IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.

61, 2023 5909013

An Unsupervised Inversion Method for Seismic

Brittleness Parameters Driven by the
Physical Equation
Yinghe Wu , Shulin Pan , Yaojie Chen, Jingyi Chen, Shengbo Yi, Dongjun Zhang, and Guojie Song

Abstract— Brittleness is an important parameter character- Index Terms— Brittleness, Fastformer network, physical equa-
izing the fracturing properties of shale reservoir, which can tion, seismic inversion, unsupervised learning.
be predicted by the prestack seismic inversion. In order to
overcome the low efficiency and ill-posed problems of the I. I NTRODUCTION
traditional prestack brittleness inversion, we propose a new
unsupervised deep learning (DL) inversion method for seismic
brittleness parameters based on the physical equation. This
method integrates DL framework and the physical equation and
W ITH the increasing difficulty in obtaining shallow and
conventional oil and gas resources, seismic explo-
ration technology for unconventional oil and gas reservoirs
provides a DL inversion strategy without actual labels. We first has become a focus of attention. As one of the important
input the original seismic data into the Fastformer network and
use the low-frequency model as the physical constraint to predict
unconventional oil and gas resources, shale gas exploration and
the brittle parameters. Then, the prediction results of brittleness development is very important. Due to the strong heterogeneity
parameters are sent to the forward modeling module (a linear of shale formation and the characteristics of low porosity
approximation equation) to calculate the synthetic seismic data. and low permeability, the large-scale hydraulic fracturing is
Next, the error between the calculated seismic data and original required to form a fracture network during the development
seismic data is used to update the network prediction results. The
network parameters are iteratively optimized to minimize the
process to obtain industrial gas brittle characteristics of rocks
error, and the brittle prediction parameters are finally output. are also a factor that cannot be ignored, so the ease of
In the whole training process, it is not necessary to use the fracturing of rocks becomes a very important indicator in
real brittle parameters as the labels. Through this method, the the evaluation of shale oil and gas reservoirs. The higher the
effect of approximate unsupervised learning is obtained. Finally, degree of rock brittleness, the easier it is to be modified during
we apply the proposed method to the synthetic data and field
data and compared with the results inverted by the traditional
the fracturing process, thus forming a connected fracture
L1 method. The experimental results show that the proposed network to improve production. The rock physics experiments
method has higher inversion accuracy and efficiency than the show that Young’s modulus (E) and Poisson’s ratio (σ ) can
traditional L1 method, which has a great potential in the practical better indicate the difficulty of rock fracturing and can be
application. used to evaluate the rock brittleness, which provide a basis
for finding the brittle zone of the reservoir.
Manuscript received 26 January 2023; revised 24 March 2023; accepted
24 April 2023. Date of publication 5 May 2023; date of current ver- At present, the prediction of reservoir brittleness mainly
sion 17 May 2023. This work was supported in part by the Technology adopts the geophysical inversion and logging data calcula-
Cooperation Project of the CNPC-SWPU Innovation Alliance under Grant tion [1]. Among them, the prestack amplitude versus offset
2020CX020000, in part by the Central Government Funds of Guiding Local
Scientific and Technological Development under Grant 2021ZYD0003, in part (AVO) inversion can use various information contained in
by the Nanchong Municipal Government-Universities Scientific Cooperation prestack seismic data to invert rich elastic parameters of the
Project under Grant SXHZ045, and in part by the Research and Innovation underground media, which has been widely used for the
Fund for Graduate Students of Southwest Petroleum University under Grant
2022KYCX040. (Corresponding authors: Shulin Pan; Yinghe Wu.) brittleness prediction. The theoretical basis of the prestack
Yinghe Wu, Yaojie Chen, and Shengbo Yi are with the School of AVO inversion is the Zoeppritz equation. However, due to
Geoscience and Technology, Southwest Petroleum University, Chengdu the complexity of the equation form, it is not convenient
610500, China (e-mail: yinghe.wu@foxmail.com; yaojie.chen@foxmail.com;
13159962121@163.com). to use, and some scholars have made approximations to it.
Shulin Pan is with the State Key Laboratory of Oil and Gas Reser- Aki and Richards [2] expressed the reflection coefficients as
voir Geology and Exploitation and the School of Geoscience and the relationship between the compressional wave (P-wave)
Technology, Southwest Petroleum University, Chengdu 610500, China
(e-mail: shulinpan@swpu.edu.cn). velocity, shear wave (S-wave) velocity, and density (ρ), which
Jingyi Chen is with the Seismic Anisotropy Group, Department provides a theoretical basis for later AVO inversions. Shuey [3]
of Geosciences, The University of Tulsa, Tulsa, OK 74104 USA rederived the Zoeppritz equation by introducing σ into the
(e-mail: jingyi-chen@utulsa.edu).
Dongjun Zhang is with the Shale Gas Research Institute, PetroChina expression for the reflection coefficients. Goodway et al. [4]
Southwest Oil & Gasfield Company, Chengdu 610051, China (e-mail: proposed a formula to express brittleness using Lame param-
zhangdj_wt@petrochina.com.cn). eters and shear modulus. Zong et al. [5] deduced the approx-
Guojie Song is with the School of Science, Southwest Petroleum University,
Chengdu 610500, China (e-mail: cylsgj@126.com). imate equations of E, σ , and ρ (YPD) based on Aki and
Digital Object Identifier 10.1109/TGRS.2023.3273302 Richard’s formula, established the linear relationship between
1558-0644 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
5909013 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023

