Physics-Informed Data-Driven Seismic Inversion: Recent Progress

and Future Opportunities

Youzuo Lin1 , James Theiler1 , and Brendt Wohlberg1
1
Los Alamos National Laboratory

November 30, 2022

Abstract
The goal of seismic inversion is to obtain subsurface properties from surface measurements. Seismic images have proven valuable,
even crucial, for a variety of applications, including subsurface energy exploration, earthquake early warning, carbon capture
and sequestration, estimating pathways of sub-surface contaminant transport, etc. These subsurface properties˜(such as wave
speed, density, impedance, and reflectivity) influence the transmission of seismic waves through the subsurface media, and well-
understood physics models (so-called “forward models”) can be used to predict what surface measurements would be made,
Posted on 30 Nov 2022 — CC-BY 4.0 — https://doi.org/10.1002/essoar.10511175/v1 — This a preprint and has not been peer reviewed. Data may be preliminary.

for any given subsurface configuration. Seismic inversion is the inverse problem: given actual surface measurements, infer what
subsurface configuration would give rise to those measurements. Like most inverse problems, seismic imaging is ill-posed. There
is more below the surface than on the surface, and many different subsurface configurations can give rise to the same surface
measurements. Because the forward model is itself computationally expensive – the inverse inference is even more so. But
recent advances in algorithms and computing provide an opportunity for remarkable progress in seismic inversion, and efficient
solutions to previously infeasible problems have been obtained using data-driven approaches (such as the deep learning methods
that were developed primarily for problems in computer vision). The excellent performance of learning-based methods arises
from its ability to exploit large amounts of high-quality training data, without the need for hand-designed features. Unlike
computer vision, however, seismic inversion is not a data-rich domain. The technical challenges and high cost of acquiring field
data prevent the accumulation of large quantities of data (and the commercial value of that data, once acquired) prevents much
of the existing data from being widely available. To alleviate the data scarcity issue and improve model generalization, there
has been growing interest in combining physics knowledge with machine learning for solving seismic inversion problems. This
review will survey methods for incorporating physics knowledge with machine learning (primarily deep neural networks) to
solve computational seismic inversion problems. We will provide a structured framework of the existing research in the seismic
inversion community, and will identify technical challenges, insights, and trends.

1
 1

Physics-Informed Data-Driven Seismic Inversion

Recent Progress and Future Opportunities

Youzuo Lin, James Theiler, and Brendt Wohlberg

Abstract

The goal of seismic inversion is to obtain subsurface properties from surface measurements. Seismic
images have proven valuable, even crucial, for a variety of applications, including subsurface energy
exploration, earthquake early warning, carbon capture and sequestration, estimating pathways of sub-
surface contaminant transport, etc. These subsurface properties (such as wave speed, density, impedance,
and reflectivity) influence the transmission of seismic waves through the subsurface media, and well-
understood physics models (so-called “forward models”) can be used to predict what surface mea-
surements would be made, for any given subsurface configuration. Seismic inversion is the inverse
problem: given actual surface measurements, infer what subsurface configuration would give rise to
those measurements. Like most inverse problems, seismic imaging is ill-posed. There is more below the
surface than on the surface, and many different subsurface configurations can give rise to the same surface
measurements. Because the forward model is itself computationally expensive – the inverse inference is
even more so.
But recent advances in algorithms and computing provide an opportunity for remarkable progress
in seismic inversion, and efficient solutions to previously infeasible problems have been obtained using
data-driven approaches (such as the deep learning methods that were developed primarily for problems in
computer vision). The excellent performance of learning-based methods arises from its ability to exploit
large amounts of high-quality training data, without the need for hand-designed features. Unlike computer
vision, however, seismic inversion is not a data-rich domain. The technical challenges and high cost of
acquiring field data prevent the accumulation of large quantities of data (and the commercial value of
that data, once acquired) prevents much of the existing data from being widely available. To alleviate
the data scarcity issue and improve model generalization, there has been growing interest in combining
physics knowledge with machine learning for solving seismic inversion problems.
This review will survey methods for incorporating physics knowledge with machine learning (pri-
marily deep neural networks) to solve computational seismic inversion problems. We will provide a

Y. Lin, J. Theiler, and B. Wohlberg are with Los Alamos National Laboratory, Los Alamos, NM, 87544 USA
Manuscript was received March, 2022.
 2

structured framework of the existing research in the seismic inversion community, and will identify
technical challenges, insights, and trends.

Index Terms

Scientific Machine Learning, Physics Informed Neural Network, Seismic Inversion, Subsurface Char-
acterization, Computational Imaging, Inverse Problems

I. I NTRODUCTION

WHILE we can’t directly observe the geology of the earth beneath our feet, knowledge of its properties
is vital for a variety of applications, including knowing where to build or drill for water or hydrocarbons.
From civil infrastructure to energy exploration, characterizing subsurface geology is crucial, but since
direct observation of the earth’s interior is nearly impossible, inference of unknown subsurface properties
primarily relies on indirect and limited geophysical measurements taken at or near the surface. Seismic
inversion attempts to reconstruct an image of the subsurface from measurements of natural or artificially
produced seismic waves that have traveled through the subsurface medium. A forward model describes
how the observations depend on the subsurface map, while the inverse problem is the infererence of that
map from the observations. The forward model of seismic wave propagation is governed by the wave
equations, in the form of a partial differential equation (PDE). A linear approximation can be utilized
to simplify the problem, at the cost of loss of accuracy and resolution, leading to travel-time inversion
methods [1]. Seismic full-waveform inversion (FWI) addresses the full non-linear problem, providing
superior inversion accuracy and resolution [2], but at considerably greater computational cost. Given the
large volume of publications in this field, our review focuses on non-linear seismic FWI techniques over
other seismic inverse problems (such as travel-time tomography) based on the following considerations:

• Superior Performance. FWI leverages extensively more waveform physics than other seismic
inversion techniques, resulting in significantly improved imaging results. The technique has been
widely applied to various subsurface applications ranging from reservoir-scale characterization [3]
to global-scale tomography of Earth’s crust and mantle [4].
• Significant Technical Impact. FWI faces similar technical challenges to many other ill-posed inverse
problems. Our focus here is on how to leverage the underlying governing wave physics (as shown
in Fig. 1a), but the techniques discussed here can be not only extendable to other wave imaging
problems but also provide potential solutions to a broader context of PDE-governed inverse problems.
• Active Research. FWI is one of the most active research topics in subsurface geophysics. A variety
of new computational methods have emerged recently (see Fig. 1b for solving the FWI problem.
 3

(a) (b)

Fig. 1: (a) Physics-informed data-driven seismic imaging. The wave equation is the governing physics.
(b) Trends of physics-guided machine learning techniques in computational seismic imaging. Six major
technical categories can be identified out of a total 107 papers from 2017 to 2021 collected from the
Web of Science and Google Scholar.

This forms an excellent foundation for us to provide a broad and in-depth review of the existing
literature.

