the data recording system, which we intend

to study in the future.

ACKNOWLEDGMENTS

The research is supported by the Joint Program of the National Natural Science Foun-

dation of China (grant no.U1806226), the National Natural Science Foundation of China

(grant nos.5179007, 51809155, 61702301), and the starting fund from Zhejiang University.

22
.

23
 REFERENCES

Alkhalifah, T., and I. Tsvankin, 1995, Velocity analysis for transversely isotropic media:

Geophysics, 60, no. 5, 1550–1566.

Araya-Polo, M., J. Jennings, A. Adler, and T. Dahlke, 2018, Deep-learning tomography:

The Leading Edge, 37, no. 1, 58–66.

Baysal, E., D. D. Kosloff, and J. W. Sherwood, 1983, Reverse time migration: Geophysics,

48, no. 11, 1514–1524.

Bunks, C., F. M. Saleck, S. Zaleski, and G. Chavent, 1995, Multiscale seismic waveform

inversion: Geophysics, 60, no. 5, 1457–1473.

Chen, Y., H. Chen, K. Xiang, and X. Chen, 2016, Geological structure guided well log

interpolation for high-fidelity full waveform inversion: Geophysical Journal International,

207, no. 2, 1313–1331.

Chen, Y., M. Zhang, M. Bai, and W. Chen, 2019, Improving the signal-to-noise ratio of

seismological datasets by unsupervised machine learning: Seismological Research Letters,

90, no. 4, 1552–1564.

Choromanska, A., M. Henaff, M. Mathieu, G. B. Arous, and Y. Lecun, 2014, The loss

surfaces of multilayer networks: Eprint Arxiv, 192–204.

Cohen, J. K., and N. Bleistein, 1979, Velocity inversion procedure for acoustic waves: Geo-

physics, 44, no. 6, 1077–1087.

Deng, L., D. Yu, et al., 2014, Deep learning: methods and applications: Foundations and

Trends R in Signal Processing, 7, no. 3–4, 197–387.

Dunkin, J., and F. Levin, 1973, Effect of normal moveout on a seismic pulse: Geophysics,

38, no. 4, 635–642.

Esteva, A., B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, 2017,

24
 Dermatologist-level classification of skin cancer with deep neural networks: Nature, 542,

115–118.

Furumura, T., B. Kennett, and H. Takenaka, 1998, Parallel 3-D pseudospectral simulation

of seismic wave propagation: Geophysics, 63, no. 1, 279–288.

George, D., and E. Huerta, 2018, Deep learning for real-time gravitational wave detection

and parameter estimation: Results with advanced LIGO data: Physics Letters B, 778,

64–70.

Hinton, G., L. Deng, D. Yu, G. E. Dahl, A. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke,

P. Nguyen, and T. N. Sainath, 2012a, Deep neural networks for acoustic modeling in

speech recognition: The shared views of four research groups: IEEE Signal Processing

Magazine, 29, no. 6, 82–97.

Hinton, G. E., S. Osindero, and Y.-W. Teh, 2006, A fast learning algorithm for deep belief

nets: Neural computation, 18, no. 7, 1527–1554.

Hinton, G. E., N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov, 2012b,

Improving neural networks by preventing co-adaptation of feature detectors: arXiv

preprint arXiv:1207.0580.

Hu, W., A. Abubakar, and T. M. Habashy, 2009a, Joint electromagnetic and seismic inver-

sion using structural constraints: Geophysics, 74, no. 6, R99–R109.

——–, 2009b, Simultaneous multifrequency inversion of full-waveform seismic data: Geo-

physics, 74, no. 2, R1–R4.

Kawaguchi, K., 2016, Deep learning without poor local minima: Advances in neural infor-

mation processing systems, 586–594.

Khan, A., J. A. D. Connolly, J. Maclennan, and K. Mosegaard, 2010, Joint inversion of

seismic and gravity data for lunar composition and thermal state: Geophysical Journal

25
 International, 168, no. 1, 243–258.

Kingma, D. P., and J. Ba, 2014, Adam: A method for stochastic optimization: arXiv

preprint arXiv:1412.6980.

Kosloff, D. D., and E. Baysal, 1982, Forward modeling by a fourier method: Geophysics,

47, no. 10, 1402–1412.

Lailly, P., and J. Bednar, 1983, The seismic inverse problem as a sequence of before stack

migrations: Conference on inverse scattering: theory and application, SIAM, 206–220.