the longitudinal wave reflection coefficient and E reflectivity, from prestack seismic data through the partial label training.
σ reflectivity, and ρ reflectivity, and obtained the direct inver- Das et al. [31] adopted convolution neural network to invert
sion method of E and σ . Guo et al. [6] also proposed a new impedance, and this method is used in actual data inversion
formula to express brittleness using E and σ . According to by training a large number of synthetic data. Wu et al. [32]
the YPD equation, Zhang et al. [7] deduced the approximate designed a deep learning (DL) network with residual modules
reflection coefficient equations of both compressional and and two attention mechanisms, which enhances the inversion
converted waves based on Eρ, σ , and ρ and provided a new accuracy of wave impedance inversion. Wang et al. [33]
method for direct inversion. Ritesh et al. [8] quantitatively put forward a seismic impedance inversion method based on
studied the lithology and brittleness of shale and believe that the cyclic generation countermeasure network, which has a
Eρ can be used as a good indicator of shale brittleness. better performance than other DL based methods in most
Zhang et al. [9] proposed a two-term approximation equation cases. At present, most of the work is focused on impedance
by replacing the density term with E and σ through E and inversion and other tasks, and few people have applied DL to
the relationship between σ and ρ. Ge et al. [10] deduced a the brittleness inversion. In addition, most of these studies
new brittleness parameter reflection coefficient approximation are supervised learning and semisupervised learning meth-
equation based on the correlation between E and σ and ods [34], [35], which require the use of appropriate training
predicted the brittleness parameters. set and training network, and the quality and quantity of
Because seismic inversion is a typical ill-posed problem, and the training set directly affect the inversion effect. However,
many scholars have carried out inversion algorithm research. it is difficult to obtain training labels from real seismic data,
Tikhonov proposed Tikhonov regularization [11], which uses which has become an important problem in the large-scale
the L2 norm to eliminate the defects of nonlinearity and application of DL methods in seismic inversion.
instability caused by the Zoeppritz equation inversion but Inspired by the above and related research, this article
will impose uniform smooth constraints on inversion results. proposes an unsupervised DL prestack seismic brittleness
Wang [12] used the L1 norm and gradient descent method parameter inversion method based on the physical equation.
for wave impedance inversion. Zhang et al. [13] proposed a Different from the traditional DL method, this method can
seismic inversion method based on L1 norm and the total directly predict the brittleness of a stacked seismic data
variational (TV) regularization. Li et al. [14] introduced a without building a training set and pretraining network, which
frequency-domain anisotropic total p-variational (ATpV) reg- avoids the problem of building a training set embedding
ularization into wave impedance inversion, which reduced stratigraphic parameter data labels in the supervised DL inver-
inversion error and improved inversion accuracy. He et al. [15] sion method. At the same time, we add the low-frequency
proposed an acoustic impedance inversion method based on physical constraints and forward equation modules to the
weighted L1 norm sparse constraint, which obtained clearer Fastformer network to improve the prediction accuracy and
impedance boundary, weakened pseudolayer phenomenon, efficiency of the brittleness index inversion. The rest of this
and improved inversion accuracy. Dai et al. [16] added the article is arranged as follows. In Section II, the principle
cross gradient of model parameters as a regular term to the of the direct inversion of prestack seismic brittleness index
amplitude-versus-angle (AVA) inversion, making the inversion based on the approximate Zoeppritz equation is introduced.
results more stable. Dai and Yang [17] proposed a regularized In Section III, the Fastformer network framework for the
AVA inversion method based on L0 norm gradient, which brittleness parameter inversion is presented, and the proposed
improves the inversion accuracy of P-wave velocity, S-wave unsupervised inversion method driven by the physical equation
velocity, and density. Chen et al. [18] proposed a direct and training process is described in detail. In Section IV, the
inversion method for brittle parameters based on weighted Lp, prediction results based on Marmousi2 theoretical data and
which combines weighted Lp and L1 regularization methods real data are presented and compared with the traditional L1
into the inversion of brittle parameters, improving the sparsity method and well logging data. In the field dataset test, the
and accuracy of the inversion results. This kind of traditional results of the indirect inversion compared with commercial
algorithm can improve the inversion accuracy to a certain software are also provided. Finally, the conclusion is drawn in
extent, but it takes a long time in the practical application. Section V.
In recent years, machine learning (ML) and DL have
shown the great potential when applied to a variety of geo- II. P RINCIPLE OF P RESTACK B RITTLENESS
physical research problems by providing automated perfor- PARAMETER I NVERSION
mances. For example, seismic first-arrival pickup, seismic We know that E represents the compressive resistance of
source location, seismic data denoising, stack velocity estima- rock, and ρ can show the lithologic characteristics of reservoir.
tion, fault and horizon recognition of seismic profile, seismic The product of E and ρ (Eρ) can be used as a more sensitive
phase classification, surface wave dispersion curve pickup parameter to characterize the brittleness of the rock [34].
and inversion, and so on [19], [20], [21], [22], [23], [24], σ is the absolute value of the ratio of transverse strain to
[25], [26], [27], [28], [29]. Some scholars have used DL meth- longitudinal strain, which is related to the properties of fluid in
ods to invert seismic parameters, and achieved good results. the reservoir and can be used as an indicator of the brittleness
Biswas et al. [30] developed a physically guided convolutional of shale gas reservoir. The high precision and σ can be
neural network, which consists of a forward network and an obtained by the prestack joint inversion to identify shale gas
inversion network, and predicted the P-wave velocity and ρ reservoir characteristics.