Seismic full-waveform inversion is a nonlinear inverse problem. Approaches for solving this problem
can be categorized into two groups: physics-based methods and machine learning-based methods. Physics-
based methods primarily rely on nonlinear optimization techniques that are computationally intensive
and highly sensitive to data coverage1,2 . Once developed, however, they can be robust and generalize
quite naturally to new datasets. In part to alleviate the computational expense of physics-driven methods,
data-driven inversion techniques have recently been proposed [5], [6], [7], [8] and reviewed [3]. The data-
driven techniques can be extremely efficient once fully trained, but the generalizability of the data-driven
approaches is limited by the size and range of the training set. In addition, because data acquisition
in subsurface geoscience is always expensive and time-consuming, and because observations can be
incomplete and are inevitably contaminated by noise, large training sets are difficult to acquire. Further,
because data-driven models learn from data directly without leveraging domain knowledge, those models
are prone to unreasonable or unrealistic predictions that do not conform to the physical mechanism.
Efforts have been made to combine the physical consistency and generalization ability of physics-
based modeling with the computational efficiency of data-driven approaches; these hybrid algorithms

1
Data coverage: the spatial extent to which a subsurface region is covered by a seismic survey.
2
We provide a summary of the seismic terminologies used in the review (see Table II).
 4

Reference Year Application Domain Physics ML PIML

Virieux and Operto [2] 2009 Exploration Geophysics ✓
Tromp [4] 2020 Global Geophysics ✓
Bergen et al. [9] 2019 General Applications in Solid Earth ✓
Gregory et al. [10] 2020 Computational Imaging and Inverse Problems ✓
Yu and Ma [11] 2021 General Applications in Solid Earth ✓
Adler et al. [3] 2021 Computational Seismic Inverse Problems ✓
Willard et al. [12] 2020 Engineering and Environmental Problems ✓
Karniadakis et al. [13] 2021 Scientific and Engineering Problems ✓

TABLE I: A summary of relevant review articles in the topics of physics-based seismic FWI for
exploration geophysics [2] and global geophysics [4], pure ML approaches for general geoscience [9], [11]
and computational imaging [10], pure ML-based seismic inverse problems [3], PIML for environmental
problems [12], and for broad scientific and engineering inverse problems [13].

are collectively referred to as physics-informed machine learning (PIML). There is ongoing work in the
seismic inversion community to develop new models that combine physics-driven methods with machine
learning techniques. This review paper will focus on this new trend. The strengths and limitations of
these techniques will be discussed based on some numerical examples.
Several review articles relevant to to this topic have recently been published. We summarize a number
of them in Table I, and categorize them into three groups: physics-based seismic imaging techniques,
machine learning methods (ML), and physics-informed machine learning (PIML).

• Physics-based. Virieux and Operto [2] provide a thorough overview of physics-based FWI techniques
in exploration geophysics, while Tromp [4] gives a different perspective in an overview focusing on
applications of FWI for global seismology problems.
• Machine Learning. We have found four particularly noteworthy review papers that focus on data-
driven methods. Bergen et al. [9] provide a summary of machine learning models for solving general
solid Earth geosciences problems ranging from earthquake physics to data visualization, with a focus
on existing supervised and unsupervised ML models developed for subsurface applications. Gregory
et al. [10] provide an overview on the broad topics of deep learning for computational imaging
and inverse problems, concentrating mainly on methodology development rather than on specific
imaging problems. In contrast, Yu and Ma [11] explore deep neural networks, and emphasize a broad
range of applications in geophysics, including geophysical data processing, interpretation, earthquake
detection, and tomography. Adler et al. [3] provide an overview of recent machine learning-based
 5

Terminology Description in Plain Language

Seismic Data Coverage The spatial extent to which a subsurface region is covered by a seismic survey.
Velocity Map A spatial map of the local speed of the seismic wave through the subsurface medium.
Well logs present the detailed record of the geologic information (such as acoustic velocity,
Well Logs electrical resistivity, gamma ray, etc.) penetrated by a borehole.
Reverse time migration is a seismic processing technique that is able to estimate
Reverse Time Migration limited subsurface structural information, such as the locations of layer boundaries.
Prestack Seismic Data Prestack seismic data is the original raw seismic record.
Post-stacked seismic data is a processed seismic record that contains seismic waveforms that
Post-stacked Seismic Data have been added together from different records to reduce noise and improve overall data quality.

TABLE II: A summary of seismic terminology used in this review.

methods for solving different computational seismic inverse problems including velocity, impedance,
reflectivity model building, and seismic bandwidth extension. The focus of their paper is mostly on
pure data-driven deep learning approaches with applications to exploration geophysics.
• PIML. The literature for physics-informed machine learning is just emerging in geoscience. Nei-
ther of the two relevant reviews [14], [13] focuses specifically on computational seismic imaging.
Willard et al. [14] discuss the use of physics-informed neural networks (PINN) in engineering and
environmental problems. Karniadakis et al. [13] provide a more high-level overview of PINN with
applications to inverse problems in different domains such as biophysics, quantum chemistry, material
sciences, and molecular simulations.

The remainder of this paper is organized as follows. Section II (Current Research and Development)
provides a high-level overview of the physics-based and data-driven computational seismic imaging
methods. Section III (Physics-Informed Data-Driven Computational Seismic Imaging) gives a summary of
existing hybrid models incorporating physics with machine learning techniques. Section IV (Challenges
and Future Research Opportunities) provides a perspective of the existing challenges and some future
opportunities. Finally, Section V (Conclusions) summarizes our paper.

II. C URRENT R ESEARCH AND D EVELOPMENT

A. Physics-Based Seismic Imaging Techniques

Physics-driven methods infer subsurface properties by directly employing the governing physics and
equations. Seismic waves are mechanical perturbations that travel in the medium at a speed governed by
 6

the acoustic/elastic impedance of the material through which they are traveling. In the time domain, the
acoustic wave equation is given by
1 ∂2
∇2 p(r, t) − p(r, t) = s(r, t) , (1)
c2 (r) ∂t2
where t denotes time, r = (x, z) represents the spatial location in Cartesian coordinates (x is the horizontal
direction and z is the depth), ∇2 is the Laplacian operator, and c(r) is the velocity map (that is, the
speed of the wave through the medium at position r ). Here, p(r, t) is the pressure wavefield, and s(r, t)
is the source term that specifies the location and time history of the source. In the context of this forward
model, if we know the velocity map c(r) and the source term s(r, t), we can use Eq. (1) to solve for
the pressure field p(r, t). Thus, we can write

d = f (m) , (2)

to express the deterministic relationship between the model m, which encapsulates the subsurface velocity
map c(r), and the observed data d, which typically consists of time histories of the pressure wavefield
p(r, t) at a discrete set of locations r , usually on the surface. The goal of the inverse problem is to infer
the subsurface model m given observed data d. This can be posed as a minimization problem [15]
n o
E(m) = min ∥d − f (m)∥22 + λ R(m) , (3)
m

where ∥d − f (m)∥22 is the data fidelity term with ||·||2 denoting the ℓ2 norm, and R(·) is the regularization
term with regularization parameter λ. Solving the minimization in Eq. (3) can be numerically challenging
or even infeasible due to the ill-posed nature of the inversion problem (i.e. even small noise levels in the
measurements d can yield large uncertainties in the model m). For the FWI problem, this can arise from
limited data coverage (too few measurements of the pressure wavefield) and in particular from cycle skip-
ping (multiple solutions matched to the same data). Many approaches have been proposed to alleviate this
problem. These includes advanced regularization [15], dynamic warping techniques [16], incorporation of
prior information [17], multiscale inversion approaches [18], and preconditioning techniques [19]. Another
issue with current computational methods for solving FWI is the high computational cost. Most of the
current techniques are based on first-order or second-order optimization techniques. Both of these types
of methods are expensive since they require computation of the gradient of the function Eq. (3), which
requires a forward wave propagation, a backward wave propagation, and full-wavefield cross-correlation.
A line search strategy, requiring additional steps of forward wave propagation, is also necessary.

