Advances in

Vol. 15, No. 3, p. 245-260, 2025

Geo-Energy Research
Invited review

Multiscale modeling for multiphase flow and reactive mass

transport in subsurface energy storage: A review
Xiaocong Lyu1,2 *, Wendong Wang3 , Denis Voskov4,5 , Piyang Liu6 , Li Chen7
1 StateKey Laboratory of Petroleum Resources and Engineering, China University of Petroleum (Beijing), Beijing, 102249, China
2 School of Petroleum Engineering, China University of Petroleum (Beijing), Beijing 102249, P. R. China
3 School of Petroleum Engineering, China University of Petroleum (East China), Qingdao 266500, P. R. China
4 Department of Geoscience and Engineering, TU Delft, Delft, 2600AA, Netherlands
5 Department of Energy Science and Engineering, Stanford University, California 94305, USA
6 School of Civil Engineering, Qingdao University of Technology, Qingdao 266500, P. R. China
7 Key Laboratory of Thermo-Fluid Science and Engineering of MOE, School of Energy and Power Engineering, Xi’an Jiaotong University,
Xi’an 710000, P. R. China

Keywords: Abstract:
Multiscale modeling Modeling of multiphase flow and reactive mass transport in porous media remains a pivotal
multiphase flow challenge in the realm of subsurface energy storage, demanding a nuanced understanding
reactive mass transfer across varying scales. This review paper presents a comprehensive overview of the latest
subsurface storage advancements in multiscale modeling techniques that address the inherent complexity of
porous media these processes. Three cutting-edge approaches are presented: hybrid multiscale simulation,
which leverages both continuum and discrete modeling frameworks to enhance model
Cited as:
fidelity; approximated physics, which simplifies complex reactions and interactions to
Lyu, X., Wang, W., Voskov, D., Liu, P.,
expedite computations without significantly sacrificing accuracy; and machine-learning-
Chen, L. Multiscale modeling for
assisted multiscale simulation, which integrates predictive analytics to refine simulation
multiphase flow and reactive mass
outputs. Each method presents distinct advantages and hurdles, collectively advancing the
transport in subsurface energy storage: A
precision and computational efficiency of subsurface modeling. Despite the substantial
review. Advances in Geo-Energy
progress, we recognize the persistent challenges, such as the need for more robust
Research, 2025, 15(3): 245-260.
coupling techniques, the balance between model complexity and computational feasibility,
https://doi.org/10.46690/ager.2025.03.07
and effectively combining machine learning with traditional physical models. Promising
directions for future work are discussed to address these challenges, aiming to push the
boundaries of current multiscale modeling capabilities.

1. Introduction move towards a low-carbon economy (Bauer et al., 2013;

Gasanzade et al., 2021). As the demand for energy grows,
Energy transition in the subsurface is crucial for achiev-
the development of efficient and scalable subsurface storage
ing sustainable development and mitigating climate change
solutions becomes paramount. These technologies not only en-
impacts. It involves the shift from traditional fossil fuel
able the effective integration of intermittent renewable energy
extraction to cleaner and renewable energy sources such as
sources into the grid but also enhance the resilience of energy
geothermal energy, carbon capture and storage, and subsurface
systems against environmental and market fluctuations (Krevor
energy storage. This transition helps reduce greenhouse gas
et al., 2023). By leveraging the existing geological formations
emissions, promotes energy security, and supports the global

∗ Corresponding author.
E-mail address: [email protected] (X. Lyu); [email protected] (W. Wang); [email protected] (D. Voskov);
[email protected] (Y. Liu); [email protected] (L. Chen).
2207-9963 © The Author(s) 2025.
Received December 20, 2024; revised January 17, 2025; accepted February 6, 2025; available online February 11, 2025.
246 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

and infrastructure, subsurface energy storage presents a viable might not always capture the complexity of real-world systems
pathway to stabilize energy supply and foster sustainable (Zhang et al., 2020). Additionally, computational constraints
energy landscapes worldwide. limit the scale of simulations, making it challenging to directly
Subsurface energy storage in porous media presents a extrapolate findings to the macroscopic levels relevant for
complex challenge due to the heterogeneous nature of geolog- industrial applications (Ma and Ranjith, 2019).
ical formations and the intricate interactions between multiple Pore-scale simulation is another way to enhance our un-
phases of fluids and solids (Bauer et al., 2017; Kabuth et al., derstanding and optimization of gas storage within geological
2017). The variability in pore structure and mineral compo- formations (Meakin and Tartakovsky, 2009; Wang et al.,
sition across different scales—from the microscopic to the 2023b; Yang et al., 2023; Xie et al., 2024). By offering
field level—significantly impacts the behavior and efficiency detailed insights into the microscopic interactions between gas
of storage systems. This complexity is further compounded and the porous rock matrix, these simulations enable us to
by the reactive transport of chemical species, which can predict the behavior of gas injection and storage with greater
alter the physical properties of the storage medium over time precision (Blunt et al., 2013; Icardi et al., 2014; Molins,
(Tarkowski, 2017; Crotogino et al., 2018; Gasanzade et al., 2015; Xie et al., 2023). Such detailed understanding aids in
2021). Additionally, the operational dynamics of injecting maximizing storage efficiency, ensuring the long-term/periodic
and extracting fluids, such as compressed air, hydrogen, or security of the stored gas, and mitigating leakage risks.
supercritical CO2 , require precise control and understanding to The complexity and heterogeneity of geological formations
prevent issues like leakage, induced seismicity, and reservoir at the microscopic level introduce significant challenges in
degradation. These factors make the modeling and optimiza- accurately modeling the physical processes involved (Liu and
tion of subsurface energy storage systems a critical area of re- Mostaghimi, 2017; Liu et al., 2020; Wang et al., 2023a).
search, necessitating advanced multiscale modeling techniques The computational resources required for these simulations
to accurately predict and enhance system performance. are substantial, making it difficult to scale up to larger, more
Given the complexities of multiphase flow and reactive practical field scales (Soulaine and Tchelepi, 2016; Chen et al.,
mass transport in heterogeneous porous media, multiscale 2022). These limitations necessitate the development of more
modeling is indispensable in environmental engineering and advanced computational methods and upscaling techniques to
geosciences to optimize energy storage technologies (Luo bridge the gap between pore-scale phenomena and field-scale
et al., 2009; Golparvar et al., 2018; De Santis et al., 2021; applications.
Ramesh Kumar et al., 2021). This approach is crucial as it The application of Darcy-scale simulation in energy stor-
spans a wide range of spatial and temporal scales—from the age represents a critical advancement in the field of zero-
microscopic scale of individual pore spaces to the macroscopic carbon process (Lyu and Voskov, 2023; Awag et al., 2024).
scale of entire geological formations (Fig. 1). Multiscale mod- By employing models that operate at the Darcy scale, one
eling captures key physical and chemical processes occurring can better understand the flow and behavior of gas within
at different scales, providing insights into the behavior of subsurface geological formations, thereby enhancing the ef-
energy storage systems (Joekar-Niasar et al., 2012; Sheng ficacy and safety of storage operations. These simulations are
and Thompson, 2013). It integrates models from molecu- pivotal for predicting the migration patterns of injected gas,
lar dynamics and pore-scale models to continuum-scale and assessing storage capacity, and evaluating potential leakage
field-scale models, allowing for a detailed representation of risks, which are essential for the long-term success of energy-
physical and chemical interactions. These interactions among storage projects aimed at mitigating climate change (Pau et al.,
the injected gases, resident fluids, and the porous matrix 2010; Martinez and Hesse, 2016; Vialle et al., 2016; Lyu et al.,
can lead to complex reactive transport processes, such as 2021b). However, the simplifications required for these models
dissolution, precipitation, and chemical reactions that alter the may overlook finer-scale geological heterogeneities and fluid
porosity and permeability of the medium (Li et al., 2020; interactions, potentially leading to inaccuracies in predictions
Subramaniam, 2020; Heinemann et al., 2021). By accurately (Ajayi et al., 2019). Furthermore, the computational demands
simulating these processes, multiscale modeling contributes of these simulations can be substantial, limiting their resolu-
significantly to optimizing gas injection strategies, assessing tion and the detail they can provide (Pau et al., 2010).
storage capacities, predicting the long-term fate of stored The development of multiscale modeling in subsurface
gases, and evaluating potential environmental impacts, thereby energy storage and multiphase flow has evolved significantly
ensuring the efficiency, safety, and sustainability of energy over the past few decades, driven by the need to capture
storage solutions (Tryggvason et al., 2013; Carrillo et al., complex fluid behaviors across different spatial and temporal
2020). scales. Early approaches primarily relied on continuum-scale
Molecular simulation has emerged as a critical tool in models, such as extensions of Darcy’s law and capillary
advancing our understanding of gas storage mechanisms at pressure formulations, to describe multiphase flow in porous
the molecular level, offering insights into the interactions, media. As computational resources advanced, pore-scale mod-
behaviors, and properties of gas when introduced into various eling techniques, including direct numerical simulations and
storage mediums (Liu and Wilcox, 2012, 2013; Ma and Ran- Lattice Boltzmann Methods (LBMs), provided more detailed
jith, 2019). Despite its significant contributions, the accuracy insights into fluid displacement and reactive transport pro-
of molecular simulations heavily depends on the quality of cesses (Blunt et al., 2002; Bijeljic and Blunt, 2007; Balhoff and
the mathematical models and the parameters used, which Wheeler, 2009; Blunt et al., 2013). The emergence of hybrid
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 247