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
WU et al.: UNSUPERVISED INVERSION METHOD FOR SEISMIC BRITTLENESS PARAMETERS 5909013

A. Eρ, σ , and ρ Approximate Reflection Coefficient Equation where Li = ln(σi ). The matrix is expressed as Rσ = DL σ ,
The reflection coefficient equation establishes the relation- where D is the difference matrix and L σ is the logarithm of
ship between seismic data and elastic parameters. In 2012, σ . According to (7) and using angle related wavelet matrix,
Zong et al. [5] proposed the approximate equation of the the angles-related forward seismic data can be expressed as:
reflection coefficient based on E, σ , and ρ S(θ1 )
 

1 2

1E  .. 
R(θ ) = sec θ − 2k 2 sin2 θ  . 
4 E S(θ N )
1 2 (2k 2 − 3)(2k 2 − 1)2

W (θ1 )a(θ1 )D W (θ1 )b(θ1 )D W (θ1 )c(θ1 )D  L Eρ 
 
+ sec θ
4 k 2 (4k 2 − 3) 1 .. .. ..
=  . . .  Lσ 

1 − 2k 2 1σ

2
+ 2k 2 sin2 θ W (θ N )a(θ N )D W (θ N )b(θ N )D W (θ N )c(θ N )D Lρ
3 − 4k 2 σ
(9)
1 1 2 1ρ
 
+ − sec θ (1)
2 4 ρ where L σ , and represent the density volume, σ and natural
where θ is the incidence angle, and k is the ratio of com- logarithm of ρ, and matrices a(θi ), b(θi ), and are the diagonal
pressional and shear wave velocity. Substitute (1Eρ/Eρ) = matrices. If diagonal matrices are defined as B, the forward
(1E/E) + (1ρ/ρ) into (2) to obtain the approximate equa- equation can be linearly simplified as
tion of reflection coefficient based on Eρ, σ , and ρ S = W B DL = AL (10)
1Eρ
 
1 2
R(θ ) = sec θ − 2k sin θ
2 2
where A = W B D. Based on the above forward equation, the
4 Eρ traditional inversion objective function can be expressed as
(2k − 3)(2k − 1)2
2 2

1
+ sec2 θ J (L) = min L ∥AL − S∥22 (11)
4 k 2 (4k 2 − 3)
1 − 2k 2 1σ

where ∥ ∥2 is the L2 norm. Based on the above forward
+ 2k 2 sin2 θ
3 − 4k 2 σ modeling framework and related introduction, together with
1ρ the initial model constraints, we can get the inversion objective
 
1 1
+ + 2k 2 sin2 θ − sec2 θ (2) function as follows:
2 2 ρ
set J (L) = min ∥AL − S∥22 + µ∥L − L 0 ∥22 + λ (wi ∥DL∥1 ) (12)
L
1
a(θ ) = sec2 θ − 2k 2 sin2 θ (3) where L 0 is the low-frequency model, µ is the parameter of the
4
1 (2k 2 − 3)(2k 2 − 1)2 low-frequency model, λ is regularization factor, is the weight
b(θ ) = sec2 θ coefficient, and ∥ ∥1 is the L1 norm. In this way, an inversion
4 k 2 (4k 2 − 3)
objective function suitable for prestack inversion is obtained.
1 − 2k 2
+ 2k 2 sin2 θ (4) The relevant formation brittleness parameters can be obtained
3 − 4k 2 by (12).
1 1
c(θ ) = + 2k 2 sin2 θ − sec2 θ (5)
2 2 III. P HYSICAL E QUATION D RIVEN U NSUPERVISED
1Eρ 1σ 1ρ L EARNING I NVERSION M ETHOD
R Eρ = , Rσ = , Rρ = . (6)
Eρ σ ρ A. Fastformer Network Structure
Then, (2) can be simplified as Traditional inversion and the established inversion objective
R(θ ) = a(θ)R Eρ + b(θ)Rσ + c(θ)Rρ (7) function are generally single trace data inversion one by one,
so the network for the time series processing is more suitable
where R Eρ is the Eρ reflectivity, Rσ is the σ reflectivity, for our task. Compared with recurrent neural networks (RNN,
and Rρ is the ρ reflectivity. Equation (7) establishes the LSTM, GRU, etc.), transformer [36] is based on the multi-
direct linear relationship between Eρ reflectivity, Poisson’s head self-attention mechanism. By capturing the interaction
specific reflectivity, density reflectivity, and the P–P reflection between inputs of each pair of positions, transformer can
coefficient. effectively model context in a long sequence. It and its variants
have achieved great success in many fields, but its efficiency
B. Inverse Equation of Brittleness Parameter is low due to its quadratic complexity to the length of the
In seismic prestack brittleness inversion, the inversion rela- input sequence. Although there are many methods to accelerate
tionship equation between seismic data and parameters (Eρ, σ , the transformer, they are either inefficient or ineffective in the
and ρ) to be inverted is generally established through reflection long sequences. Wu et al. [37] proposed that FastFormer is a
coefficient equation (7). Taking σ as an example transformer variant that can realize context modeling under the
σi+1 − σi 1 1σ 1 linear complexity. The network structure is shown in Fig. 1.
Rσ = ≈ = 1 ln(σ ) Fastformer first combines the input attention query matrix into
σi+1 + σi 2 σ 2
1 1 a global query vector and then uses elementwise multiplication
≈ [ln(σi+1 ) − ln(σi )] = (L i+1 − L i ) (8) to learn the global context-dependent key matrix for the
2 2
Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
5909013 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023

Fig. 2. Unsupervised method training process based on physical equation

driving.