B. Data-Driven Seismic Imaging Techniques

We begin this section with the remark that purely data-driven techniques are unlikely ever to work
for seismic imaging; there just isn’t enough data. A lot of wells have been drilled over the last century,
 7

Fig. 2: An illustration of InversionNet, a data-driven seismic waveform inversion solver [6].

and so we have a reasonable idea about what the subsurface is like in general. But for a given site,
there is precious little in the way of direct observations of the subsurface. When we speak of data-driven
methods being applied to the seismic imaging problem, we are virtually always speaking of simulated
data. As with traditional “physics-based” methods described in the previous subsection, we assume we
know the governing equation for acoustic waves, and that we can readily simulate the function f (·)
described in Eq. (2). Given this, the key idea for data-driven methods is to create many samples of image
pairs: of velocity model m-images along with matched samples of waveform data d-images. There is
some art to the creation of the m-images – these are chosen to be representative of what we think the
subsurface might be like. But once we have a collection of m-images, we can use d = f (m) to create
the associated d-images. With all of this simulated data, we can at last apply some data-driven analysis.
A large image-to-image deep neural net (as shown in Fig. 2) is used to learn a direct mapping of d to
m, which can finally be used to map an actual data observation do to an estimate of the subsurface map.
Mathematically, this can be posed as

m
c(d) = gθ∗ (d)
X (4)
s.t. θ ∗ (Φs ) = arg min L(gθ (di ), mi ) ,
θ
(mi ,di )∈Φs

where: θ represents the trainable weights in the inversion network gθ (·); L(·, ·) is a loss function; and
Φs is the training dataset with paired samples of (mi , di ).
A distinctive feature of this approach is that, in contrast with traditional inversion, it produces a
 8

“whole function” gθ∗ (·) ≈ f −1 (·), not just a single inversion m

cobs ≈ f −1 (dobs ). Nonetheless, experience
with data-driven methods indicates that they can be trained in substantially less time than it takes for a
single traditional inversion [6]; and once trained, the application to a specific observation dobs is nearly
instantaneous. The reason for this efficiency is not fully understood, but some factors that potentially
contribute include: the simulations are only of the forward model, and no gradients (or Hessians) are
computed; the simulations are restricted to plausible models, so no time is spent (as may be the case for
traditional optimizers) in implausible regions of solution space; traditional optimization is inherently serial,
whereas the many simulations can be readily parallelized, and neural network architectures naturally lend
themselves to efficient massive parallelization. Further, because the function f −1 (d) is learned directly
and implicit regularization is imposed via training on a large volume of simulations, we do not see the
ill-posedness that traditional inversion often exhibits; on the other hand, solutions tend to look like the
models m that they were trained on, so the generalization ability of this approach is limited.
Most recent explorations have focused on data-driven end-to-end reconstruction. For instance, GeoDNN [7]
used a fully connected network with 8 layers for FWI. InversionNet [6] introduced a encoder-decoder
network with modern convolutional neural network (CNN) features. ModifiedFCN [8] employed a fully
connected convolutional network (FCN) design and SeisInvNet [5] enhanced each seismic waveform with
auxiliary knowledge from neighborhood traces for better spatial correspondence.

III. P HYSICS -I NFORMED DATA -D RIVEN C OMPUTATIONAL S EISMIC I MAGING

In this section we discuss a variety of computational techniques that have been developed for incor-
porating physics in seismic imaging methods.

A. Physics Knowledge

Physics knowledge can be represented in many different ways. What makes seismic imaging problems
fundamentally different from those of the computer vision community are the physical laws underlying
those problems. These physical laws are “invariant” in the sense that the spatial and temporal dynamics
can be characterized by those laws throughout the physical process. Here we provide two different
representations of physics knowledge that has been utilized to enhance the machine learning model.

1) Imaging Physics. The most directly relevant physics is expressed by the governing equations
underlying the imaging process. For seismic imaging, this is the partial differential equation in
Eq. (1) that describes the propagation of acoustic waves through the subsurface.
2) Physical Properties of the Solution Space. To the extent that we understand the geophysics of
the subsurface, we can constrain the solution space to be consistent with those physical properties,
 9

Fig. 3: An illustration of PIML-based seismic imaging techniques, which can be grouped into two
categories: those built on ML models (blue hexagons) and those built on physics models (green hexagons).

and we can more effectively choose “plausible” samples from the solution space to train our neural
networks. Among these properties are the tendencies for materials to stratify into distinct layers,
the existence of faults as distinct geological formations, and the overall variability of densities and
sound speeds in subsurface material, etc. We can also make use of dynamical properties of the
subsurface phenomena that we are trying to measure. Consider CO2 sequestration as one example.
After injection of super-critical CO2 into the subsurface, the CO2 plume will migrate over time,
meaning that the spatial spreading of CO2 should gradually increase over time. If a model can be
designed so that the images are constrained to exhibit dynamic spreading (and not, say, compaction)
of the CO2 plume, then this would improve the overall performance [20].

We provide a taxonomy of physics-informed learning-based seismic inversion techniques (see Fig. 3)

with two major categories of approaches: one being built on ML models (blue hexagons) and the other
built on physics models (green hexagons). There are three groups of techniques within each category.
We will discuss all six groups in the following sections.

B. Machine Learning Models Incorporating Physics

In this section, we describe methods identified with blue hexagons in Fig. 3. These models are all
fundamentally data-driven (in most cases, they are deep neural networks), but in each case with some
aspect of the appropriate physics incorporated. In particular, we will discuss how to embed physics
 10

knowledge through data augmentation techniques, how to constrain the learning model using the governing
wave equation, and how to learn underlying solution space distribution to help with the inversion.
1) Simulations and Data Augmentation: As noted previously in Section II-B, we can leverage our
knowledge of the imaging physics to simulate large quantities of paired data (m, d) where m is a
plausible subsurface velocity map and d is the data that would be observed if m accurately modeled
the subsurface. What we did not say in Section II-B, however, is what we meant by “plausible.” or how
a large sample of plausible subsurface maps might be generated. In general to do this requires domain
knowledge of the subsurface rock physics, and the creation of a large idealized set of subsurface structure
maps (to be used as training sets) is commonly known as “velocity building” in the subsurface geophysics
community.
While this can be effective, it may not ultimately be scalable, because it often leads to very large
training sets – for example, more than 60,000 velocity maps were created for training in [6] – that
still may not encompass the full variability of what might be in the subsurface. One approach is to use
auxiliary knowledge of the subsurface physics to reduce this variability. Well logs3 from drilled wells
has been used to build a large-scale subsurface data set for training a deep learning model [21], and
pseudo-wells based on the well statistics (such as porosity, saturation, mineralogy, and thickness) have
also been created to train the neural network models [22]. Additionally, generative models (such as the
variational auto-encoder and the generative adversarial network) have been leveraged to synthesize more
realistic simulations. A variational auto-encoder with physics-informed regularization (shown in Fig. 4a)
was designed to synthesize realistic velocity maps [23]. Domain adaptation is another technique that was
shown to be effective to overcome the data scarcity issue. A style transfer network (see (Fig. 4b) was
developed to convert from natural images to physically meaningful subsurface velocity maps [24].
It is possible, albeit less straightforward, to augment the training set from the seismic data perspective
rather than just produce more simulations form a broader variety of velocity maps. A couple of authors
have shown the benefits of enhancing supervised training with this kind of data augmentation [25],
[26]. Specifically, Gomez et al. [25] developed adaptive data augmentation techniques based on active
learning, where physically realistic simulated seismic data are created to enrich the existing training
set. Ideas of domain adaptation have also been applied to enrich the quality of seismic data in training
set. Alkhalifah et al. [26] construct a high-quality synthetic seismic dataset with more realistic features
from the target seismic data (such as noise and artifacts). Their example applications on passive seismic
imaging demonstrate the efficacy of their proposed techniques.