Injection well

>700 m to
surface
Gas

100 m
Caprock: Dissolution: Residual trapping: Injection: Gas-brine-rock: Structural geology:
⚫ Diffusion ⚫ P/T change ⚫ Pore structure ⚫ P/T change ⚫ Fluid-rock interaction ⚫ Fault leakage
⚫ Capillary leakage ⚫ Multicomponent ⚫ Multiphase processes ⚫ Multiphase processes ⚫ Rock dissolution and ⚫ Far and near field
⚫ Fracturing impurity ⚫ Relative permeability and ⚫ Stress/strain changes precipitation stress change
⚫ Buoyancy pressure ⚫ Single/multiple phases capillary hysteresis ⚫ Wellbore corrosion ⚫ Formation damage ⚫ Reactivation
⚫ salinity ⚫ Salt precipitation ⚫ Particle movement ⚫ Overpressure

Fig. 1. Gas storage in porous media highlighting all geological uncertainties (modified from Heinemann et al. (2021)).

multiscale frameworks, which integrate molecular dynamics, modeling techniques continue to be of paramount importance.
pore-scale modeling, and continuum-scale formulations, has Advances in computational methods, increased understanding
further refined predictive capabilities for subsurface systems of underlying physical and chemical processes, and improved
(Balhoff et al., 2007; Tartakovsky et al., 2008; Scheibe et al., data availability from experimental and field studies are driving
2015). progress in this area. This paper concisely outlines the latest
More recently, approximated physics-based approaches and developments, pinpoints challenges, and offers insights into
machine-learning-assisted multiscale simulations have gained the future of subsurface energy storage. Its goal is to signifi-
traction, offering computationally efficient alternatives to tra- cantly advance the adoption of multi-scale numerical coupling
ditional physics-driven methods (Zhu et al., 2022). Machine simulation technology within this field.
Learning (ML) techniques, including deep learning, Con-
volutional Neural Networks (CNNs), and Physics-Informed 2. Mathematical models in different scales
Neural Networks (PINNs), have been increasingly applied Understanding the simultaneous flow of multiple phases
to model complex subsurface processes such as multiphase through porous media in different scales is crucial in various
flow, reactive transport, and reservoir characterization (Wang fields, e.g., hydrogeology, petroleum engineering, and energy
et al., 2021b; Amini et al., 2022; Zhou et al., 2022; Marcato storage. The multiscale modeling approach integrates infor-
et al., 2023). These approaches enable rapid surrogate model- mation from macroscopic to microscopic scales, capturing the
ing, reducing the computational cost of traditional numerical heterogeneity of porous media and providing insights into fluid
simulations while preserving key physical constraints. Data behavior at both large and small scales (Fig. 2). This holistic
assimilation techniques, such as Bayesian inference and en- perspective enables us to simulate and analyze multiphase
semble learning, have also been integrated with ML models flow phenomena accurately, offering valuable insights for
to enhance uncertainty quantification and parameter estimation optimizing resource recovery and environmental management.
in subsurface energy applications (Tang et al., 2021b; Wu
et al., 2023a). Furthermore, hybrid artificial-intelligent-physics 2.1 Molecular dynamic simulation of micro- and
approaches leverage domain knowledge to ensure physically nanoscale flow
consistent predictions, bridging the gap between purely data-
Molecular dynamics simulation is a computational tech-
driven models and conventional physics-based simulations
nique used to study the behavior of molecules and atoms
(Karimpouli et al., 2022). As AI-driven methods continue
within complex hydrocarbon systems. It involves the simula-
to evolve, their role in subsurface modeling is expanding,
tion of the motion and interactions of individual particles over
offering new opportunities for optimizing energy storage, CO2
time, providing insights into the thermodynamic and kinetic
sequestration, and hydrocarbon recovery.
properties of the materials. The motions of each particle can
Due to the advantages and limitations of different-scale
be written as follows:
simulations, the development and refinement of multiscale
248 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

Molecules Core

Pore Field

Fig. 2. Schematic representations of porous media with different scales.

where u, C, and p denote the velocity vector of field, solute

Nm Nw
→
− →
− →
− → − concentration, and pressure, respectively. The subscript j and
mr̈i = fi = ∑ F ik + ∑ F ikw + F sou i (1)
k=1,k6=i k6=i,k=1 n p represent the phase index and the number of phases in
→
− the system. D is the molecular diffusion coefficient, which
where subscript i(k) denotes particle i(k), i is the unit vector
depends on the type of molecules. R is the source term to
in x-coordinate. m is the particle mass and r̈i is the position
describe the reaction in the system. The interplay between
of the particle i. fi is the total force. Nm is the total number of
the fluid and the pore structure is facilitated through the ap-
particles in the simulation system, and Nw is the total number
→
− plication of appropriate boundary conditions. Typically, pore-
of particles on the solid wall. F ik is the molecular force that
scale models exhibit resolutions in the micron range, enabling
particle k exerts on particle i due to Lennard-Jones potential.
→
− precise characterization of microscale phenomena that are
F ikw represents the molecular force between particle i and all
→
− beyond the scope of Darcy-scale models.
the solid wall particles k. F sou is the external force, which
makes the fluid deviate from the equilibrium, e.g., gravity, 2.3 Darcy-scale simulation of multiphase flow in
pressure, or electric force. When a two-body potential model porous media
is applied, the interaction force between a pair of molecules
comes from the following relation: The Darcy-scale simulation of multiphase flow in porous
media is a complex numerical modeling approach used to
→
− understand and predict the behavior of fluids as they move
F ik = −∇U(rik ) (2)
where U(rik ) is the potential energy function, and rik is the through porous geological formations. This type of simulation
magnitude of the distance between two arbitrary particles. is based on Darcy’s law, which describes the flow of a fluid
The Lennard-Jones pair-wise potential energy function can through a porous medium. In multiphase flow scenarios, differ-
be expressed as: ent fluids necessitate sophisticated mathematical models and
computational techniques to accurately represent the physical
"   6 #
σ 12 σ processes involved:
U(rik ) = 4ε − (3)
rik rik np
! np
∂
where σ is a molecular length scale, and ε is an interaction φ ∑ xc j ρ j s j + ∇ · ∑ (xc j ρ j u j + s j ρ j Jc j )
∂t j=1 j=1
strength parameter. The movement of the fluid particles can be
np
predicted by these basic physics laws, with special conditions,
+ ∑ xc j ρ j qej = 0, c = 1, 2, . . . , nc (7)
e.g., the boundary and initial conditions. j=1

2.2 Pore-scale modeling of multiphase flow in kr j

u j = −K (∇p j − ρ j g∇z), j = 1, · · · , n p (8)
porous media µj
pw = pn − pc (9)
Pore-scale simulation is one important way to understand np
the complicated interaction between fluids and pore structures ∑ sj = 1 (10)
j=1
during fluid flow and solute transport. The conventional incom-
pressible Navier-Stokes (NS) equations and the convection- Jc j = −φ Dc j ∇xc j (11)
diffusion equation are used to describe the multiphase flow where u, p and s are the Darcy velocity, pressure, and satura-
and reactive mass transport in porous media: tion, respectively. D is the effective diffusion coefficient. c and
j are the index of component and phase, respectively. These
∇ · uj = 0, j = 1, · · · , n p (4) variables denoted by a bar represent parameters at the Darcy
1 scale, obtained through measurement or upscaling techniques.
∂t uj + uj · ∇uj = − ∇p j + µ j ∇2 uj , j = 1, · · · , n p (5) φ and k are formation porosity and permeability. ρ, xc j , kr ,
ρj
and µ are phase density, component mole fraction in a phase,
∂t C + uj · ∇C = D∇2C + R, j = 1, · · · , n p (6)
relative permeability, and kinematic viscosity, respectively. g
and z are gravitational acceleration and vertical depth. Jc j is
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 249