low-frequency model L 0 , and calculate the partial error of the

network predicted PL and low-frequency model L 0 for con-
straint. Then, PL is sent to the forward modeling module and
the forward operator A for matrix multiplication to obtain the
Fig. 1. Fastformer network structure. forward seismic data S ′ , calculate the errors of S and S ′ , and
backpropagate the two parts of errors to the network to update
attention key and global query vectors. The additive attention the network parameter w through iteration, so that the errors
is then merged into the global key vectors. The interaction of S and S ′ are minimized, and the optimal network parameter
between global key and attention is modeled through the w ′ is obtained. Finally, the seismic brittleness parameters with
element product, and global context aware attention is learned high accuracy can be directly calculated by using the network
using the linear transformation. Finally, they are combined with optimal parameters. At this time, the MSE loss function
with the attention query matrix to form the final output. Thus, of the network is
the computational complexity can be reduced to linear, and the
context information in the input sequence can be effectively Loss = ∥A PL − S∥2 + µ ∗ ∥PL − L 0 ∥2 (13)
captured. For the additive attention network learning global
where µ is the parameter of low-frequency constraint term,
query and key vectors, its time and memory overhead are
generally set to 0.0001–0.001 (the empirical value obtained
both O(N .d), and the total number of parameters is 2hd (h
through multiple experiments, and it should not be too large or
is the number of attention headers). In addition, the time cost
too small). By introducing the low-frequency constraints and
and memory cost of the element product are O(N .d), and the
embedding forward equation modules into the DL network,
total complexity is O(N .d)2 , which is more efficient than the
the combination of the DL framework and traditional inversion
standard Transformer complexity O(N 2 .d).
equations is realized, so that the DL network is transformed
from data driving to the physical equation driving, which
B. Unsupervised Learning Inversion Method achieves the effect of the approximate unsupervised learning
The general supervised learning method takes the complete and avoids the problem of requiring real formation parameters
seismic data as the input of the neural network, and the as labels. In the training process of network model, the
corresponding seismic inversion parameters are taken as labels, disappearance of gradient will seriously affect the convergence
which are expected to be output. Use a large number of of network model. In order to solve this problem, the layer
training datasets to pretrain the neural network, learn the normalization (LN) is added to the original network, with
characteristics of the training dataset, and then use the trained LN set before the network attention module and between the
network model to predict the seismic data to achieve the residual connections. The data in the network are normalized
purpose of inversion. However, in actual seismic data, it is by executing the LN. The above processing can effectively
difficult to obtain enough effective training samples, and the accelerate the learning speed of the neural network and
quality of datasets often has a great impact on the prediction increase the stability of the network as a whole while solving
results, so the application of this method is limited. In this arti- the gradient dispersion problem. Because the output data
cle, we propose an unsupervised prestack seismic brittleness of the network change violently and the seismic brittleness
parameter inversion method based on the physical equation parameters are sparse, in order to improve the overall inversion
driving. Its seismic brittleness parameter inversion process is effect, the L1 regularization is added to the loss function to
shown in Fig. 2. make the inversion results closer to the real value. The final
The proposed unsupervised learning method includes three loss function is
parts: inversion module, forward modeling module, and
Loss = ∥A PL − S∥2 + µ ∗ ∥PL − L 0 ∥2 + λ ∗ ∥w∥1 (14)
low-frequency constraint module. The actual training and
prediction steps are as follows: first, input the original seismic where λ is the regularization parameter, generally set to
data S into the inversion module embedding the Fastfomer 0.0001–0.001µ (empirical value obtained after many experi-
network, introduce forward equation operator A = W B D and ments), which decreases with increasing number of training

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
WU et al.: UNSUPERVISED INVERSION METHOD FOR SEISMIC BRITTLENESS PARAMETERS 5909013

Fig. 3. (a)–(c) 10◦ , 20◦ , and 30◦ multiangle seismic data.

epochs (the more training rounds, the smaller the setting), selected theoretical model is representative. In order to meet
and w is a network parameters. Our network architecture is the inversion requirement, take the logarithm of the real model
built on the PyTorch platform and uses general processing in Fig. 4 to obtain the theoretical model shown in Fig. 5, and
unit (GPU) (4 × 1080Ti) parallelism to speed up training. obtain the low-frequency initial model shown in Fig. 6 on the
We select the ADAM optimizer to train the network and set its basis of the model to be inverted. Construct forward operator
superparameters (initial learning rate = 1e–3, batch_size = 32, A, the multiangle seismic data and low-frequency model are
and decay_weight = 1e–5). When it comes to training epochs, sent into the network, and 50 epochs are trained, with a total
it is not better with much long training epochs. Generally, each of 1200 iterations.
batch size of data is iterated about 30–50 times to achieve The change of the loss function curve of the network model
convergence basically. with the number of iterations and the change of the error
Eρ inversion (prediction) accuracy A Eρ , σ inversion accu- between the network prediction model and the real model with
racy Aσ , and ρ inversion accuracy Aρ are defined as the number of iterations are shown in Fig. 7. It can be seen
 n
!, that as the number of iterations increases, the loss function
gradually decreases, and the predicted brittleness parameters
X
abs( pr e Eρ (i) − label Eρ (i))/label Eρ (i)


 A Eρ = 1− n
gradually approach the real model. After the training, the



 i
brittleness parameters are directly predicted by the saved

 n
!,
 X
Aσ = 1 − abs( pr eσ (i) − labelσ (i))/labelσ (i) n network model.
As shown in Fig. 8, the direct inversion results of the


 i !,
n

network basically reflect the change trend of the brittleness

 X
abs( pr eρ (i) − labelρ (i))/labelρ (i)

 Aρ = 1 − n parameters. Compared with Fig. 5(a) and (b), Eρ and σ


i
directly inverted through the network are of high accuracy,
(15)
and the inversion results in the model fault area are also
where A is the accuracy, pr e is the inversion result, is the relatively stable, with a high degree of coincidence with
actual value, and abs is the absolute value operator. the real model, good horizontal continuity, and high vertical
stratigraphic resolution. Fig. 8(c) shows the density inversion
IV. E XPERIMENTS results. Compared with Fig. 5(c), the density inversion results
can better reflect the lithological characteristics of the forma-
A. Synthetic Data Example tion. To sum up, the inversion results error of the method
To verify the inversion effect of the method, a part of proposed in this article is small and the accuracy is high.
Marmousi2 model is selected to generate reflection coef- In order to further compare the inversion results, we carried out
ficients of three incident angles, which are 10◦ , 20◦ , and the inversion using the traditional L1 method and compared
30◦ , respectively. Then, use the convolution of 30-Hz zero it with the network inversion results. Fig. 9(a)–(c) shows the
phase Ricker wavelet and reflection coefficient to synthesize traditional inversion results. In order to more directly highlight
the multiangle seismic data volume. There are 750 traces in the difference between the two methods, the absolute errors
total, with a sampling interval of 1 ms. Each trace includes between the two different inversion results and the real model
500 sampling points. Fig. 3(a)–(c) shows the angle seismic to be inverted are calculated, respectively. It can be observed
data volume with the incidence angles 10◦ , 20◦ , and 30◦ , from the inversion errors in Fig. 10 that in most regions,
respectively (three angle data are single angle gathers). The the inversion results of unsupervised depth learning method
three parameters of the theoretical model are shown in Fig. 4. have reduced the errors, especially in the red box, it can be
Fig. 4(a) is the real Eρ model, Fig. 4(b) is the real σ model, observed that the inversion results of this proposed method
and Fig. 4(c) is the real ρ model. The selected model shows have significantly reduced errors, the results are more stable,
a large variation of σ and Eρ laterally in the 400 ms section, and are more consistent with the original model.
which is more similar to the characteristics of the actual In order to analyze the inversion results in more detail, the
high-quality shale reservoir fracturing section. Therefore, the 150th and 600th traces are extracted from the real model and