3
Well logs present a detailed record of the geologic information penetrated by a borehole.
 11

(a) (b)

(c)

Fig. 4: Schematic illustration of three physics-informed data augmentation techniques to generate

physically realistic samples. (a) VAE model with physics-informed loss to synthesize velocity maps [23].
(b) Style transfer model to leverage existing natural images to generate velocity maps [24]. (c) An adaptive
active learning method to augment training set with unlabeled seismic data [25].

As a demonstration, we provide a brief overview of the technique developed in [25]. Four steps
(illustrated in Fig. 4c) are proposed including:

1) Estimate an approximate solver G(θ̂, dobs );

b obs = G(θ̂, dobs );
2) Generate approximate velocity maps from unlabeled data m
3) Create seismic data using forward model dbobs = f (m
cobs );
4) Add new pairs to the original training set.

The augmented dataset plays a key role in model accuracy because it will not only carry useful physics
information but also provides examples of velocity maps that are consistent with the target geology
feature of interest. Furthermore, the full augmentation process can be applied in an iterative fashion by
re-training the approximate solver G(θ̂, dobs ) based on the extended training set in order to generate
new approximate velocity maps m
cobs . This approach allows further refinement of the mapping between
velocity and seismic subdomains. In Fig. 5, we further provide visualization of the inverted velocity
maps using both physics-based FWI and data-driven approaches as proposed in [25] to characterize tiny
 12

0 0 0 0 0
Dept h (m )

Dept h (m )

Dept h (m )

Dept h (m )

Dept h (m )
176 176 176 176 176

352 352 352 352 352

528 528 528 528 528

705 705 705 705 705

0 501 1002 1503 2005 0 501 1002 1503 2005 0 501 1002 1503 2005 0 501 1002 1503 2005 0 501 1002 1503 2005
Dist ance (m ) Dist ance (m ) Dist ance (m ) Dist ance (m ) Dist ance (m )

(a) (b) (c) (d) (e)

Fig. 5: (a) Ground truth. Inversions and errors obtained using (b) physics-based FWI (0.069, 0.972),
data-driven model trained on (c) the large and medium subsets (0.0620, 0.990), (d) large and augmented
medium subsets (augmented once) (0.0134, 0.993), and (e) large and augmented medium subsets (aug-
mented twice) (0.0122, 0.994). Error is in the format of (MAE, SSIM).

CO2 leakage plumes at Kimberlina sequestration site [27]. We observe that the network trained over
augmented subsets obtains a more accurate estimate of the plume shape and location.
2) Physics-Informed Neural Networks: Physics-informed neural networks (PINNs) [28] are learning
methods for solving scientific inverse and forward problems given governing physics PDEs. This technique
has been widely adopted in various science and engineering domains. The PINN framework is a popular
choice in seismic imaging for incorporating the governing wave equations [29], [30], [31], [32]. In the
following, we select and discuss one these techniques, as representative of the group.
Jin et al. [30] developed an Unsupervised Physical-informed Full-Waveform Inversion (UPFWI), which
can be posed as
X
c(d) = gθ∗ (d), where θ ∗ (Φu ) = arg min
m L(f (gθ (di )), di ) , (5)
θ di ∈Φu

where Φu represents an unsupervised dataset that contains seismic data di alone. Finite differences
were used to approximate the forward modeling of PDE as a differentiable operator (from velocity map
to seismic data), and then model its inversion by CNN (from seismic data to velocity map). Hence,
the supervised data-driven inversion task in Eq. (4) is transformed into the unsupervised seismic data
reconstruction task in Eq. (5), where the PINN plays an essential role. Comparing the network architecture
of UPFWI (Fig. 6) to InversionNet (Fig. 2), the major difference is the explicit incorporation of wave
equation. It constrains the output of the CNN (the orange network in Fig. 6), when passing through an
explicit wave equation, to match the original input (i.e., seismic data). Some additional physical constraints
are imposed to further enhance the imaging results: one is to leverage a learning-based perceptual loss
function by extracting style and content features of the output seismic data via VGG16 network, and the
other is to confine the resulting dimension of the CNN output, which should be the same as that of the
 13

Fig. 6: Schematic illustration of UPFWI structure, which comprises a CNN to learn an inverse mapping
and a differentiable operator to approximate the forward modeling of PDE [30].

FlatFault CurvedFault
Supervision Method
MAE ↓ MSE ↓ SSIM ↑ MAE ↓ MSE ↓ SSIM ↑
InversionNet [6] 15.83 2156.00 0.9832 23.77 5285.38 0.9681
Supervised VelocityGAN [33] 16.15 1770.31 0.9857 25.83 5076.79 0.9699
H-PGNN+ [29] 12.91 1565.02 0.9874 24.19 5139.60 0.9685
Unsupervised UPFWI 14.60 1146.09 0.9895 23.56 3639.96 0.9756

TABLE III: Quantitative results evaluated in MAE, MSE and SSIM. UPFWI yields better inversion
accuracy comparable to that of supervised baselines.

expected velocity maps. The former helps in imaging the reflection layers of the velocity maps while the
latter is helpful to some degree to constrain the solution space.
We provide a performance comparison of UPFWI with two pure data-driven methods (i.e., Inversion-
Net [6] and VelocityGAN [33]), and a supervised physics-informed method H-PGNN [29] using two
different datasets (FlatFault and CurvedFault [34]). UPFWI performs comparably to all three baseline
methods on both FlatFault and CurvedFault. The velocity maps inverted by different methods are shown
in Figure 7, which is consistent with the quantitative analysis in Table III.
3) Other Physics-Informed Constraints: Another interesting way of incorporating physics is to estimate
the velocity maps based on a learned representation of the underlying distribution of the physical
parameters. Specifically, the generative adversarial network (GAN) is a popular network architecture
that has been used to characterize data distribution. A number of recent works are representative of this
 14

Fig. 7: Comparison of different methods on inverted velocity maps of FlatFault (top) and CurvedFault (bot-
tom). For FlatFault, UPFWI reveals more accurate details at layer boundaries and the slope of the fault
in deep region. For CurvedFault, UPFWI reconstructs the geological anomalies on the surface that best
match the ground truth.

category [35], [33], [36]. Mathematically, these ideas can be formulated as:

c(z ∗ ) = gθ∗ (z ∗ )
m

s.t. z ∗ (d) = arg min L(f (gθ∗ (z)), d) (6)

z
X
θ ∗ (Φm ) = arg min LGAN (gθ (αi ), mi ) ,
θ mi ∈Φm

where Φm is a training dataset including numerous velocity maps; and z and αi are tensors sampled
from the normal distribution. The iterative optimization is then performed on z to draw a velocity map
sampled from the adaptive distribution. It is worthwhile mentioning a major difference in learning strategy
comparing Eq. (6) with Eq. (5). The optimal weights θ ∗ (Φu ) are Eq. (5) is obtained based on the dataset
Φu for a subdomain, whereas the optimal θ ∗ (d) in Eq. (6) is optimized per sample d. After training
with sufficient seismic data, Eq. (5) yields a mapping from seismic data to velocity maps that can be
directly used for inference without any optimization. In contrast, the optimal weights in Eq. (6) are sample
dependent, requiring optimization for any new sample before inference.