Lecun, Y., Y. Bengio, and G. Hinton, 2015, Deep learning: Nature, 521, no. 7553, 436.

Li, S., B. Liu, Y. Ren, Y. Chen, S. Yang, Y. Wang, and P. Jiang, 2020, Deep-learning

inversion of seismic data: IEEE Transactions on Geoscience and Remote Sensing, 58, no.

3, 2135–2149.

Li, S., B. Liu, X. Xu, L. Nie, Z. Liu, J. Song, H. Sun, L. Chen, and K. Fan, 2017, An

overview of ahead geological prospecting in tunneling: Tunnelling and Underground Space

Technology, 63, 69–94.

Li, W., 2018, Classifying geological structure elements from seismic images using deep

learning, in SEG Technical Program Expanded Abstracts 2018: Society of Exploration

Geophysicists, 4643–4648.

Liu, B., Q. Guo, S. Li, B. Liu, Y. Ren, Y. Pang, L. Liu, and P. Jiang, 2020a, Deep learning

inversion of electrical resistivity data: IEEE Transactions on Geoscience and Remote

Sensing, 58, no. 8, 5715–5728.

Liu, B., Y. Pang, D. Mao, J. Wang, Z. Liu, N. Wang, S. Liu, and X. Zhang, 2020b, A

rapid four-dimensional resistivity data inversion method using temporal segmentation:

Geophysical Journal International, 221, no. 1, 586–602.

Malehmir, A., D. Raymond, B. Gilles, U. Milovan, J. Christopher, W. Donald John, M.

26
 Bernd, and C. Geoff, 2012, Seismic methods in mineral exploration and mine planning: A

general overview of past and present case histories and a look into the future: Geophysics,

77, no. 5, WC173–WC190.

McCulloch, W. S., and W. Pitts, 1943, A logical calculus of the ideas immanent in nervous

activity: The Bulletin of Mathematical Biophysics, 5, no. 4, 115–133.

Mittet, R., R. Sollie, and K. Hokstad, 1995, Prestack depth migration with compensation

for absorption: Geophysics, 60, no. 5, 1485–1494.

Moseley, B., A. Markham, and T. Nissen-Meyer, 2018, Fast approximate simulation of

seismic waves with deep learning: arXiv preprint arXiv:1807.06873.

Nurindrawati, F., and J. Sun, 2019, Estimating total magnetization directions using convo-

lutional neural networks, in SEG Technical Program Expanded Abstracts 2019: Society

of Exploration Geophysicists, 2163–2167.

Pratt, R. G., 1999, Seismic waveform inversion in the frequency domain, part 1: Theory

and verification in a physical scale model: Geophysics, 64, no. 3, 888–901.

Pratt, R. G., and R. M. Shipp, 1999, Seismic waveform inversion in the frequency domain,

part 2: Fault delineation in sediments using crosshole data: Geophysics, 64, no. 3, 902–

914.

Pratt, R. G., and M. Worthington, 1990, Inverse theory applied to multi-source cross-hole

tomography. part 1: Acoustic wave-equation method: Geophysical Prospecting, 38, no.

3, 287–310.

Puzyrev, V., 2019, Deep learning electromagnetic inversion with convolutional neural net-

works: Geophysical Journal International, 218, no. 2, 817–832.

Qu, S., E. Verschuur, and Y. Chen, 2019, Full-waveform inversion and joint migration

inversion with an automatic directional total variation constraint: Geophysics, 84, no. 2,

27
 R175–R183.

Ren, Y., X. Xu, S. Yang, L. Nie, and Y. Chen, 2020, A physics-based neural-network way

to perform seismic full waveform inversion: IEEE Access, 8, 112266–112277.

Rosenblatt, F., 1958, The perceptron: a probabilistic model for information storage and

organization in the brain: Psychological Review, 65, no. 6, 386–408.

Röth, G., and A. Tarantola, 1994, Neural networks and inversion of seismic data: Journal

of Geophysical Research: Solid Earth, 99, no. B4, 6753–6768.

Rumelhart, D. E., G. E. Hinton, and R. J. Williams, 1988, Learning representations by

back-propagating errors: Nature, 323, no. 6088, 696–699.

Saad, O., and Y. Chen, 2020, Deep denoising autoencoder for seismic random noise atten-

uation: Geophysics, 85, no. 4, V367–V376.

Santurkar, S., D. Tsipras, A. Ilyas, and A. Madry, 2018, How does batch normalization

help optimization?, in Advances in Neural Information Processing Systems 31: Curran

Associates, Inc., 2483–2493.