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
5909013 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023

Fig. 4. (a)–(c) True Eρ, σ , and ρ.

Fig. 5. (a)–(c) True Eρ, σ , and ρ after logarithm.

Fig. 6. (a)–(c) Low frequency initial Eρ, σ , and ρ.

Fig. 7. (a) Loss function curve. (b) Absolute error between the prediction model and real model.

inversion results for comparative analysis, and the inversion the blue curve is the initial model, the black curve is the
results of the network are smoothed. The comparison results inversion results of the traditional L1 method, and the red
are shown in Fig. 11, where the green curve is the real model, curve is the inversion results of the unsupervised DL. From the

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
WU et al.: UNSUPERVISED INVERSION METHOD FOR SEISMIC BRITTLENESS PARAMETERS 5909013

Fig. 8. (a)–(c) Eρ, σ , and ρ obtained by UBPD.

Fig. 9. (a)–(c) Eρ, σ , and ρ obtained by the traditional L1 direct inversion method.

Fig. 10. (a)–(c) Absolute errors of Eρ, σ , and ρ of the UBPD inversion and the real model. (d)–(f) Absolute errors of Eρ, σ , and ρ of the traditional L1
inversion method and the real model.

single trace inversion results, it can be seen that the single trace is higher than that of the traditional L1 method, that is, the
inversion results of the unsupervised DL method are more con- method used in this article is closer to the real model, with
sistent with the real model single trace. Finally, the final result high precision and strong stability, indicating the effectiveness
can be obtained by the logarithmic inverse transformation of of the method. Compared with Eρ and σ , the ρ parameter
the network inversion results. From Table I, it can be seen that itself is slightly less sensitive to the inversion equation, so the
the inversion accuracy of the unsupervised learning method density prediction performance is relatively poor. At the same

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
5909013 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023

Fig. 11. (a)–(c) 150th trace inversion comparison on Eρ, σ , and ρ. (d)–(f) 600th trace inversion comparison on Eρ, σ , and ρ.

Fig. 12. (a)–(c) 10◦ , 20◦ , and 30◦ add noisy multi angle seismic data.

TABLE I TABLE II
I NVERSION ACCURACY C OMPARISON (750 T RACES AVERAGES ) S YNTHETIC DATA C ALCULATION E FFIENCY C OMPARISON

after adding noise. Input the noisy multiangle seismic data

time, under the condition of the same amount of inversion and the same low-frequency model into the network, train
data, as shown in Table II, the proposed method has higher 50 epochs, and iterate a total of 1200 times until the network
computing efficiency than the traditional algorithm (the time converges. After training, the stored network model is used
spent by unsupervised brittleness inversion method based on to directly predict the brittleness parameters. Figs. 13 and 14
the physical equation driving (UBPD) includes the time spent show the inversion results of the unsupervised method and the
in training. When the amount of inversion data is larger, the traditional L1 method after adding noise.
computational efficiency of UBPD will be more obvious.). It can be seen that noise has a certain influence on both
If there is a GPU graphics card with stronger computing inversion methods. Among them, the inversion results of
power, the acceleration efficiency can be further improved by the traditional L1 method are more fuzzy, and the noise
using compute unified device architecture (CUDA). has a great influence on the L1 method. In order to more
Finally, in order to verify the robustness of the proposed clearly reflect the differences between the two methods, the
method to noise, Gaussian white noise with an SNR of 10 was absolute errors between the two different inversion results
added to the synthetic seismic data for noise suppression and the actual model to be inverted were calculated. From
experiments. Fig. 12(a)–(c) shows the multiangle seismic data the absolute error Fig. 15, in the red box, it is clear that

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
WU et al.: UNSUPERVISED INVERSION METHOD FOR SEISMIC BRITTLENESS PARAMETERS 5909013

Fig. 13. (a)–(c) After adding noise, Eρ, σ , and ρ obtained by UBPD.

Fig. 14. (a)–(c) After adding noise, Eρ, σ , and ρ obtained by the traditional L1 direct inversion method.

Fig. 15. (a)–(c) After adding noise, the absolute errors of Eρ, σ , and ρ of the UBPD inversion and the real model. (d)–(f) After adding noise, the absolute
errors of Eρ, σ , and ρ of the traditional L1 inversion method and the real model.

the error value of the new method is slightly smaller than higher inversion accuracy compared to the traditional meth-
that of the traditional L1 method. Experiments show that the ods, which further illustrates the advantages of the proposed
proposed unsupervised method has certain noise resistance and method.

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
5909013 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023

Fig. 16. (a)–(c) 12◦ , 24◦ , and 36◦ multiangle seismic data.

Fig. 17. (a)–(c) Low-frequency initial Eρ, σ , and ρ.