C. Physics Models Enhanced by Machine Learning

Three groups of methods are included in this category, as shown in the green hexagons in Fig. 3. These
methods are built on physics-based iterative approaches, with machine learning techniques introduced to
 15

Fig. 8: Schematic illustration of the idea of reparameterization, which combines neural network and PDE
of seismic waves. The neural network is used to parameterize the velocity maps while the PDE is to
calculate the forward modeling and data misfit term (Image courtesy of [37] and used with permission.)

enhance the computational techniques. Specifically, we will discuss how to learn useful prior knowledge
using deep learning models for physics-based methods, how to leverage ML to estimate gradient, and
how to learn an effective loss function using ML for physics-based methods.
1) Prior Knowledge Learning: A large body of work has demonstrated the importance of prior knowl-
edge in improving the performance of physics-based FWI approaches. This prior knowledge includes the
existing parameter distribution [37], [38], [39], [40], use of the low-frequency component for a reasonable
initial guess [41], [42], [43], and others. Here our focus is on exploiting the known parameter distribution,
for which a number of data-driven methods have been proposed. Those techniques originate from the
general idea of deep internal learning that trains deep neural networks on a single training sample. An
example would be the deep image prior [44] in which deep learning models are utilized to learn useful
prior knowledge via re-parameterization. When applying similar ideas to seismic inversion problems, it has
already been shown that a couple of the major benefits are expected including one, the techniques mitigate
local minima issue due to the sparse representation of the velocity maps using re-parameterization; two,
the re-parameterization can be seen as an implicit regularization, which in turn helps alleviate the ill-
posedness of the inversion problem [39]. Schematically, the network structures of those methods usually
consist of two components: a neural network and a PDE, which is illustrated in Fig. 8. The neural
network is used to re-parameterize the velocity maps while the PDE is to calculate the forward modeling
procedure and obtain the data misfit. Mathematically, seismic inversion with re-parameterization can be
 16

Fig. 9: Schematic illustration of the idea of (Image courtesy of [45] and used with permission.)

formulated as
m
c(d) = gθ∗ (d) (α)
(7)
s.t. θ ∗ (d) = arg min L(f (gθ (α)), d) ,
θ

where α is a random tensor drawn from a normal distribution, i.e., α ∼ N (0, σ 2 ). Different network
architectures, gθ (a), have been explored to study their regularization implicit effect in those existing
work including CNN-domain FWI [38], DNN-FWI [39] and FCN-FWI [37]. As shown in Eq. (7), these
methods involve iterative optimization, with θ ∗ (d) in Eq. (7) being optimized for seismic measurements
d for a particular time and site, and requiring re-training in order to compute the inversion for a different
input d . Unlike those subdomain approaches, the optimal weights θ ∗ (Φu ) in Eq. (4) are obtained based
on the dataset Φu , thus one network would be applicable to multiple seismic samples once fully trained.
Here the training of the networks, gθ (a), is usually carried out in two steps. The networks are first pre-
trained on an expert initial guess of the solution. After pre-training, the network is used to reparameterize
the velocity and the network parameters and is updated iteratively via the computation of the gradient.
Comparable results to physics-based inversion methods have been reported in the aforementioned works.
There are, however, some technical challenges associated with these methods. One is the heavy reliance on
the initial guess of the solution. Eq. (7) requires an expert initial guess, which is crucial for optimization
and expensive to acquire [39]. An unsatisfactory initial guess would potentially mislead the training and
yield a completely wrong inversion. Two is the high computational cost in training, which is due to the
nature of the per sample learning strategy. The optimal network parameters, θ in Eq. (7), will not be
able to be generalized to other data without a full re-training.
 17

2) Unrolling-Based Techniques: The computational techniques solving Eq. (3) are based on iterative
methods. The main idea behind this group of methods is to unroll an iterative schemes with a small number
of iterations, replacing critical components of the scheme, such as the computation of the gradient of the
data fidelity term, with a trainable module. Mathematically, this can be represented as

mk+1 = mk + αk ∇E(mk ) ≈ Ψkθ [mk , ∇E(mk )] , k = 0, 1, . . . , N − 1 , (8)

where αk is the line search step size at each iteration; ∇E(mk ) is the gradient term; and Ψkθ is a
neural network with learnable parameters θ . With an appropriately trained network of Ψkθ , the gradient
of FWI will be obtained efficiently. However, there are two major issues in employing Eq. (8). One is
the irreversible loss of some information when calculating the gradient based on the cross-correlation
of the forward and backward propagated wavefields. A heuristic interpretation is that the original raw
seismic waveform data are mapped from high-dimensional space into a low-dimensional space where
the gradient resides in, resulting in unavoidable lost of some information of the observed data. Two is
that the accuracy of Eq. (8) depends on a good initial guess of the solution. A variety of approaches
have been developed to address these issues [45], [46], [47]. Zhang and Gao [45] developed a two-stage
learning strategy described below (illustrated in Fig. 9):
• Stage 1 – a network is designed to estimate background velocity map mback from input

mback = Γθ [msmooth , I(msmooth )] , (9)

where Γθ is an encoder-decoder neural network with learnable parameter θ (shown in the upper
panel of Fig. 9); mback is the reconstructed background velocity map; msmooth is the smoothed
background velocity map; and I(·) is the reverse time migration operator4 using prestack seismic
data5 .
• Stage 2 – the high resolution velocity map is updated iteratively from the background velocity map
as
mk+1 = Λkθ [mk , I(mback ), Istack (mback ), ∇R(mk )] , k = 0, 1, . . . , N − 1, (10)

where Λkθ is a residual neural network with learnable parameter θ (shown in the lower panel of
Fig. 9); Istack (·) is the reverse time migration operator using the post-stacked seismic data6 ; and
∇R(mk ) is the gradient of the regularization in Eq. (3).

4
Reverse time migration is a seismic processing technique that is able to estimate limited subsurface structural information,
such as the locations of layer boundaries.
5
Prestack seismic data is the original raw seismic record.
6
Post-stacked seismic data is a processed seismic record that contains seismic waveforms that have been added together
from different records to reduce noise and improve overall data quality.
 18

Fig. 10: Schematic illustration of the hybrid machine learning inversion framework (Image courtesy of
[48] and used with permission).