Digital fluids Digital rocks Digital field

(Molecular model) (Geometry)

Hydrodynamic
Reservoir
properties
simulation
Thermodynamic
Time
properties
Lattice
Lab. How do we get the
Boltzmann properties for such
experiments
Method complex materials?
Computational chemistry
Molecular Formation water (Brine)
Quantum dynamics Oil, Gas, CO2, Kerogen,
mechanics Mineral Surface Chemistry,
Additives, e.g., Surfactants
Space
Fig. 3. Schematic diagram of multiscale simulations.

the diffusion-dispersion tensor of component c in phase j. simulation provides detailed insights into the microscopic
interactions and behaviors of individual particles or molecules
3. Bridging-scale techniques (Fig. 4). This microscopic information is essential for ac-
3.1 Hybrid multiscale simulation curately capturing phenomena at the nanoscale, where the
specific interactions between molecules significantly influence
The transition across scales in porous media modeling re- the overall behavior of the system. On the other hand, the
mains a fundamental challenge, particularly in linking molec- LBM, a mesoscopic approach, excels at simulating fluid flows
ular dynamics with pore-scale representations and further at larger scales by treating the fluid as a collection of discrete
extending these descriptions to continuum-scale models. Hy- particles and tracking their distribution functions over a lattice
brid multiscale simulation, which amalgamates continuum and mesh. By coupling these two methodologies, the multiscale
atomistic perspectives, offers a powerful tool for understanding simulation leverages the strengths of both: the detailed repre-
and predicting the behavior of complex systems across differ- sentation of particle interactions in molecular simulation for
ent disciplines (Fig. 3). The efficiency of such a hybrid model accurate boundary and initial condition definitions, and the
is significantly influenced by the comparative dimensions of efficient handling of fluid flow at larger scales in LBM. This
the molecular dynamics and continuum regions, alongside the synergistic integration allows for the simulation of complex
magnitude of the interface through which data exchange occurs systems with high fidelity, spanning from the molecular up
between the two subdomains. to the macroscopic level, thus bridging the gap between
Regarding the advanced simulation approaches for molec- microscopic mechanisms and their macroscopic manifestations
ular simulation, AB Initio Molecular Dynamics (AIMD) and in fluid dynamics and material science. Similarly, the algebraic
quantum mechanics/molecular mechanics methods stand out multiscale solver, initially used for continuum scale, leverages
due to their ability to provide a detailed picture of molecular an algebraic approach that requires only the fine-scale system
systems at the atomic level. AIMD combines quantum me- and a superimposed multiscale coarse grid, allowing it to
chanics principles with classical molecular dynamics, allowing be applied to arbitrary geological models without additional
for the simulation of molecular systems with electrons explic- preprocessing (Hajibeygi et al., 2011; Hajibeygi and Tchelepi,
itly included. This method is particularly useful for studying 2014; Wang et al., 2014). A key innovation is its hybrid
chemical reactions, material properties at the atomic scale, and two-stage strategy: a global solver operates exclusively at the
the behavior of molecules in various environments. It provides coarse scale, reducing computational costs, while various fine-
insights that are unattainable through classical molecular- scale local preconditioners, including correction functions,
dynamics simulations alone. AIMD and quantum mechan- block incomplete lower–upper factorization, and ILU, enhance
ics/molecular mechanics methods embody the forefront of convergence and numerical stability. Notably, algebraic multi-
computational chemistry and materials science, enabling the scale solver demonstrates superior efficiency and flexibility,
exploration of complex molecular phenomena with remarkable particularly in handling complex geological heterogeneity,
detail and accuracy. making it a significant advancement in multiscale reservoir
Multiscale simulation, integrating molecular simulation simulation techniques.
with the LBM or Computational Fluid Dynamics (CFD) Combining pore-scale modeling with reservoir simulation
methods, represents a cutting-edge approach in computational allows for more comprehensive and accurate descriptions of
physics and engineering to study complex fluid dynamics fluid flow and transport phenomena. By integrating these two
and transport phenomena across different scales (Smith and scales, one can leverage the detailed insights gained from
Theodorakis, 2024). In this hybrid framework, molecular pore-scale models to inform and enhance the parameterization
works is that they do not embed well in 3D Euclidean space – i.e. phase, or individual pores containing porous fillers such as
one invariably needs to have some throats connecting non-adja- clay). In this work, we use f to denote the fraction of macrop-
250Real porous media might
cent pores. Lyu, X., pores
have et al. on
Advances in scales,
different Geo-Energy Research, ores 2025, 15(3):
or grains 245-260
that contain the micropores.
but the pores nevertheless span a finite size range and are embed- (3) Choose a scaling factor b for microporous regions (ratio of the
ded in a Euclidean space. macro to micro length scale, noting that different regions
Many packing algorithms (such as sphere-packing algorithm can have different scaling factors). Although we used integer
used in this work, [35]) are capable of creating heterogeneous grain scaling factors in this work, the choice of b is arbitrary and
packings with a wide distribution of grain sizes. Pore-network non-integer numbers can be used as well.
model construction for these packings is typically done via Dela- (4) Rescale and map a microporosity network onto each desig-
unay tessellation (DT) of the grain centers. DT, however, struggles nated microporous region; retain the pores and throats of
with creating meaningful pores e.g. where very numerous small the mapped network that fall within the microporous region
grains touch a large grain and results in many distorted pores. and connect them with all the existing pores and throats
We have exemplified this situation elsewhere [36]. Voro++ library (macro or micro). Repeat the procedure until all of the
(http://math.lbl.gov/voro++/) can create weighted Voronoi tessel- microporous regions are processed.
lation (dual of DT) for packings with wide range of sizes and it
could be extended to create pore throat networks (private commu- Steps (1)–(4) result in a multiscale network that contains both
nication with Dr. C. Rycroft). Alas, to our knowledge no successful macroporosity and microporosity.
attempt exists to date. Before we proceed with a more detailed explanation of the algo-
rithm, we comment on the origin of the micronetwork. This work
Fig. 4.pore
2.3. Two-scale A domain
networkdecomposition
construction coupling arrangement, which includes: (1) an averaged
uses networks region at
representative of the bottommedia
granular boundary
on to
both scales.
determine the CFD boundary, and (2) a constrained region at theThis top isboundary
simply aofmatter
the domain where a termination
of convenience, however, andforcewe
is do not in-
applied to prevent molecules from escaping. Additionally,
We will refer to the pore-throat network that corresponds to a buffer zone is incorporated to ensure that the constrained molecular
tend to imply that microporosity is granular in nature: most are
region doesporosity
the inter-granular not directly
as interact with the and
macronetwork, averaged
any region
networkused fornot
boundary
granular determination
at all. The (Smith and Theodorakis,
micronetwork that is to2024).
be rescaled onto

Fig. 5.ofSchematic
Fig. 2. Schematic of two-scale
two-scale pore pore network
network construction for (a)construction for(b)
grain-filling, and (a)pore-filling
grain-filling and (b) pore-filling
microporosity. Intergranularmicroporosity (Mehmani
(‘‘macro’’) network pores areand
shown as larger
blue circles.Prodanović,
For designated2014).
microporous regions, a network similar to what is observed in available images/data is rescaled with an appropriate length ratio and mapped into
the microporous region, in (a) grains are microporous, in (b) pores are microporous. Throats connecting the two networks across the known boundary of the two are identified
in order to get a single network that contains both length scales and are shown in red. The key feature is that the connections between two length scales can be non-planar
(and arbitrarily complex). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
of Darcy-scale models. For example, pore-scale simulations enhance the predictive capabilities for two-phase flow systems
can be used to calculate effective permeability and relative (Raeini et al., 2012; Mehmani and Prodanović, 2014).
permeability curves, which are then upscaled and incorpo- Coupling reservoir simulation and molecular simulation
rated into Darcy-scale simulations to improve their accuracy on massively parallel High-Performance-Computing (HPC)
in predicting macroscopic flow behavior. One notable study systems represents a sophisticated computational approach
that highlights the potential of this multi-scale approach is designed to bridge the gap between macroscopic reservoir
the work by Blunt et al. (2013), which utilized pore-scale behaviors and microscopic fluid interactions (Bao et al., 2016).
imaging and modeling to derive petrophysical properties of This approach leverages the immense computational power
carbonate rocks. Their findings demonstrated that pore-scale of HPC systems to simultaneously run large-scale reservoir
simulations could provide detailed characterizations of pore simulations, which predict fluid flow and phase behavior on
connectivity and distribution, which are critical for accurate the scale of kilometers, and molecular simulation simulations,
Darcy-scale modeling. Another significant contribution is the which model the physical and chemical interactions between
research by Bijeljic et al. (2013), where they employed multi- molecules at the nanometer scale (Fig. 6). By integrating these
scale imaging techniques to capture the pore structure and two levels of simulation, a more comprehensive understanding
subsequently performed flow simulations to understand the of subsurface fluid dynamics can be achieved, encompassing
impact of pore-scale heterogeneities on macroscopic transport both the large-scale effects of reservoir geometry and hetero-
properties. The integration of pore-scale and Darcy-scale mod- geneity, as well as the microscopic influences of fluid viscosity
els has also been effectively applied in studying multiphase and wettability. The parallel nature of HPC systems allows
flow in porous media (Fig. 5). For instance, a method is for the distribution of computational tasks across thousands
developed to compute relative permeability from pore-scale of processors, significantly reducing the time required for
simulations and demonstrated that incorporating these detailed simulations and enabling the handling of complex models
pore-scale data into Darcy-scale models could significantly that were previously infeasible, thus providing insights into
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 251

RST

Lookup table Data available?