B. Field Data Example

The proposed method is applied to the inversion of brit-
tleness parameters in a field data. These data are collected
from a shale work area in Sichuan Basin, China. We have
shown some angle superimposed profiles. Fig. 16(a)–(c) shows
the seismic data with incident angles of 12◦ , 24◦ , and 36◦ ,
respectively. (The actual data here are angle superimposed
data, e.g., 12◦ data represent data superimposed within the
range of 0◦ –12◦ . The maximum angle of this data is 36◦ .) The
profile has 800 traces of data, with a sampling time period of
1200–1920 ms and a sampling interval of 4 ms. The upper and Fig. 18. Loss function curve.
lower boundary lines of the profile represent two horizons of
geological interpretation, and the black dotted line represents
the well location. traditional L1 method, and commercial software, respectively.
First, extracting statistical wavelet from the original seismic The inversion resolution of the commercial software is slightly
data to construct the forward operator A, and then, the low- lower, the overall inversion results are consistent with the
frequency model, as shown in Fig. 17(a)–(c), is obtained by other two methods, and the inversion results of some regions
interpolating the data of multiple wells in the work area. are slightly different from the other two methods. Compared
Next, the multiangle seismic data and the logarithmic original with the traditional L1 method, the overall resolution of the
low-frequency model are sent into the network, and 50 epochs inversion results of the unsupervised DL method has been
are trained, with a total of 1200 iterations. The loss function greatly improved. In order to facilitate observation, some
curve of the network model is shown in Fig. 18. It can data in the black box are magnified. It is obvious that the
be observed from the figure that the loss function gradually inversion results of the proposed method have better horizontal
decreases with the increase of the number of iterations. After continuity, and the vertical resolution is effectively improved,
the training, the stored network model is used to directly which can effectively depict the strati-graphic characteristics.
predict the brittleness parameters. In order to further verify the reliability of inversion, the
For better comparison, we also provide indirect inversion inversion results of well bypass are extracted and compared
results of a commercial software. Figs. 19–21 show the with the logging data. The inversion comparison results of
final inversion results of the unsupervised learning method, the three methods are shown in Fig. 22. It can be seen in

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
WU et al.: UNSUPERVISED INVERSION METHOD FOR SEISMIC BRITTLENESS PARAMETERS 5909013

Fig. 19. (a)–(c) Eρ, σ , and ρ obtained by UBPD.

Fig. 20. (a)–(c) Eρ, σ , and ρ obtained by the traditional L1 direct inversion.

Fig. 21. (a)–(c) Eρ, σ , and ρ obtained by the commercial software indirect inversion.

Fig. 22 that the trend of the results of unsupervised learning TABLE III
method and traditional L1 method is relatively consistent with F IELD DATA C ALCULATION E FFICIENCY C OMPARISON
the logging data, but it is not difficult to observe carefully, the
inversion results of the unsupervised learning method is closer
to the logging data, and the resolution is significantly higher.
However, the indirect inversion results of the commercial
software are generally consistent with the logging data. This
is because the inversion parameters are not directly inverted,
but converted from other parameters obtained from inversion,
which results in an increase in the error of inversion results area, the unsupervised learning method can more accurately
and a decrease in inversion accuracy. Therefore, the reliability invert the characteristics of the reservoir, with less error and
of the direct inversion method is better than that of the higher accuracy, which further proves the correctness and
indirect inversion method. According to the inversion results reliability of the proposed method. Similarly, in the case of
and the actual exploration in the work area, the location the same amount of inversion data, as shown in Table III,
of the reservoir section is located at 1650–1680 ms, and the proposed method is more efficient than the traditional
the red box is a high-quality shale reservoir area. In this algorithm, indicating the advantages of the proposed method.

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
5909013 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 61, 2023