An advantage of the two-stage learning strategy (Eqs. (9) and (10)) over the standard unrolling method
as in Eq. (8) is that the neural network input and output are in the same space. In addition, more
physical information (i.e., reverse time migration images) are used, which not only provides more details
to reconstruct the velocity maps, but also alleviates the model reliance on a good initial guess of the
solution. As mentioned in [45], the prediction performance of the model heavily relies on the quality
and quantity of the samples and labels. Some performance improvements are reported in the literature
due to the incorporation of additional physical information, but this is more computationally expensive
because of the nature of the iterative procedure.
3) Learning Loss Functions: The loss function plays a critical role in physics-based inversion tech-
niques. An inappropriately selected loss function will make the iteration prone to non-convexity/multiple
minima problems and over-sensitivity to the initial guess. Traditional methods to tackle this challenge
are based on hand-crafted functions, which may succeed in some datasets but fail in others [2]. To
overcome this issue, ML approaches have been utilized to learn a data fidelity function [48], [49], [50].
Specifically, a learning-based data fidelity function that measures high-level features has been shown
 19

to be effective in mitigating the local minima issue. Following this, Chen and Saygin [48] develop a
multi-scale hybrid machine learning inversion method by defining a data fidelity function that measures
the latent space features of the observed and predicted data (see Fig. 10). The velocity maps are updated
iteratively by minimizing the learning-based data fidelity term. A different perspective is to use a network
to approximate either the data fidelity function or the regularization term. Sun and Alkhalifah [49] employ
a meta-learning methodology to learn a robust data fidelity function for FWI. From the same group, Zhang
and Alkhalifah [50] trained a deep neural network to first estimate the distribution of facies (the overall
characteristics of a rock unit that reflect its origin and differentiate the unit from others around it), which
will be used as regularization to constrain the iterative procedure of the FWI. We select the work in
[48] to represent this group of approaches due to its promising performance and applicability to real
data. Their idea is illustrated in Fig. 10. Rather than using the traditional loss term (i.e., ∥d − f (m)∥22
in Eq. (3)), a new loss function is proposed
n o
E(m) = min ∥Wθ (d) − Wθ (f (m))∥22 , (11)
m

where Wθ is a neural network with learnable parameter θ used to extract useful latent space features from
both real data, d and predicted data f (m). The gradient of the network can be obtained via automatic
differentiation and the adjoint-state method. The proposed learning-based loss function in Eq. (11) is found
to be less prone to entrapment in local minima compared to the traditional loss in Eq.(3). Furthermore,
the proposed inversion method can recover the subsurface velocity structure in different scales. A low
dimensional latent space contains the kinematic information of the seismic waveforms, while large-
dimensional latent space features preserve the dynamic information from the waveforms. Leveraging
that, in turn, further improves the imaging resolution significantly. Experiments with both synthetic and
real data tests have been carried out to demonstrate the performance of the proposed method [48].

D. Open Data and Benchmarks

The availability of an extensive database is vital for machine learning-based inversion techniques.
Unlike some data-rich domains (such as computer vision), this is problematic for seismic inversion in
two ways. One is that data are often proprietary and hence challenging to share, but the second and
more important problem is that acquiring correctly labeled data is extremely difficult or even impossible
as the true velocity maps are never known exactly. There have been some recent effort to address this
urgent need for more and better open data for the R&D of seismic inversion. The U.S. Department of
Energy (DOE) created the Energy Data eXchange (EDX), which is a virtual library and data laboratory
built to find, connect, curate, use and re-use data to advance fossil energy and environmental R&D [51].
 20

Supervised

se
lap
e- Model Complexity
Ti m Kimberlina3D FlatVEL

Kimberlina2D CurvedVEL
Style-transfer

Kimberlina3D
CurvedFault Style-transfer
Kimberlina2D

FlatFault CurvedFault

FlatFault
UUnn
ssuu
ppee CurvedVEL
rrvvii
sseedd FlatVEL

Fig. 11: Overview of the datasets in OpenFWI [34].

EDX supports the entire life cycle of data by offering secure, private collaborative workspaces for ongoing
research projects until they mature and become catalogued, curated, and published. Recently, Los Alamos
National Laboratory created and released OpenFWI, a collection of large-scale open-source datasets,
particularly suited for benchmarking seismic inversion algorithms [34]. The purpose of OpenFWI is to
facilitate diversified, rigorous, and reproducible research on data-driven seismic inversion. As shown in
Fig. 11, OpenFWI includes datasets of multiple scales, encompasses diverse domains/applications, and
covers various levels of model complexity. Along with the dataset, an extensive study on performance
benchmarks using both physics-based inversion and data-driven techniques are provided.

IV. C HALLENGES AND F UTURE R ESEARCH O PPORTUNITIES

A summary of the inversion techniques mentioned in this review is provided in Table IV. This collection
of papers is by no means exhaustive, but it does represent some of the major R&D efforts in the past five
years. While physics-informed data-driven techniques have already shown great potential in addressing
some of the existing issues in seismic inversion (such as high computational cost, non-unique solutions,
etc), new challenges have been encountered during the development of those techniques. Here we briefly
discuss some of these, and propose potential future research topics.

1) Generalization and Model Robustness. These two topics are critical to any machine learning prob-
lems, but especially to seismic imaging. Generalization formally refers to the issue of performance
on out-of-sample data (data that is not in the training set), but for seismic imaging, the concern
 21

Learning Strategy Supervision Physics Knowledge

Method
Data-Driven Hybrid Supervised Unsupervised Augmentation Governing Eq. Priors Unrolling Loss
InversionNet [6] ✓ ✓
Araya-Polo et al. [7] ✓ ✓
ModifiedFCN [8] ✓ ✓
Li et al. [5] ✓ ✓
Renan et al. [25] ✓ ✓ ✓ ✓
Yang et al. [23] ✓ ✓ ✓ ✓
Feng et al. [24] ✓ ✓ ✓ ✓
UPFWI [30] ✓ ✓ ✓ ✓
VelocityGAN [33] ✓ ✓ ✓
H-PGNN+ [29] ✓ ✓ ✓
NNFWI [37] ✓ ✓ ✓
He & Wang [39] ✓ ✓ ✓
Wu & McMechan [38] ✓ ✓ ✓
Mosser et al. [35] ✓ ✓ ✓
Richardson et al. [36] ✓ ✓ ✓
Zhang & Gao [45] ✓ ✓ ✓
Zhang et al. [47] ✓ ✓ ✓
Chen & Saygin [48] ✓ ✓ ✓
Zhang & Alkhalifah [50] ✓ ✓ ✓
Sun & Alkhalifah [49] ✓ ✓ ✓

TABLE IV: Summary of machine learning based seismic inversion techniques addressed in this paper.

extends to data that is drawn “out of distribution,” or from a drifted distribution. Robustness refers
to the performance degradation of a model with perturbed noisy test data. Incorporation of physics
seeks to enhance both the generalization and the robustness, though it is difficult to explicitly
demonstrate the effectiveness of the models on field data, which are usually contaminated with a
significant amount of noise and artifacts. We remark that Deng et al. [52] used theoretical arguments
to provide a rigorous derivation of worst-case error bounds in seismic imaging.
2) Scientific “Sim2Real.” Given that data are scarce in subsurface geoscience, full-physics simula-
tions provide some compensation for this lack of data. How to bridge the “knowledge gap” (the
gap between physical simulations and real applications) by utilizing machine learning and the
underlying physics is a scientifically important question that needs to be carefully addressed. A
similar challenge also exists in other scientific disciplines (such as robotics), where the notion of
Sim2Real has been proposed to transfer knowledge learned in simulation to the real data [53]. That
 22

leads to two potential solutions: one is the pure model-based simulators and the other is machine
learning methods (such as generative models). The former usually simplifies the complex physics
resulting in an unavoidable reality gap between the simulation and real data. The latter may be
used as a compensating tool to learn underlying physics directly out of data. We are beginning to
see new research and development has been carried out in both areas. However, the feasibility of
these approaches on real applications with field data needs to be validated.
3) Penalties and Constraints. For most of the methods in Table IV, the physics knowledge is embed-
ded in the loss function as soft constraints via penalties such as regularization terms. Mathematically,
these problems can be generally posed as