Coupler
Preprocessing for MD
simulations

MDS Scheduler

Fig. 6. Flow chart of the computational framework for molecular simulation and reservoir-simulation toolbox (modified from
Bao et al. (2016)).

optimizing carbon sequestration strategies with unprecedented stantaneously compared to the timescale of transport processes
detail and accuracy. (Steefel et al., 2015). This significantly reduces the computa-
tional burden as it avoids the need to solve complex differential
3.2 Approximated physics and advanced equations governing reaction kinetics. These techniques have
nonlinear solvers proven effective in various applications, from groundwater
Due to the highly nonlinear properties across different contamination remediation to enhanced oil recovery.
scales of time and space, one useful technique is to simplify The hybrid analytical-numerical method, developed by
the complex, underlying physical laws governing small-scale Flyer and Fokas (2008), offers another possibility to solve
interactions to make large-scale simulations computationally complex problems in multiphase flow and reactive mass trans-
feasible without significantly sacrificing accuracy. Upscaling port. This technique is particularly valuable in scenarios where
techniques, aiming to derive macroscopic properties from exact solutions are available for specific parts of a problem
microscopic behavior, help in bridging the gap between pore- domain, or where boundary conditions and singularities pose
scale phenomena and field-scale applications, ensuring that the challenges that purely numerical methods might struggle to
detailed interactions at the pore level are adequately repre- handle efficiently. By integrating analytical solutions for parts
sented in larger-scale models. For instance, homogenization of the domain where they are applicable, this approach can
techniques and volume averaging are commonly employed significantly enhance the accuracy and efficiency of the simu-
to obtain effective parameters, e.g., porosity, permeability, lation. It provides a way to impose rigorous solutions to sub-
and reactive surface areas that can be used in Darcy-scale sets of a problem, ensuring that the overall numerical solution
models (Zhang et al., 2021). Another improvement in reactive adheres more closely to the physical reality. Furthermore, this
mass transport is the development of a nonlinear element method can reduce computational costs by simplifying aspects
balance formulation that seamlessly integrates thermodynamic of the model where high precision is crucial but difficult to
and chemical equilibria within a reactive-compositional flow achieve numerically (Meunier et al., 2017; De Barros et al.,
model (Kala and Voskov, 2020; Ahusborde et al., 2024; 2019).
de Hoop et al., 2024). Unlike traditional molar-based for- Though these approximations can predict the phase dy-
mulations, which often struggle with numerical instability namics in porous media, the underlying physics is expressed
and inefficiencies in handling complex phase behavior, this either by empirical correlations or simple models to reduce the
approach leverages the consistent reduction of component nonlinearity, which may cause discrepancies in the predictions.
conservation equations to element conservation equations, en- One accurate and efficient approach, named Operator-Based
suring a more robust and computationally efficient framework. Linearization (OBL), is proven to solve the coupled problems
A key innovation is the incorporation of chemical equilibrium efficiently and accurately (Voskov, 2017; Lyu et al., 2021a).
constraints directly into the multiphase multicomponent neg- Conventionally, the simulations require assessing physical
ative flash calculations, enabling the simultaneous solution of properties and their derivatives in relation to nonlinear vari-
phase and chemical equilibria. By introducing the Equilibrium ables, which is a process known for its complexity and time
Rate Annihilation matrix and modifying the multiphase flash consumption. However, the OBL approach introduces a novel
equations to work with element mole fractions, the method tactic for simplifying the calculation of nonlinear physics by
provides a generalized treatment of precipitation and dissolu- parameterizing the state-dependent properties in the physical
tion reactions in multiphase flow systems. In addition, some space either during the preprocessing process or adaptively
reactive transport simulators couple solute transport equations with a few supporting points (Khait and Voskov, 2018; Lyu
with chemical kinetics. These simulators often implement and Voskov, 2023). In the course of a simulation, the physical
approximations such as local equilibrium assumptions, where properties at the current timestep are assessed using a multi-
it is presumed that chemical reactions reach equilibrium in- linear interpolation method (Fig. 7). This approach enhances
252 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

𝝌𝒄𝒋 𝝎𝟏 𝝌𝒄𝒋 𝝎𝟐
(a)
(b)

𝝌𝒄𝒋 𝝎

𝝌𝒄𝒋 𝝎𝟑 𝝌𝒄𝒋 𝝎𝟒

Fig. 7. Schematic description of operator parameterization and interpolation process for χ operators with a predefined OBL
resolution (modified from Voskov (2017); Lyu et al. (2021a)). ω 1 to ω 4 are four supporting points, and ω is the current
thermodynamic state of a given control volume at a given timestep in the simulation. (a) Operator distribution and (b)
interpolation.

(a) (b)
＋
l II II H

l : J-1
| 」
Fig. 8. Examples of data structures used by AMR codes. (a) Block-structured and (b) point-structured (modified from Cant
et al. (2022)).

the efficiency of the linearization process (Voskov, 2017). By 2018; de Hoop et al., 2021).
simplifying the assembly of the Jacobian matrix, it utilizes Domain decomposition methods partition the computa-
partial derivatives of the physical properties with respect to tional domain into smaller subdomains, solving the nonlinear
nonlinear variables, which are then directly calculated as co- equations within each subdomain either sequentially or in
efficients in the multi-linear interpolation process. One can find parallel. These methods, including the Schwarz alternating
more details about the applications of the OBL approach for method and the balancing domain decomposition by con-
multi-phase multi-component systems with gravity (Khait and straints, facilitate scalable and efficient solutions for large-
Voskov, 2018), capillarity (Lyu et al., 2021a), thermal (Wang scale multiscale simulations. By addressing nonlinearity within
et al., 2020), and diffusion effects (Lyu et al., 2021c). This localized regions, domain decomposition methods can handle
approach can bridge the gap between detailed microscopic de- complex boundary conditions and heterogeneous properties
scriptions and emergent macroscopic behaviors. Approximated more effectively. Another efficient technique, named Operator
physics enables the exploration of a vast range of applications, splitting method, decomposes the original problem into sim-
providing insights into the behavior of complex systems across pler sub-problems that can be solved sequentially. For instance,
multiple levels of organization. the adomian decomposition method splits nonlinear Partial
Nonlinear solvers address the inherent nonlinearity in the Differential Equations (PDEs) into linear and nonlinear parts,
governing equations of fluid flow and transport in porous solving them iteratively. These methods are particularly useful
media. These equations, such as the Navier-Stokes equations for handling stiff reactions and transport processes in porous
or the Richards equation, exhibit strong nonlinearities due media. Nonlinear solvers applied within each sub-problem
to coupling between pressure, saturation, and permeability. ensure that the overall solution remains accurate and stable
Advanced nonlinear solvers are designed to handle these (Estep et al., 2008; Lu et al., 2022; Vasilyeva, 2023).
complexities effectively when the classical Newton-Raphson
approach suffers from convergence issues. Adaptive Mesh 3.3 High-performance computing and
Refinement (AMR) techniques dynamically adjust the com- machine-learning-assisted multiscale simulation
putational grid to resolve fine-scale features where they are HPC has become an indispensable tool in modern scientific
most needed, such as near sharp fronts or interfaces in porous and engineering research, enabling the solution of complex
media (Fig. 8). By refining the mesh adaptively, AMR methods and computationally intensive problems. HPC involves the
can achieve high resolution in critical regions while maintain- use of supercomputers and parallel processing techniques to
ing coarser grids elsewhere, thus optimizing computational perform computations at speeds far beyond that of standard
resources. Nonlinear solvers combined with AMR, such as computing systems. The ability to process vast amounts of
the nonlinear multigrid-AMR methods, have demonstrated data and perform numerous calculations simultaneously makes
substantial improvements in both accuracy and efficiency for HPC crucial in multiscale simulations (Gauthier et al., 2021;
multiscale simulations (Berger and Oliger, 1984; Cusini et al.,
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 253

PDE (ν)
NN (x, t; θ) ∂
σ σ ∂t

x σ σ
∂ ∂u ∂u ∂ 2u
u +u –ν
∂x ∂t ∂x ∂x2
t σ σ

σ σ ∂2
∂x2

<ε ? Loss

Done Y

Fig. 9. The physics-informed neural network algorithm (Karniadakis et al., 2021).