[6] Z. Guo, X.-Y. Li, C. Liu, X. Feng, and Y. Shen, “A shale rock
physics model for analysis of brittleness index, mineralogy and porosity
in the Barnett Shale,” J. Geophy. Eng., vol. 10, no. 2, Apr. 2013,
Art. no. 025006.
[7] G. Z. Zhang, B. Y. Du, H. S. Li, H. Z. Chen, Z. Z. Li, and X. Y. Yin,
“The method of joint pre-stack inversion of PP and P-SV waves in
shale gas reservoirs,” Chin. J. Geophys., vol. 57, no. 12, pp. 4141–4149,
Dec. 2014.
[8] K. S. Ritesh and C. Satinder, “Determining brittleness from seismic
data,” AAPG Explorer, vol. 36, no. 10, pp. 40–44, 2015.
[9] R. Zhang, X. T. Wen, J. X. Yang, L. Li, and S. Liu, “Two-term reflection
coefficient equation with Young’s modulus and Poisson ratio and its
pre-stack seismic inversion,” Oil Geophys. Prospecting, vol. 54, no. 1,
pp. 145–153, 2019.
[10] Z. Ge, S. Pan, J. Liu, J. Zhou, and X. Pan, “Estimation of brittleness
and anisotropy parameters in transversely isotropic media with vertical
axis of symmetry,” IEEE Trans. Geosci. Remote Sens., vol. 60, 2022,
Art. no. 5904509.
[11] A. N. Tikhonov, “Regularization of incorrectly posed problems,” Soviet
Math. Doklady, vol. 4, no. 6, pp. 1624–1627, 1963.
Fig. 22. Comparison between the partial inversion results of well bypass
and well data (green line—well, blue line—commercial software, black [12] Y. Wang, “Seismic impedance inversion using ℓ1 -norm regularization
line—traditional L1 method, and red line—UBPD method). and gradient descent methods,” Jiip, vol. 18, no. 7, pp. 823–838,
Dec. 2010.
[13] F. Zhang, R. Dai, and H. Liu, “Seismic inversion based on L1-norm
V. C ONCLUSION misfit function and total variation regularization,” J. Appl. Geophys.,
In this article, we proposed an unsupervised prestack seis- vol. 109, pp. 111–118, Oct. 2014.
[14] S. Li, Y. He, Y. Chen, W. Liu, X. Yang, and Z. Peng, “Fast multi-trace
mic brittleness parameter inversion method driven by the phys- impedance inversion using anisotropic total p-variation regularization in
ical equation. Compared with the conventional DL method, the frequency domain,” J. Geophys. Eng., vol. 15, no. 5, pp. 2171–2182,
it does not require actual label training. In the synthetic data Oct. 2018.
[15] L. He, H. Wu, X. Wen, and J. You, “Seismic acoustic impedance
test set, the brittleness parameters inverted by the unsupervised inversion using reweighted L1-norm sparse constraint,” IEEE Geosci.
DL Fastformer network are basically consistent with the real Remote Sens. Lett., vol. 19, pp. 1–5, 2022.
brittleness parameters and have a certain anti-noise ability, [16] R. Dai, C. Yin, and D. Peng, “AVA inversion of pre-stack seismic data
which verifies the correctness of this method. The inversion with cross-gradient constraints,” J. Appl. Geophys., vol. 199, Apr. 2022,
Art. no. 104594.
results of the field data show that the inversion results have [17] R. Dai and J. Yang, “Amplitude-versus-angle (AVA) inversion for pre-
high resolution and are in good agreement with the well data stack seismic data with L0-Norm-Gradient regularization,” Mathematics,
in the work area, which further proves the practicability and vol. 11, no. 4, p. 880, Feb. 2023.
[18] Y. Chen, S. Pan, Y. Wu, Z. Wei, and G. Song, “Direct inversion method
reliability of the method. of brittleness parameters based on reweighted Lp-norm,” Appl. Sci.,
The proposed method is suitable for almost all approximate vol. 13, no. 1, p. 246, Dec. 2022.
prestack parameter inversion and avoids the disadvantages of [19] S. Y. Yuan, J. Liu, S. Wang, T. Wang, and P. Shi, “Seismic waveform
the traditional direct inversion methods to some extent. At the classification and first-break picking using convolution neural networks,”
IEEE Geosci. Remote Sens. Lett., vol. 15, no. 2, pp. 1–5, Jan. 2018.
same time, the DL of the physical equation guarantees the [20] S. M. Mousavi and G. C. Beroza, “Bayesian-deep-learning estimation
fidelity of the inversion results, makes full use of the powerful of earthquake location from single-station observations,” IEEE Trans.
learning ability of the network and parallel acceleration of Geosci. Remote Sens., vol. 58, no. 11, pp. 8211–8224, Nov. 2020.
[21] S. Wang, Y. Li, N. Wu, Y. Zhao, and H. Yao, “Attribute-based double
GPU, and has high computing efficiency and great application constraint denoising network for seismic data,” IEEE Trans. Geosci.
potential. The proposed method has better inversion effective- Remote Sens., vol. 59, no. 6, pp. 5304–5316, Jun. 2021.
ness under the same conditions. It should be noted that the [22] G. Fabien-Ouellet and R. Sarkar, “Seismic velocity estimation: A deep
extraction of wavelet in the application of the actual data has recurrent neural-network approach,” Geophysics, vol. 85, no. 1,
pp. 21–29, Jan. 2020.
a certain impact on the actual application effect, and when the [23] X. Wu, L. Liang, Y. Shi, and S. Fomel, “FaultSeg3D: Using synthetic
initial model becomes worse, it will affect the final inversion data sets to train an end-to-end convolutional neural network for 3D
results. We will further pay attention to these issues in the seismic fault segmentation,” Geophysics, vol. 84, no. 3, pp. 35–45,
May 2019.
follow-up study. [24] F. Wang, X. Wu, and H. Wang, “Seismic horizon identification using
semi-supervised learning with virtual adversarial training,” IEEE Trans.
R EFERENCES Geosci. Remote Sens., vol. 60, 2022, Art. no. 4508611.
[1] J. Xu, K. Y. Liu, and Q. Z. Wu, “Evaluation and prediction of shale [25] E. Tolstaya and A. Egorov, “DL for automated seismic facies classifi-
brittleness in the Jiaoshiba area,” Geophys. Prospecting Petroleum, cation,” Interpretation, vol. 10, no. 2, pp. 31–40, Mar. 2022.
vol. 58, no. 3, pp. 456–460, Jan. 2019. [26] X. Zhang, Z. Jia, Z. E. Ross, and R. W. Clayton, “Extracting dispersion
[2] K. Aki and P. G. Richards, Quantitative Seismology. San Francisco, CA, curves from ambient noise correlations using deep learning,” IEEE
USA: W. H. Freeman, 1999, pp. 5–43. Trans. Geosci. Remote Sens., vol. 58, no. 12, pp. 8932–8939, Dec. 2020.
[3] R. T. Shuey, “A simplification of the Zoeppritz equations,” Geophysics, [27] Z. Wang, C. Sun, and D. Wu, “Automatic picking of multi-mode surface-
vol. 50, no. 4, pp. 609–614, Apr. 1985. wave dispersion curves based on machine learning clustering methods,”
[4] B. Goodway, M. Perez, J. Varsek, and C. Abaco, “Seismic petrophysics Comput. Geosci., vol. 153, Aug. 2021, Art. no. 104809.
and isotropic-anisotropic AVO methods for unconventional gas explo- [28] J. Hu, H. Qiu, H. Zhang, and Y. Ben-Zion, “Using deep learning to
ration,” Lead. Edge, vol. 29, no. 12, pp. 1500–1508, Dec. 2010. derive shear-wave velocity models from surface-wave dispersion data,”
[5] Z. Y. Zong, X. Y. Yin, F. Zhang, and G. C. Wu, “Reflection coefficient Seismolog. Res. Lett., vol. 91, no. 3, pp. 1738–1751, May 2020.
equation and pre-stack seismic inversion with Young’s modulus and [29] Y. Wu, S. Pan, J. Chen, G. Song, and Q. Gou, “A surface-wave inversion
Poisson ratio,” Chin. J. Geophys., vol. 55, no. 11, pp. 3786–3794, method based on FHLV loss function in LSTM,” IEEE Geosci. Remote
Nov. 2012. Sens. Lett., vol. 19, pp. 1–5, 2022.