θ ∗ = arg min {L(gθ (m), d) + λ R(f (d), m)} , (12)

θ

where m and d are the inputs and predictions, respectively; gθ (·) represents the mapping re-
lationship fitted by the network model with learnable parameters θ ; f (·) is the mathematical
representation of the domain knowledge (such as the governing PDEs); and λ is the regularization
parameter connecting the loss term L(·) and the regularization term R(·). With soft constraint as
posed in Eq. (12), the physics knowledge can be only weakly enforced in the prediction results,
meaning that they can not ensure the predictions do not violate the underlying physics constraints.
To overcome this limitation of soft constrained approaches, different optimization techniques are
recently proposed to incorporate physics knowledge as hard constraints, which ensures a given
governing equation is strictly satisfied in a given domain [54].
4) Uncertainty Quantification. Considering issues of generalization, model robustness, and informa-
tion gap between synthetic and field data, it is important to quantitatively evaluate the model perfor-
mance on both inversion accuracy and the associated confidence. Uncertainty quantification (UQ)
has been widely used as a tool to achieve this. There has been some recent work studying data-
driven seismic inversion with UQ to provide insights into the various UQ induced by different
perturbations including measurement noise and model parameters [37].

V. C ONCLUSIONS

The problem of seismic full-waveform inversion, first formulated more than 35 years ago [55], continues
to be a major area of research in subsurface geoscience. From characterizing fossil-fuel reservoirs in
subsurface energy exploration to imaging global subduction zones in earthquake seismology, there is a
demand for algorithms that are faster, more accurate, and scalable to very large datasets. As demonstrated
by the research efforts highlighted in this review, scientific machine learning is showing great potential
 23

on all of these fronts. Serious challenges remain – direct subsurface measurements are difficult and
necessarily sparse; generalizing learned solutions to regimes outside their training is inherently fraught
with uncertainty; network robustness when handling inevitable measurement noise and artifacts – but
we believe that there are ample opportunities for physics-informed machine learning to enable exciting
progress in this science-rich and data-starved field.

ACKNOWLEDGMENT

This work was funded by the Los Alamos National Laboratory (LANL) - Laboratory Directed Research
and Development program under project numbers 20210542MFR and 20200061DR, and by the U.S.
Department of Energy (DOE) Office of Fossil Energy’s Carbon Storage Research Program via the Science-
Informed Machine Learning to Accelerate Real Time Decision Making for Carbon Storage (SMART-CS)
Initiative.

R EFERENCES

[1] Albert Tarantola. Inverse problem theory and methods for model parameter estimation. SIAM, 2005.
[2] Jean Virieux and Stephane Operto. An overview of full-waveform inversion in exploration geophysics. Geophysics,
74(6):WCC1–WCC26, 2009.
[3] Amir Adler, Mauricio Araya-Polo, and Tomaso Poggio. Deep learning for seismic inverse problems: Toward the acceleration
of geophysical analysis workflows. IEEE Signal Processing Magazine, 38:89–119, 2021.
[4] Jeroen Tromp. Seismic wavefield imaging of earth’s interior across scales. Nature Review Earth and Environment, 1:40–53,
2020.
[5] Shucai Li, Bin Liu, Yuxiao Ren, Yangkang Chen, Senlin Yang, Yunhai Wang, and Peng Jiang. Deep-Learning Inversion
of Seismic Data. IEEE Transactions on Geoscience and Remote Sensing, 58(3):2135–2149, March 2020.
[6] Yue Wu and Youzuo Lin. InversionNet: An efficient and accurate data-driven full waveform inversion. IEEE Transactions
on Computational Imaging, 6:419–433, 2019.
[7] Mauricio Araya-Polo, Joseph Jennings, Amir Adler, and Taylor Dahlke. Deep-learning tomography. The Leading Edge,
37(1):58–66, 2018.
[8] Wenlong Wang, Fangshu Yang, and Jianwei Ma. Velocity model building with a modified fully convolutional network. In
SEG, pages 2086–2090, 2018.
[9] Karianne Bergen, Paul Johnson, Maarten Hoop, and Gregory Beroza. Machine learning for data-driven discovery in solid
earth geoscience. Science, 363, 2019.
[10] Gregory Ongie, Ajil Jalal, Christopher Metzler, Richard Baraniuk, Alexandros Dimakis, and Rebecca Willett. Deep learning
techniques for inverse problems in imaging. IEEE Journal on Selected Areas in Information Theory, 1:39 – 56, 2020.
[11] Xiwei Yu and Jianwei Ma. Deep learning for geophysics: Current and future trends. Reviews of Geophysics,
59:e2021RG000742, 2021.
[12] Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar. Integrating physics-based modeling with
machine learning: A survey. arXiv preprint arXiv:2003.04919, 2020.
 24