Ghassemzadeh et al., 2021). One of the primary components leading to breakthroughs for multiscale phenomena (Zhuang
of HPC is the architecture of supercomputers, which includes et al., 2024b).
thousands of interconnected processors working in parallel. Training deep neural networks often hinges on the avail-
These processors can perform quadrillions of calculations per ability of large datasets, which may not always be accessible
second. For instance, the Graphics Processing Units (GPUs) for specific scientific inquiries. As an alternative, these net-
offer substantial improvements in computational speed and ef- works can be effectively trained by leveraging supplementary
ficiency. GPUs are designed to handle parallel tasks efficiently, information derived from the application of physical laws.
making them particularly well-suited for the large-scale sim- This innovative technique is called the PINNs, which integrate
ulations and matrix operations inherent in reservoir modeling. traditional physical models with advanced ML algorithms
Studies have shown that GPUs can accelerate computational (Wang and Lin, 2020; Karniadakis et al., 2021). PINNs ensure
tasks by an order of magnitude compared to CPUs, which that simulations are not only data-driven but also adhere to fun-
translates into more timely and cost-effective decision-making damental physical laws, enabling more accurate and efficient
processes in reservoir management (Esler et al., 2022). predictions of complex phenomena (Fig. 9). By capturing the
Even with significant advancements in simulating mul- intricate behaviors and interactions within porous materials at
tiphysics challenges through the numerical breakdown of multiple scales, this approach significantly enhances our ability
PDEs, integrating noisy data into current algorithms still to design and optimize energy storage processes. Recently,
poses difficulties. In addition, it is intricate to address high- Physics-Informed Kolmogorov-Arnold Networks (PIKANs)
dimensional issues steered by parameterized PDEs. Further- have attracted attention due to their advantages over tradi-
more, addressing inverse problems that involve undiscovered tional PINNs, making them a powerful alternative for solving
physical principles is frequently exorbitantly costly. This ne- complex physics-governed problems (Jacob et al., 2024; Rigas
cessitates distinct approaches and complex codes. Machine- et al., 2024; Toscano et al., 2024). The primary strength of
learning-assisted multiscale simulation has emerged as a trans- PIKANs lies in their architecture, which is inspired by the
formative approach for analyzing multiphase flow and reactive Kolmogorov-Arnold representation theorem. Unlike standard
mass transport in porous media, e.g., CNNs (Han and Jentzen, PINNs that rely on deep fully connected networks to approx-
2017; Zha et al., 2022; Zhuang et al., 2024a). These neural imate solutions to differential equations, PIKANs decompose
networks enhance the efficiency and accuracy of simula- the solution space into a hierarchical structure, enabling a
tions that span multiple scales, from atomic to macroscopic more interpretable and efficient representation of complex
levels, by learning patterns and predicting outcomes from functions. This decomposition reduces the number of required
vast datasets. Recurrent neural networks, with their ability to parameters, leading to faster convergence and improved gen-
handle sequential data, are invaluable for simulations involving eralization, particularly in high-dimensional problems where
time-dependent processes. On the other hand, CNNs excel PINNs often struggle due to the curse of dimensionality.
in analyzing spatial data, making them ideal for simulations Traditional PINNs operate as black-box models, where the
related to image recognition, material science, and fluid dy- learned functions lack transparency and do not explicitly
namics, where the spatial distribution of elements is crucial. reveal the underlying structure of the solution (Toscano et al.,
By integrating these ML models into multiscale simulations, 2024). PIKANs, however, leverage the Kolmogorov-Arnold
one can tackle complex, real-world problems more effectively, structure to provide a more modular and hierarchical break-
254 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

Table 1. Comparison of different approaches for multiscale modeling.

Approach Advantages Limitations Potential Applications

Hybrid - Multiphase flow in porous media

- Captures physics across multiple scales - High computational cost
multiscale - CO2 sequestration
Simulation - Accurate representation of pore- and continuum-scale processes - Requires sophisticated coupling strategies
- Hydrocarbon recovery modeling
- Reactive transport modeling
Approximated - Reduces computational complexity - Potential loss of physical fidelity
physics - Enhanced oil recovery simulations
- Provides stable solutions for complex nonlinear systems - May require calibration with full-physics models
- Large-scale reservoir modeling
- Enables large-scale, high-resolution simulations - Requires significant computational resources - High-resolution energy storage simulations
HPC
- Supports real-time data processing - Can be expensive and energy-intensive - Large-scale multiphase flow modeling
- Data-driven reservoir modeling
ML-assisted - Accelerates computation using surrogate models - Requires large training datasets
simulation - Uncertainty quantification
- Can learn complex patterns from data - Limited interpretability in some cases
- CO2 and hydrogen storage forecasting

down of the target function, allowing for better physical recognition, uncertainty quantification, and real-time forecast-
insight and easier identification of dominant patterns in the ing but require extensive training datasets and may suffer from
learned representations. This aspect is particularly beneficial limited interpretability. Ultimately, the choice of modeling ap-
in fields like porous media flow, subsurface transport, and proach depends on the specific application, balancing accuracy,
geophysical modeling, where understanding the interactions computational efficiency, and scalability for challenges such
between variables is crucial for improving predictive accu- as CO2 sequestration, hydrogen storage, and enhanced oil
racy. In addition, computational efficiency is another area recovery.
where PIKANs outperform PINNs. The hierarchical nature
of PIKANs ensures that the network learns low-dimensional 4. Challenges and perspectives
representations effectively before reconstructing the full func- Although significant advancements have been made in
tion, leading to better-conditioned optimization problems and multiscale modeling for multiphase flow and reactive transport
reducing the likelihood of vanishing or exploding gradients in porous media, several critical challenges persist. Addressing
that commonly hinder deep PINNs architectures. Furthermore, these issues is essential to improving model accuracy, com-
PIKANs demonstrate improved robustness in handling multi- putational efficiency, and practical applicability in subsurface
scale problems, a domain where PINNs often require extensive energy storage and carbon sequestration.
tuning of hyperparameters and network depth to capture small-
scale features. Since PIKANs break down complex mappings 4.1 Robust coupling techniques
into simpler sub-functions, they naturally accommodate multi- A key difficulty in coupling techniques arises from the in-
scale structures without excessive parameterization. herent scale separation between microscopic and macroscopic
Table 1 shows the comparison of different approaches processes. Traditional numerical methods, such as finite differ-
for multiscale modeling for multiphase flow and reactive ence or finite element approaches, struggle to resolve fine-scale
mass transport in porous media. Hybrid multiscale simulations details without significantly increasing computational expense
provide a rigorous framework for capturing physics across (Soulaine, 2024). On the other hand, upscaling techniques
multiple scales, accurately representing pore- and continuum- like volume averaging or homogenization simplify fine-scale
scale interactions. However, their high computational cost and processes but often fail to preserve essential microscale hetero-
the necessity for sophisticated coupling strategies make them geneities, leading to errors in macroscopic predictions (Zhang
challenging to implement for large-scale reservoir simula- et al., 2021). The complexity is further exacerbated in reactive
tions. In contrast, the approximated physics approach aims transport problems, where chemical reactions occurring at the
to reduce computational complexity by simplifying governing pore scale can have a significant impact on macroscopic flow
equations while maintaining stability and accuracy. Although behavior. If these interactions are not accurately accounted for,
these methods enhance computational efficiency, they may the model may fail to capture important phenomena such as
sacrifice physical fidelity and often require calibration with mineral precipitation, dissolution, or wettability alteration, all
full-physics models to ensure reliability. of which influence the long-term performance of subsurface
HPC has significantly expanded the capabilities of mul- energy storage systems.
tiscale simulations by enabling high-resolution, large-scale In addition, the nonlinearity is introduced by multiphase
modeling through parallel computing and GPU acceleration. interactions, phase changes, and geochemical reactions (Cai
While HPC facilitates real-time data processing and enhances et al., 2024). In highly heterogeneous reservoirs, fluid flow
the accuracy of multiphase flow and reactive transport simula- and reaction kinetics vary significantly across different rock
tions, it demands substantial computational resources, making formations, requiring adaptive modeling approaches. However,
it costly and energy-intensive. Meanwhile, machine-learning- ensuring numerical stability while capturing these complex
assisted multiscale simulations offer a promising alternative interactions remains a major hurdle, particularly when dealing
by leveraging surrogate models and data-driven approaches with large-scale simulations with millions of computational
to accelerate computations. These methods excel in pattern grid cells.
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 255