Shi, Y., X. Wu, and S. Fomel, 2019, Saltseg: Automatic 3D salt segmentation using a deep

convolutional neural network: Interpretation, 7, no. 3, SE113–SE122.

Shin, C., and Y. H. Cha, 2008, Waveform inversion in the Laplace domain: Geophysical

Journal International, 173, no. 3, 922–931.

Shin, C., and Y. Ho Cha, 2009, Waveform inversion in the Laplace—Fourier domain: Geo-

physical Journal International, 177, no. 3, 1067–1079.

Simonyan, K., and A. Zisserman, 2014, Very deep convolutional networks for large-scale

image recognition: arXiv preprint arXiv:1409.1556.

Sirgue, L., O. Barkved, J. Dellinger, J. Etgen, U. Albertin, and J. Kommedal, 2010, The-

matic set: Full waveform inversion: The next leap forward in imaging at Valhall: First

28
 Break, 28, no. 4, 65–70.

Sun, J., Z. Niu, K. A. Innanen, J. Li, and D. O. Trad, 2020, A theory-guided deep-learning

formulation and optimization of seismic waveform inversion: Geophysics, 85, R87–R99.

Tarantola, A., 1984, Inversion of seismic reflection data in the acoustic approximation:

Geophysics, 49, no. 8, 1259–1266.

Tsai, K. C., W. Hu, X. Wu, J. Chen, and Z. Han, 2018, First-break automatic picking

with deep semisupervised learning neural network, in SEG Technical Program Expanded

Abstracts 2018: Society of Exploration Geophysicists, 2181–2185.

Virieux, J., H. Calandra, and R.-E. Plessix, 2011, A review of the spectral, pseudo-spectral,

finite-difference and finite-element modelling techniques for geophysical imaging: Geo-

physical Prospecting, 59, no. 5, 794–813.

Voulodimos, A., N. Doulamis, A. Doulamis, and E. Protopapadakis, 2018, Deep learning

for computer vision: A brief review: Computational Intelligence and Neuroscience, 2018,

1–13.

Wang, Z., E. P. Simoncelli, and A. C. Bovik, 2003, Multiscale structural similarity for

image quality assessment: The Thrity-Seventh Asilomar Conference on Signals, Systems

& Computers, 2003, IEEE, 1398–1402.

Woodward, M. J., 1992, Wave-equation tomography: Geophysics, 57, no. 1, 15–26.

Wu, X., L. Liang, Y. Shi, and S. Fomel, 2019, Faultseg3d: using synthetic datasets to train

an end-to-end convolutional neural network for 3D seismic fault segmentation: Geo-

physics, 84, no. 3, 1–36.

Wu, Y., and Y. Lin, 2018, InversionNet: A Real-Time and Accurate Full Waveform Inversion

with CNNs and continuous CRFs: arXiv preprint arXiv:1811.07875.

Yang, F., and J. Ma, 2019, Deep-learning inversion: a next generation seismic velocity-

29
 model building method: Geophysics, 84, no. 4, 583–599.

Yang, Y., Y. Li, and T. Liu, 2009, 1D viscoelastic waveform inversion for Q structures from

the surface seismic and zero-offset VSP data: Geophysics, 74, no. 6, WCC141–WCC148.

Yu, S., J. Ma, and W. Wang, 2018, Deep learning tutorial for denoising: arXiv preprint

arXiv:1810.11614.

Yuan, P., W. Hu, X. Wu, J. Chen, and H. Van Nguyen, 2019, First arrival picking using

u-net with lovasz loss and nearest point picking method, in SEG Technical Program

Expanded Abstracts 2019: Society of Exploration Geophysicists, 2624–2628.

Zhang, G., C. Lin, and Y. Chen, 2020, Convolutional neural networks for microseismic

waveform classification and arrival picking: Geophysics, 85, no. 4, WA227–WA240.

Zhang, G., Z. Wang, and Y. Chen, 2018, Deep learning for seismic lithology prediction:

Geophysical Journal International, 215, no. 2, 1368–1387.

30
 LIST OF TABLES

1 Detailed media parameters of numerical models.

2 Loss of prediction results in validation set.

3 Loss of prediction results in test set.

31
 LIST OF FIGURES

1 The workflow of the deep learning based velocity model inversion algorithm. Through

the proposed velocity model and designed method, a large number of models are provided.