[13] George Karniadakis, Ioannis Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine
learning. Nature Reviews Physics, 3:422–440, 2021.
[14] Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar. Integrating scientific knowledge with
machine learning for engineering and environmental systems. arXiv, 2003.04919, 2020.
[15] Youzuo Lin and Lianjie Huang. Acoustic- and elastic-waveform inversion using a modified Total-Variation regularization
scheme. Geophysical Journal International, 200:489–502, 2015.
[16] Yong Ma and Dave Hale. Wave-equation reflection traveltime inversion with dynamic warping and full-waveform inversion.
Geophysics, 78(6):R223–R233, 2013.
[17] Zhigang Zhang and Lianjie Huang. Double-difference elastic-waveform inversion with prior information for time-lapse
monitoring. Geophysics, 78(6):R259–R273, 2013.
[18] Khiem Tran, Michael McVay, Michael Faraone, and David Horhota. Sinkhole detection using 2D full seismic waveform
tomography. Geophysics, 78(5):R175–R183, 2013.
[19] Antoine Guitton, Gboyega Ayeni, and Esteban Dı́az. Constrained full-waveform inversion by model reparameterization.
Geophysics, 77(2):R117–R127, 2012.
[20] Shihang Feng, Xitong Zhang, Brendt Wohlberg, Neill P. Symons, and Youzuo Lin. Connect the dots: In situ 4-D seismic
monitoring of CO2 storage with spatio-temporal CNNs. IEEE Transactions on Geoscience and Remote Sensing, 60:1–16,
2022.
[21] Zhaoqi Gao, Chuang Li, Bing Zhang, Xiudi Jiang, Zhibin Pan, Jinghuai Gao, and Zongben Xu. Building large-scale density
model via a deep-learning-based data-driven method. Geophysics, 86:IJF – V89, 2021.
[22] Jonathan Downton, Olivia Collet, and Daniel Hampson. Theory-guided data science-based reservoir prediction of a north
sea oil field. The Leading Edge, 39:689 – 764, 2020.
[23] Yuxin Yang, Xitong Zhang, Qiang Guan, and Youzuo Lin. Making invisible visible: Data-driven seismic inversion with
spatio-temporally constrained data augmentation. IEEE Transactions on Geoscience and Remote Sensing, 2022 (Accepted).
[24] Shihang Feng, Youzuo Lin, and Brendt Wohlberg. Multiscale data-driven seismic full-waveform inversion with field data
study. IEEE Transactions on Geoscience and Remote Sensing, pages 1–14, 2021.
[25] Rojas-Gomez Renan, Jihyun Yang, Youzuo Lin, James Theiler, and Brendt Wohlberg. Physics-consistent data-driven
waveform inversion with adaptive data augmentation. IEEE Geoscience and Remote Sensing Letters, 19:1–5, 2022.
[26] Tariq Alkhalifah, Hanchen Wang, and Oleg Ovcharenko. MLReal: Bridging the gap between training on synthetic data
and real data applications in machine learning. arXiv, 2021.
[27] Preston Jordan and Jeffrey Wagoner. Characterizing construction of existing wells to a CO2 storage target: The Kimberlina
site, California. Technical report, U.S. Department of Energy - Office of Fossil Energy, March 2017.
[28] Maziar Raissi, Paris Perdikaris, and George Karniadakis. Physics-informed neural networks: A deep learning framework for
solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics,
378:686 – 707, 2019.
[29] Jian Sun, Kristopher A Innanen, and Chao Huang. Physics-guided deep learning for seismic inversion with hybrid training
and uncertainty analysis. Geophysics, 86(3):R303–R317, 2021.
[30] Peng Jin, Xitong Zhang, Yinpeng Chen, Sharon Huang, Zicheng Liu, and Youzuo Lin. Unsupervised learning of full-
waveform inversion: Connecting CNN and partial differential equation in a loop. In The Tenth International Conference
on Learning Representations (ICLR), 2022.
[31] Chao Song, Tariq Alkhalifah, and Umair Bin Waheed. Solving the frequency-domain acoustic VTI wave equation using
physics-informed neural networks. Geophysical Journal International, 225:846 – 859, 2021.
 25

[32] Majid Rasht-Behesht, Christian Huber, Khemraj Shukla, and George Karniadakis. Physics-informed neural networks
(PINNs) for wave propagation and full waveform inversions. arXiv preprint arXiv:2108.12035, 2021.
[33] Zhongping Zhang and Youzuo Lin. Data-driven seismic waveform inversion: A study on the robustness and generalization.
IEEE Transactions on Geoscience and Remote sensing, 58(10):6900–6913, 2020.
[34] Chengyuan Deng, Yinan Feng, Shihang Feng, Peng Jin, Xitong Zhang, Qili Zeng, and Youzuo Lin. OpenFWI: Benchmark
seismic datasets for machine learning-based full waveform inversion. arXiv preprint arXiv:2111.02926, 2021.
[35] Lukas Mosser, Olivier Dubrule, and Martin J Blunt. Stochastic seismic waveform inversion using generative adversarial
networks as a geological prior. Mathematical Geosciences, 52(1):53–79, 2020.
[36] Alan Richardson. Generative adversarial networks for model order reduction in seismic full-waveform inversion. arXiv
preprint arXiv:1806.00828, 2018.
[37] Weiqiang Zhu, Kailai Xu, Eric Darve, Biondo Biondi, and Gregory Beroza. Integrating deep neural networks with full-
waveform inversion: Reparametrization, regularization, and uncertainty quantification. Geophysics, 87:1JF – WA102, 2021.
[38] Yulang Wu and George A McMechan. Parametric convolutional neural network-domain full-waveform inversion.
Geophysics, 84(6):R881–R896, 2019.
[39] Qinglong He and Yanfei Wang. Reparameterized full-waveform inversion using deep neural networks. Geophysics,
86(1):V1–V13, 2021.
[40] Yuxiao Ren, Xinji Xu, Senlin Yang, Lichao Nie, and Yangkang Chen. A physics-based neural-network way to perform
seismic full waveform inversion. IEEE Access, 8:112266 – 112277, 2020.
[41] Wenyi Hu, Yuchen Jin, Xuqing Wu, and Jiefu Chen. Progressive transfer learning for low-frequency data prediction in
full-waveform inversion. Geophysics, 86:1JA – X4, 2021.
[42] Yuchen Jin, Wenyi Hu, Shirui Wang, Yuan Zi, Xuqing Wu, and Jiefu Chen. Efficient progressive transfer learning for
full-waveform inversion with extrapolated low-frequency reflection seismic data. IEEE Transactions on Geoscience and
Remote Sensing, 80, 2021.
[43] Hongyu Sun and Laurent Demanet. Extrapolated full-waveform inversion with deep learning. Geophysics, 85:1MJ – Z13,
2020.
[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. arXiv preprint arXiv:1711.10925, 2020.
[45] Wei Zhang and Jinghuai Gao. Deep-learning full-waveform inversion using seismic migration images. IEEE Transactions
on Geoscience and Remote Sensing, 60, 2021.
[46] Wei Zhang, Jinghuai Gao, Xiudi Jiang, and Wenbo Sun. Consistent least-squares reverse time migration using convolutional
neural networks. IEEE Transactions on Geoscience and Remote Sensing, 60, 2021.
[47] Wei Zhang, Jinghuai Gao, Zhaoqi Gao, and Hongling Chen. Adjoint-driven deep-learning seismic full-waveform inversion.
IEEE Transactions on Geoscience and Remote Sensing, 59:8913–8932, 2020.
[48] Yuqing Chen and Erdinc Saygin. Seismic inversion by hybrid machine learning. Journal of Geophysical Research: Solid
Earth, 126:e2020JB021589, 2021.
[49] Bingbing Sun and Tariq Alkhalifah. ML-misfit: Learning a robust misfit function for full-waveform inversion using machine
learning. arXiv preprint arXiv:2002.03163, 2020.
[50] Zhendong Zhang and Tariq Alkhalifah. Regularized elastic full-waveform inversion using deep learning. Geophysics,
84:R741 – R751, 2019.
[51] United States Department of Energy. Energy Data eXchange (EDX). https://edx.netl.doe.gov/.
[52] Chengyuan Deng and Youzuo Lin. On the robustness and generalization of deep learning driven full waveform inversion.
arXiv preprint arXiv:2111.14220, 2021.
 26

[53] Stephen James, Paul Wohlhart, Mrinal Kalakrishnan, Dmitry Kalashnikov, Alex Irpan, Julian Ibarz, Sergey Levine, Raia
Hadsell, and Konstantinos Bousmalis. Sim-to-real via sim-to-sim: Data-efficient robotic grasping viarandomized-to-
canonical adaptation networks. In IEEE International Conference on Computer Vision, 2019.
[54] Yuntian Chen, Dou Huang, Dongxiao Zhang, Junsheng Zeng, Nanzhe Wang, Haoran Zhang, and Jinyue Yan. Theory-
guided hard constraint projection (HCP): A knowledge-based data-driven scientific machine learning method. Journal of
Computational Physics, 445, 2021.
[55] Albert Tarantola. Inversion of seismic reflection data in the acoustic approximation. Geophysics, 49(8):1259–1266, 1984.