To overcome these challenges, research efforts are focused critical reaction pathways that influence long-term system
on developing more robust and efficient coupling strategies behavior (Abd and Abushaikha, 2021; Ladd and Szymczak,
that balance computational feasibility with model accuracy. 2021). Similarly, reduced-order models, while computationally
Hybrid multiscale simulation techniques offer a promising efficient, often struggle to generalize beyond the conditions
solution by integrating continuum-based approaches with dis- under which they were trained (Ardakani et al., 2024). Finding
crete models, allowing for localized fine-scale resolution while the right balance between model simplification and physical
maintaining computational efficiency. For example, hybrid realism remains a key challenge in multiscale modeling. More-
schemes can employ pore-network models to resolve mi- over, real-world applications require models to be adaptable to
croscale interactions in specific regions of interest while using different reservoir conditions, such as variations in rock prop-
continuum-based solvers for large-scale flow simulations. This erties, fluid compositions, and operating conditions. Fixed-
approach reduces computational costs while preserving critical resolution models may either be too computationally expensive
fine-scale details. The study by Korneev and Battiato (2016) for large-scale simulations or too simplistic to capture essential
on the homogenization of pore-scale reactive transport models dynamics. Developing a modeling framework that dynamically
is an example of how advanced mathematical techniques adjusts its complexity based on local conditions is crucial for
can be used to derive effective parameters for Darcy-scale improving both efficiency and accuracy.
simulations from detailed pore-scale data. In addition, the Adaptive resolution techniques, where model complexity
advancement of adaptive multiscale methods, where model varies depending on local flow conditions and reaction rates,
resolution dynamically changes based on local conditions, is one possible way to achieve this goal (Liu and Zhang,
also provides one possibility to resolve these challenges. For 2021; Yang et al., 2021). For example, fine-scale resolution
instance, fine-scale simulations can be triggered in regions can be selectively applied to regions experiencing rapid phase
experiencing rapid phase transitions or chemical reactions, transitions or strong chemical reactions, while coarser grids
while coarser grids are used in less dynamic regions. This are used in less dynamic regions. This approach significantly
adaptive refinement approach ensures that computational re- reduces computational costs without sacrificing accuracy in
sources are allocated efficiently without sacrificing accuracy in critical areas. Adaptive meshing techniques, such as dynamic
critical areas. Despite advancements in modeling frameworks, grid refinement and coarsening, offer a practical way to
bridging the gap between scales while preserving the fidelity implement this concept in large-scale simulations. Hybrid
of the underlying physics remains an ongoing research frontier modeling, another promising approach, combines high-fidelity
(Blunt et al., 2013; Lasseux et al., 2016; Soulaine, 2024). and reduced-order models in a single framework (De Santis
et al., 2021; Li and Voskov, 2021). In this strategy, detailed
4.2 Balancing model complexity and physics-based models are used in regions where high accuracy
computational feasibility is required, while simplified surrogate models approximate
High-fidelity models strive to capture detailed physics, less critical areas. Hybrid multiscale methods, such as do-
such as interfacial tension effects, pore-scale transport phe- main decomposition techniques, enable the seamless transition
nomena, geochemical reactions, and phase behavior, to en- between different levels of model complexity, ensuring com-
hance predictive accuracy. However, incorporating such in- putational efficiency without compromising essential physical
tricate details into large-scale reservoir simulations leads to details (Tang et al., 2021a).
significant computational demands, often rendering these mod- 4.3 Machine-learning integration
els impractical for real-world applications. A major challenge
arises from the need to resolve fine-scale heterogeneities in The integration of ML into multiscale modeling for subsur-
porous media, such as fractures, grain boundaries, and capil- face flow and reactive transport has emerged as a powerful tool
lary interactions, while maintaining computational efficiency to enhance computational efficiency and predictive accuracy.
at the reservoir scale (Duran et al., 2021). Traditional fully However, several fundamental challenges must be addressed
coupled numerical methods, including finite element and finite to ensure the reliability and practical application of ML in
volume approaches, become prohibitively expensive when large-scale subsurface simulations.
applied to large domains with high resolution. Additionally, One of the primary challenges is the lack of interpretability
reactive transport processes introduce strong nonlinearities due in deep learning models. Traditional physics-based models are
to coupled chemical reactions, which further increase the built on well-established equations governing multiphase flow
computational cost by requiring small time steps and fine and transport processes, providing transparency and explain-
spatial discretization (Abd and Abushaikha, 2021; Baqer and ability in their predictions (Yan et al., 2022). In contrast,
Chen, 2022). ML models, especially neural networks, function as black
Approximated physics methods offer a way to alleviate boxes, making it difficult to understand their decision-making
the computational burden by simplifying governing equations, process. This opacity raises concerns regarding their appli-
reducing the number of variables, or employing surrogate cability in critical scenarios, such as reservoir management,
models to approximate fine-scale processes. However, the carbon sequestration, or groundwater remediation, where in-
challenge lies in ensuring that these approximations do not correct predictions could lead to costly or even hazardous
introduce significant errors that undermine model reliability. consequences (Amini et al., 2022; Wu et al., 2023a; Jacob
Simplified reaction networks, for example, may fail to capture et al., 2024). Meanwhile, another major obstacle is the need
256 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

for extensive and high-quality training data. ML models re- be explored to enrich training datasets and improve the robust-
quire large datasets covering a wide range of conditions to ness of ML predictions. Moreover, uncertainty quantification
generalize well to unseen scenarios. In subsurface modeling, techniques must be incorporated into ML-assisted subsurface
generating such datasets is expensive and time-consuming, as models to assess the confidence level of predictions. Bayesian
it often involves running high-fidelity numerical simulations neural networks and ensemble learning methods can help
or collecting real-world experimental data. Moreover, many quantify uncertainties, providing insights into when ML pre-
subsurface processes involve rare or extreme events, such as dictions are reliable and when they require correction via tradi-
sudden permeability changes due to mineral precipitation or tional numerical solvers (Yin et al., 2021; Abbaszadeh Shahri
breakthrough of injected fluids, which are difficult to capture et al., 2022).
in training datasets. Data sparsity and distribution shifts be-
tween training and real-world conditions further degrade the 5. Summary
predictive performance of the model (Wang et al., 2021a; Xu Multiscale modeling techniques tailored for the compre-
et al., 2022). hensive understanding of multiphase flow and reactive mass
Additionally, ensuring generalizability across different geo- transport phenomena within porous media are crucial for sub-
logical formations, fluid properties, and operational conditions surface energy storage. Great advances have been achieved in
remains a significant challenge. ML models trained on a recent years, e.g., hybrid multiscale simulation, approximated
specific set of geological features may fail to extrapolate to physics, and machine-learning-assisted multiscale simulation,
reservoirs with different permeability distributions, fracture elucidating their respective merits and challenges in enhancing
networks, or fluid compositions. Unlike traditional numerical both the precision and computational efficiency of model-
solvers that incorporate governing equations applicable to a ing endeavors. Despite significant strides, persistent hurdles
broad range of conditions, ML-based models often struggle such as the refinement of coupling methodologies, the deli-
with scenarios that deviate from the training distribution, cate equilibrium between model intricacy and computational
leading to unreliable predictions. Hybridizing ML with tradi- tractability, and the fusion of ML algorithms with physical
tional physics-based approaches introduces new complexities. models are acknowledged. Consequently, prospective avenues
While PINNs and other hybrid models aim to incorporate for surmounting these challenges, with the ultimate goal of
fundamental physical principles into ML architectures, design- expanding the horizons of multiscale modeling capabilities,
ing such models requires balancing data-driven learning with are still to be resolved.
equation-driven constraints. Ensuring numerical stability and
convergence in these hybrid approaches is still an active area Acknowledgements
of research. We acknowledge the National Natural Science Foundation
One possible way to overcome this challenge is to develop of China (Nos. 52304054 and 52274056), National Natu-
physics-informed and hybrid machine-learning models that ral Science Foundation of China Joint Fund Project (No.
integrate prior physical knowledge into data-driven frame- U24B6003), and National Key R&D Program of China (No.
works (Zhu et al., 2022; Wu et al., 2023b). PINNs and other 2023YFB4104200).
constrained ML techniques enforce governing equations as
part of the learning process, reducing reliance on extensive Conflict of interest
training data and improving generalizability. By embedding
The authors declare no competing interest.
physics-based loss functions, these models can ensure that
predictions adhere to fundamental conservation laws, enhanc- Open Access This article is distributed under the terms and conditions of
ing their reliability in real-world applications. In this process, the Creative Commons Attribution (CC BY-NC-ND) license, which permits
we also can develop the adaptive fusion of ML models with unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
numerical solvers, where ML accelerates certain computa-
tionally expensive components (such as subgrid-scale physics References
or iterative solvers), while traditional numerical techniques
ensure robustness and accuracy (Kochkov et al., 2021). For Abbaszadeh Shahri, A., Shan, C., Larsson, S. A novel
example, ML can be used to rapidly approximate fine-scale approach to uncertainty quantification in groundwater
permeability fields or reaction rates, while finite-volume or table modeling by automated predictive deep learning.
finite-element solvers handle macroscopic flow and transport Natural Resources Research, 2022, 31(3): 1351–1373.
equations. This approach retains the interpretability of physics- Abd, A. S., Abushaikha, A. S. Reactive transport in porous
based models while benefiting from ML’s efficiency. media: A review of recent mathematical efforts in model-
Transferring learning and domain adaptation strategies to ing geochemical reactions in petroleum subsurface reser-
enhance model generalizability is also one promising way voirs. SN Applied Sciences, 2021, 3(4): 401.
to improve the performance of ML (Oyewole et al., 2022). Ahusborde, E., Amaziane, B., de Hoop, S., et al. A benchmark
By pre-training ML models on a diverse set of simulated or study on reactive two-phase flow in porous media: Part
experimental datasets and fine-tuning them for specific reser- ii-results and discussion. Computational Geosciences,
voirs, ML-assisted simulations can become more adaptable 2024, 28: 1–18.
to varying geological conditions. Synthetic data augmentation Ajayi, T., Gomes, J. S., Bera, A. A review of CO2 storage in
techniques, such as generative adversarial networks, can also geological formations emphasizing modeling, monitoring
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 257