After the observation system is determined, we obtain the seismic data by performing the

wavefield simulation. We use seismic data as input and the velocity model as the label to

train the designed network. After training, given the input seismic data, we can directly

obtain the network’s predicted model.

2 Multiple interfaces made up of y.

3 Curves and the corresponding models in different construction ways.

4 The average wave velocity at each depth of dense-layer models in the training set.

5 Fault model design workflow.

6 Salt body model design workflow.

7 Designed dense-layer, fault, and salt body velocity models samples.

8 Comparison between the previous SeisInvNet and the proposed method on the En-

coder part. (a) presents the SeisInvNet, i.e., which only encodes the information from the

common-shot gather. (b) presents the proposed method which encodes information from

both the common-shot gather and the common-receiver gather.

9 The proposed method used for model prediction.

10 The velocity model and its corresponding simulated data. (a) denotes the velocity

model; (b), (c), and (d) represent the recorded data from the 1st, 10th, and 20th shot,

respectively.

11 Loss curves on training set and validation set during the training process. Among

them, the red line is derived by the proposed method, and the blue line is obtained by

the SeisInvNet. (a) denotes mean-squared error for the simulated velocity inversion in the

32
training set; (b) denotes MSSIM for the simulated velocity inversion in the training set; (c)

denotes mean-squared error for the simulated velocity inversion in the validation set; (d)

denotes MSSIM for the simulated velocity inversion in the validation set.

12 The comparison between the ground truth, prediction results of the proposed

method, and prediction results of SeisInvNet with the velocity curves for the dense-layer

models in testing set.

13 The comparison between the ground truth, prediction results of the proposed

method, and prediction results of SeisInvNet with the velocity curves for the fault models

in testing set.

14 The comparison between the ground truth, prediction results of the proposed

method, and prediction results of SeisInvNet with the velocity curves for the salt body

models in testing set.

15 The inversion results from data with zero missing shot, two missing shots, four

missing shots, and eight missing shots, respectively. (a) represents the ground truth. (b)-

(e) represent the results of our proposed method. (f)-(i) represent the prediction results of

SeisInvNet.

16 Prediction results for fault with large dips and faults close to the model boundary.

17 Network generalization ability test. (a) and (c) represent 10 and 12 layer velocity

models, respectively; (b) and (d) represent the prediction results of the proposed method,

respectively.

18 Comparison between the proposed method and full waveform inversion. (a) repre-

sents the true velocity model; (b) represents the initial model of the full waveform inversion;

(c) represents the full waveform inversion result; (d) represents the prediction result of the

proposed method.

33
 19 Preliminary test results of elastic data inversion problem. (a), (b) and (c) repre-

sent the received data from the 1st, 10th, and 20th shot, respectively; (d) and (e) represent

P-wave velocity model and S-wave velocity model, respectively; (f) and (g) represent the

corresponding preliminary prediction results of the proposed method, respectively.

34
Figure 1: The workflow of the deep learning based velocity model inversion algorithm.

Through the proposed velocity model and designed method, a large number of models

are provided. After the observation system is determined, we obtain the seismic data by

performing the wavefield simulation. We use seismic data as input and the velocity model

as the label to train the designed network. After training, given the input seismic data, we

can directly obtain the network’s predicted model.

Liu et al., 2020 – GEO-2020

35
Figure 2: Multiple interfaces made up of y.

Liu et al., 2020 – GEO-2020

36
Figure 3: Curves and the corresponding models in different construction ways.

Liu et al., 2020 – GEO-2020

37
 4.0
5-layer model
6-layer model
3.5 7-layer model
8-layer model
Velocity (km/s)

9-layer model
3.0

2.5

2.0

1.5
0 0.2 0.4 0.6 0.8 1.0
Depth (km)

Figure 4: The average wave velocity at each depth of dense-layer models in the training set.

Liu et al., 2020 – GEO-2020

38
Layer model Fault movement Fault model
(a) (b) (c)

Figure 5: Fault model design workflow.

Liu et al., 2020 – GEO-2020

39
Figure 6: Salt body model design workflow.

Liu et al., 2020 – GEO-2020

40
Figure 7: Designed dense-layer, fault, and salt body velocity models samples.

Liu et al., 2020 – GEO-2020

41
 (a)

(b)

Figure 8: Comparison between the previous SeisInvNet and the proposed method on the

Encoder part. (a) presents the SeisInvNet, i.e., which only encodes the information from the

common-shot gather. (b) presents the proposed method which encodes information from

both the common-shot gather and the common-receiver gather.