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.
WU et al.: UNSUPERVISED INVERSION METHOD FOR SEISMIC BRITTLENESS PARAMETERS 5909013

[30] R. Biswas, M. K. Sen, V. Das, and T. Mukerji, “Prestack and post- Jingyi Chen received the Ph.D. degree in solid
stack inversion using a physics-guided convolutional neural network,” geophysics from the Institute of Geology and
Interpretation, vol. 7, no. 3, pp. 161–174, Aug. 2019. Geophysics, Chinese Academy of Sciences, Beijing,
[31] V. Das, A. Pollack, U. Wollner, and T. Mukerji, “Convolutional neural China, in 2005.
network for seismic impedance inversion,” Geophysics, vol. 84, no. 6, He is currently a Decker Dawson Professor of
pp. 869–880, Nov. 2019. geophysics with The University of Tulsa, Tulsa,
[32] B. Wu, Q. Xie, and B. Wu, “Seismic impedance inversion based on OK, USA, where he is also the Director of the
residual attention network,” IEEE Trans. Geosci. Remote Sens., vol. 60, Seismic Anisotropy Group. He was a Post-Doctoral
2022, Art. no. 4511117. Fellow with Memorial University of Newfoundland,
[33] Y.-Q. Wang, Q. Wang, W.-K. Lu, Q. Ge, and X.-F. Yan, “Seismic St. John’s, NL, Canada, before joining The Univer-
impedance inversion based on cycle-consistent generative adversarial sity of Tulsa. His research interests focus on both
network,” Petroleum Sci., vol. 19, no. 1, pp. 147–161, Feb. 2022. conventional and unconventional reservoirs including numerical simulation
[34] H. Zhang, G. Zhang, J. Gao, S. Li, J. Zhang, and Z. Zhu, “Seis- of seismic wave propagation, reverse time migration (RTM), full waveform
mic impedance inversion based on geophysical-guided cycle-consistent inversion (FWI), DL, amplitude versus offset (AVO), vertical seismic profile
generative adversarial networks,” J. Petroleum Sci. Eng., vol. 218, (VSP), microseismic, multiple-component seismic data processing, and reser-
Nov. 2022, Art. no. 111003. voir characterization. His research has been supported by oil/gas companies,
[35] L. Song, X. Yin, Z. Zong, and M. Jiang, “Semi-supervised learning professional society, and the National Science Foundation.
seismic inversion based on spatio-temporal sequence residual mod-
eling neural network,” J. Petroleum Sci. Eng., vol. 208, Jan. 2022,
Art. no. 109549.
[36] A. Vaswani et al., “Attention is all you need,” in Proc. NIPS, 2017,
pp. 5998–6008.
[37] C. Wu, F. Wu, T. Qi, Y. Huang, and X. Xie, “Fastformer: Additive
attention can be all you need,” 2021, arXiv:2108.09084.

Shengbo Yi was born in Qitaihe, Heilongjiang,

China, in 1996. He received the bachelor’s degree
in exploration technology and engineering from
Yinghe Wu was born in Bazhong, Sichuan, China, Northeast Petroleum University, Daqing, China,
in 1999. He received the bachelor’s degree in explo- in 2018. He is currently pursuing the master’s degree
ration technology and engineering from Southwest in geological resources and geological engineer-
Petroleum University, Chengdu, China, in 2020, ing from Southwest Petroleum University, Chengdu,
where he is currently pursuing the Ph.D. degree in China.
geological resources and geological engineering. His research interests include geophysical inver-
His research interests include seismic data pro- sion, fluid identification, and fractures identification.
cessing and imaging, geophysical inversion, passive
imaging, and deep learning (DL).

Shulin Pan was born in Dezhou, Shandong,

China, in 1979. He received the M.S. and Ph.D.
degrees in geophysical prospecting and information Dongjun Zhang was born in Wugang, Hunan,
technology from the Chengdu University of Tech- China, in 1982.
nology, Chengdu, Sichuan, China, in 2005 and 2008, He is currently a Senior Engineer of PetroChina
respectively. Southwest Oil & Gasfield Company, Chengdu,
Since 2011, he has been a Professor with China. His research interests include the research
Southwest Petroleum University, Chengdu. His and application of shale gas seismic data processing
research interests include automatic first-arrival time and interpretation methods.
picking, Seismic data interpolation and ampli-
tude versus incident and azimuthal angle (AVAZ)
inversion. His research has been supported by the Science and Tech-
nology Cooperation Project of the CNPC-SWPU Innovation Alliance
(2020CX020000), the Open Projects Fund of the State Key Laboratory of Oil
and Gas Reservoir Geology and Exploitation (Grant no. PLN201733), and
the National Natural Science Foundation of China (Grant nos. 41204101 and
41704134).

Yaojie Chen was born in Changzhi, Shanxi, Guojie Song received the M.S. and Ph.D. degrees
China, in1998. He received the bachelor’s degree from the Department of Mathematical Science,
in resource exploration engineering from the Shanxi Tsinghua University, Beijing, China, in 2008 and
Institute of Energy, Jinzhong, China, in 2020. 2011, respectively.
He is currently pursuing the master’s degree in He is currently a Professor with Southwest
geological resources and geological engineering with Petroleum University, Chengdu, China. His research
Southwest Petroleum University, Chengdu, China. interests include forward modeling and inversion of
His research interests include geophysical inver- seismic wave and deep learning.
sion and reservoir prediction.

Authorized licensed use limited to: McFarlin Library The University of Tulsa. Downloaded on November 19,2024 at 17:31:08 UTC from IEEE Xplore. Restrictions apply.