and capacity estimation approaches. Petroleum Science, Cant, R., Ahmed, U., Fang, J., et al. An unstructured adap-
2019, 16: 1028–1063. tive mesh refinement approach for computational fluid
Amini, D., Haghighat, E., Juanes, R. Physics-informed neural dynamics of reacting flows. Journal of Computational
network solution of thermo–hydro–mechanical processes Physics, 2022, 468: 111480.
in porous media. Journal of Engineering Mechanics, Carrillo, F. J., Bourg, I. C., Soulaine, C. Multiphase flow
2022, 148(11): 04022070. modeling in multiscale porous media: An open-source
Ardakani, S. H., Zingaro, G., Komijani, M., et al. An efficient micro-continuum approach. Journal of Computational
reduced order model for nonlinear transient porous media Physics: X, 2020, 8: 100073.
flow with time-varying injection rates. Finite Elements Chen, L., He, A., Zhao, J., et al. Pore-scale modeling of
in Analysis and Design, 2024, 241: 104237. complex transport phenomena in porous media. Progress
Awag, M., Mackay, E., Ghanbari, S. A numerical analysis in Energy and Combustion Science, 2022, 88: 100968.
of background flow velocity effects on long-term post- Crotogino, F., Schneider, G.-S., Evans, D. J. Renewable
injection migration of CO2 plumes in tilted storage energy storage in geological formations. Proceedings of
aquifers. Advances in Geo-Energy Research, 2024, 11(2): the Institution of Mechanical Engineers, Part A: Journal
103–114. of Power and Energy, 2018, 232(1): 100–114.
Balhoff, M. T., Thompson, K. E., Hjortsø, M. Coupling Cusini, M., Fryer, B., van Kruijsdijk, C., et al. Algebraic
pore-scale networks to continuum-scale models of porous dynamic multilevel method for compositional flow in
media. Computers & Geosciences, 2007, 33(3): 393–410. heterogeneous porous media. Journal of Computational
Balhoff, M. T., Wheeler, M. F. A predictive pore-scale model Physics, 2018, 354: 593–612.
for non-darcy flow in porous media. SPE Journal, 2009, De Barros, F., Colbrook, M., Fokas, A. A hybrid analytical-
14(4): 579–587. numerical method for solving advection-dispersion prob-
Bao, K., Yan, M., Allen, R., et al. High-performance modeling lems on a half-line. International Journal of Heat and
of carbon dioxide sequestration by coupling reservoir Mass Transfer, 2019, 139: 482–491.
simulation and molecular dynamics. SPE Journal, 2016, de Hoop, S., Jones, E., Voskov, D. Accurate geothermal
21(03): 0853–0863. and chemical dissolution simulation using adaptive mesh
Baqer, Y., Chen, X. A review on reactive transport model and refinement on generic unstructured grids. Advances in
porosity evolution in the porous media. Environmental Water Resources, 2021, 154: 103977.
Science and Pollution Research, 2022, 29(32): 47873– de Hoop, S., Voskov, D., Ahusborde, E., et al. A benchmark
47901. study on reactive two-phase flow in porous media: Part
Bauer, S., Beyer, C., Dethlefsen, F., et al. Impacts of the i-model description. Computational Geosciences, 2024,
use of the geological subsurface for energy storage: An 28(1): 175–189.
investigation concept. Environmental Earth Sciences, De Santis, A., Colombo, M., Hanson, B. C., et al. A general-
2013, 70: 3935–3943. ized multiphase modelling approach for multiscale flows.
Bauer, S., Dahmke, A., Kolditz, O. Subsurface energy stor- Journal of Computational Physics, 2021, 436: 110321.
age: Geological storage of renewable energy—capacities, Duran, O., Devloo, P. R., Gomes, S. M., et al. A multiscale
induced effects and implications. Environmental Earth mixed finite element method applied to the simulation
Sciences, 2017, 76: 1–4. of two-phase flows. Computer Methods in Applied
Berger, M. J., Oliger, J. Adaptive mesh refinement for Mechanics and Engineering, 2021, 383: 113870.
hyperbolic partial differential equations. Journal of Com- Esler, K., Gandham, R., Patacchini, L., et al. A graphics
putational Physics, 1984, 53(3): 484–512. processing unit–based, industrial grade compositional
Bijeljic, B., Blunt, M. J. Pore-scale modeling of transverse reservoir simulator. SPE Journal, 2022, 27(01): 597–612.
dispersion in porous media. Water Resources Research, Estep, D., Ginting, V., Ropp, D., et al. An a posteriori–a priori
2007, 43(12): W12S11. analysis of multiscale operator splitting. SIAM Journal
Bijeljic, B., Raeini, A., Mostaghimi, P., et al. Predictions of on Numerical Analysis, 2008, 46(3): 1116–1146.
non-fickian solute transport in different classes of porous Flyer, N., Fokas, A. A hybrid analytical–numerical method for
media using direct simulation on pore-scale images. solving evolution partial differential equations. i. the half-
Physical Review E, 2013, 87(1): 013011. line. Proceedings of the Royal Society A: Mathematical,
Blunt, M. J., Bijeljic, B., Dong, H., et al. Pore-scale imaging Physical and Engineering Sciences, 2008, 464(2095):
and modelling. Advances in Water resources, 2013, 51: 1823–1849.
197–216. Gasanzade, F., Pfeiffer, W. T., Witte, F., et al. Subsurface
Blunt, M. J., Jackson, M. D., Piri, M., et al. Detailed physics, renewable energy storage capacity for hydrogen, methane
predictive capabilities and macroscopic consequences for and compressed air–a performance assessment study
pore-network models of multiphase flow. Advances in from the north german basin. Renewable and Sustainable
Water Resources, 2002, 25(8-12): 1069–1089. Energy Reviews, 2021, 149: 111422.
Cai, J., Qin, X., Xia, X., et al. Numerical modeling of Gauthier, D. J., Bollt, E., Griffith, A., et al. Next generation
multiphase flow in porous media considering micro-and reservoir computing. Nature Communications, 2021,
nanoscale effects: A comprehensive review. Gas Science 12(1): 1–8.
and Engineering, 2024, 131: 205441. Ghassemzadeh, S., Gonzalez Perdomo, M., Haghighi, M.,
258 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