Liu et al., 2020 – GEO-2020

42
Figure 9: The proposed method used for model prediction.

Liu et al., 2020 – GEO-2020

43
 (a) (b)

(c) (d)

Figure 10: The velocity model and its corresponding simulated data. (a) denotes the

velocity model; (b), (c), and (d) represent the recorded data from the 1st, 10th, and 20th

shot, respectively.

Liu et al., 2020 – GEO-2020

44
 0.8 1
The proposed method The proposed method

MSSIM Loss curve on training set

MSE Loss curve on training set

SeisInvNet SeisInvNet
0.8
0.6

0.6
0.4
0.4

0.2
0.2

0 0
0 50 100 150 200 0 50 100 150 200
Epochs Epochs

(a) (b)

MSSIM Loss curve on validation set 0.2

The proposed method
0 SeisInvNet

0.8

0.6

0.4

0.2

0
0 50 100 150 200
Epochs

(c) (d)

Figure 11: Loss curves on training set and validation set during the training process. Among

them, the red line is derived by the proposed method, and the blue line is obtained by the

SeisInvNet. (a) denotes mean-squared error for the simulated velocity inversion in the

training set; (b) denotes MSSIM for the simulated velocity inversion in the training set; (c)

denotes mean-squared error for the simulated velocity inversion in the validation set; (d)

denotes MSSIM for the simulated velocity inversion in the validation set.

Liu et al., 2020 – GEO-2020

45
Figure 12: The comparison between the ground truth, prediction results of the proposed

method, and prediction results of SeisInvNet with the velocity curves for the dense-layer

models in testing set.

Liu et al., 2020 – GEO-2020

46
Figure 13: The comparison between the ground truth, prediction results of the proposed

method, and prediction results of SeisInvNet with the velocity curves for the fault models

in testing set.

Liu et al., 2020 – GEO-2020

47
Figure 14: The comparison between the ground truth, prediction results of the proposed

method, and prediction results of SeisInvNet with the velocity curves for the salt body

models in testing set.

Liu et al., 2020 – GEO-2020

48
Figure 15: The inversion results from data with zero missing shot, two missing shots, four

missing shots, and eight missing shots, respectively. (a) represents the ground truth. (b)-

(e) represent the results of our proposed method. (f)-(i) represent the prediction results of

SeisInvNet.

Liu et al., 2020 – GEO-2020

49
Figure 16: Prediction results for fault with large dips and faults close to the model boundary.

Liu et al., 2020 – GEO-2020

50
Figure 17: Network generalization ability test. (a) and (c) represent 10 and 12 layer velocity

models, respectively; (b) and (d) represent the prediction results of the proposed method,

respectively.

Liu et al., 2020 – GEO-2020

51
Figure 18: Comparison between the proposed method and full waveform inversion. (a)

represents the true velocity model; (b) represents the initial model of the full waveform

inversion; (c) represents the full waveform inversion result; (d) represents the prediction

result of the proposed method.

Liu et al., 2020 – GEO-2020

52
Figure 19: Preliminary test results of elastic data inversion problem. (a), (b) and (c) repre-

sent the received data from the 1st, 10th, and 20th shot, respectively; (d) and (e) represent

P-wave velocity model and S-wave velocity model, respectively; (f) and (g) represent the

corresponding preliminary prediction results of the proposed method, respectively.

Liu et al., 2020 – GEO-2020

53
 Hyperparameter

Batchsize 36

The initial 5×10−5

learning rate

Dropout rate 0.2

Epochs 200

Table 1: Detailed media parameters of numerical models.

54
 Loss The proposed SeisInvNet

method

LM SE in training 3.736 ×10−4 5.007×10−4

set

LM SSIM in 7.357×10−3 1.105×10−2

training set

LM SE in 2.659×10−4 2.698×10−4

validation set

LM SSIM in 6.335×10−2 6.694×10−2

validation set

Table 2: Loss of prediction results in validation set.

55
 Metric Dense-layer model Fault model Salt body model

MAE 0.014703 0.021852 0.017017

MSE 0.000744 0.001431 0.001289

The propose method
SSIM 0.0.905838 0.844608 0.893321

MSSIM 0.981307 0.955735 0.967640

MAE 0.015075 0.022245 0.018588

MSE 0.000854 0.001553 0.001498

SeisInvNet
SSIM 0.905952 0.843426 0.880394

MSSIM 0.978274 0.950534 0.961678

Table 3: Loss of prediction results in test set.

56

View publication stats