et al. A data-driven reservoir simulation for natural gas Modeling & Simulation, 2016, 14(4): 1301–1318.
reservoirs. Neural Computing and Applications, 2021, Krevor, S., De Coninck, H., Gasda, S. E., et al. Subsurface
33(18): 11777–11798. carbon dioxide and hydrogen storage for a sustainable
Golparvar, A., Zhou, Y., Wu, K., et al. A comprehensive energy future. Nature Reviews Earth & Environment,
review of pore scale modeling methodologies for multi- 2023, 4(2): 102–118.
phase flow in porous media. Advances in Geo-Energy Ladd, A. J., Szymczak, P. Reactive flows in porous media:
Research, 2018, 2(4): 418–440. Challenges in theoretical and numerical methods. An-
Hajibeygi, H., Karvounis, D., Jenny, P. A hierarchical fracture nual Review of Chemical and Biomolecular Engineering,
model for the iterative multiscale finite volume method. 2021, 12(1): 543–571.
Journal of Computational Physics, 2011, 230(24): 8729– Lasseux, D., Parada, F. J. V., Porter, M. L. An improved
8743. macroscale model for gas slip flow in porous media.
Hajibeygi, H., Tchelepi, H. A. Compositional multiscale Journal of Fluid Mechanics, 2016, 805: 118–146.
finite-volume formulation. SPE Journal, 2014, 19(02): Li, L., Li, X., Zhu, Z., et al. Numerical modeling of multiphase
316–326. flow in gas stirred ladles: From a multiscale point of view.
Han, J., Jentzen, A. Deep learning-based numerical meth- Powder Technology, 2020, 373: 14–25.
ods for high-dimensional parabolic partial differential Li, L., Voskov, D. A novel hybrid model for multiphase
equations and backward stochastic differential equations. flow in complex multi-scale fractured systems. Journal of
Communications in Mathematics and Mtatistics, 2017, Petroleum Science and Engineering, 2021, 203: 108657.
5(4): 349–380. Liu, H., Zhu, Z., Patrick, W., et al. Pore-scale numerical
Heinemann, N., Alcalde, J., Miocic, J. M., et al. Enabling simulation of supercritical CO2 migration in porous and
large-scale hydrogen storage in porous media–the scien- fractured media saturated with water. Advances in Geo-
tific challenges. Energy & Environmental Science, 2021, Energy Research, 2020, 4(4): 419–434.
14(2): 853–864. Liu, M., Mostaghimi, P. High-resolution pore-scale simulation
Icardi, M., Boccardo, G., Marchisio, D. L., et al. Pore-scale of dissolution in porous media. Chemical Engineering
simulation of fluid flow and solute dispersion in three- Science, 2017, 161: 360–369.
dimensional porous media. Physical Review E, 2014, Liu, X., Zhang, S. Development of adaptive multi-resolution
90(1): 013032. mps method for multiphase flow simulation. Computer
Jacob, B., Howard, A. A., Stinis, P. Spikans: Separable Methods in Applied Mechanics and Engineering, 2021,
physics-informed kolmogorov-arnold networks. arXiv 387: 114184.
preprint arXiv:2411.06286, 2024. Liu, Y., Wilcox, J. Molecular simulation of CO2 adsorption in
Joekar-Niasar, V., Van Dijke, M., Hassanizadeh, S. M. micro-and mesoporous carbons with surface heterogene-
Pore-scale modeling of multiphase flow and transport: ity. International Journal of Coal Geology, 2012, 104:
Achievements and perspectives. Transport in Porous 83–95.
Media, 2012, 94(2): 461–464. Liu, Y., Wilcox, J. Molecular simulation studies of CO2 ad-
Kabuth, A., Dahmke, A., Beyer, C., et al. Energy storage sorption by carbon model compounds for carbon capture
in the geological subsurface: Dimensioning, risk analysis and sequestration applications. Environmental Science &
and spatial planning: The angus+ project. Environmental Technology, 2013, 47(1): 95–101.
Earth Sciences, 2017, 76: 1–17. Lu, R., Nagel, T., Poonoosamy, J., et al. A new operator-
Kala, K., Voskov, D. Element balance formulation in reactive splitting finite element scheme for reactive transport
compositional flow and transport with parameterization modeling in saturated porous media. Computers &
technique. Computational Geosciences, 2020, 24(2): Geosciences, 2022, 163: 105106.
609–624. Luo, K., Xia, J., Monaco, E. Multiscale modeling of
Karimpouli, S., Tahmasebi, P., Saenger, E. H. Ultrasonic multiphase flow with complex interactions. Journal of
prediction of crack density using machine learning: A Multiscale Modelling, 2009, 1(01): 125–156.
numerical investigation. Geoscience Frontiers, 2022, Lyu, X., Khait, M., Voskov, D. Operator-based linearization
13(1): 101277. approach for modeling of multiphase flow with buoyancy
Karniadakis, G. E., Kevrekidis, I. G., Lu, L., et al. Physics- and capillarity. SPE Journal, 2021, 26(4): 1858–1875.
informed machine learning. Nature Reviews Physics, Lyu, X., Voskov, D. Advanced modeling of enhanced CO2
2021, 3(6): 422–440. dissolution trapping in saline aquifers. International
Khait, M., Voskov, D. Adaptive parameterization for solving Journal of Greenhouse Gas Control, 2023, 127: 103907.
of thermal/compositional nonlinear flow and transport Lyu, X., Voskov, D., R.Rossen, W. Numerical investigations
with buoyancy. SPE Journal, 2018, 23(2): 522–534. of foam-assisted CO2 storage in saline aquifers. Inter-
Kochkov, D., Smith, J. A., Alieva, A., et al. Machine learning– national Journal of Greenhouse Gas Control, 2021, 108:
accelerated computational fluid dynamics. Proceedings 103314.
of the National Academy of Sciences, 2021, 118(21): Lyu, X., Voskov, D., R.Rossen, W. Numerical investigations
e2101784118. of foam-assisted CO2 storage in saline aquifers. Inter-
Korneev, S., Battiato, I. Sequential homogenization of reac- national Journal of Greenhouse Gas Control, 2021, 108:
tive transport in polydisperse porous media. Multiscale 103314.
 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260 259

Ma, Z., Ranjith, P. G. Review of application of molecular dy- in Porous Media, 2016, 113: 431–456.
namics simulations in geological sequestration of carbon Steefel, C. I., Yabusaki, S. B., Mayer, K. U. Reactive transport
dioxide. Fuel, 2019, 255: 115644. benchmarks for subsurface environmental simulation.
Marcato, A., Santos, J. E., Boccardo, G., et al. Prediction of Computational Geosciences, 2015, 19(3): 439.
local concentration fields in porous media with chemical Subramaniam, S. Multiphase flows: Rich physics, challenging
reaction using a multi scale convolutional neural network. theory, and big simulations. Physical Review Fluids,
Chemical Engineering Journal, 2023, 455: 140367. 2020, 5(11): 110520.
Martinez, M. J., Hesse, M. A. Two-phase convective CO2 Tang, H., Haynes, R., Houzeaux, G. A review of domain
dissolution in saline aquifers. Water Resources Research, decomposition methods for simulation of fluid flows:
2016, 52(1): 585–599. Concepts, algorithms, and applications. Archives of
Meakin, P., Tartakovsky, A. M. Modeling and simulation of Computational Methods in Engineering, 2021, 28(3):
pore-scale multiphase fluid flow and reactive transport 841–873.
in fractured and porous media. Reviews of Geophysics, Tang, M., Liu, Y., Durlofsky, L. J. Deep-learning-based
2009, 47(3): RG3002. surrogate flow modeling and geological parameterization
Mehmani, A., Prodanović, M. The effect of microporosity on for data assimilation in 3D subsurface flow. Computer
transport properties in porous media. Advances in Water Methods in Applied Mechanics and Engineering, 2021,
Resources, 2014, 63: 104–119. 376: 113636.
Meunier, F., Draye, X., Vanderborght, J., et al. A hybrid Tarkowski, R. Perspectives of using the geological subsurface
analytical-numerical method for solving water flow equa- for hydrogen storage in poland. International Journal of
tions in root hydraulic architectures. Applied Mathemat- Hydrogen Energy, 2017, 42(1): 347–355.
ical Modelling, 2017, 52: 648–663. Tartakovsky, A. M., Redden, G., Lichtner, P. C., et al. Mixing-
Molins, S. Reactive interfaces in direct numerical simulation induced precipitation: Experimental study and multiscale
of pore-scale processes. Reviews in Mineralogy and numerical analysis. Water Resources Research, 2008,
Geochemistry, 2015, 80(1): 461–481. 44(6): W06S04.
Oyewole, I., Chehade, A., Kim, Y. A controllable deep transfer Toscano, J. D., Oommen, V., Varghese, A. J., et al. From pinns
learning network with multiple domain adaptation for to pikans: Recent advances in physics-informed machine
battery state-of-charge estimation. Applied Energy, 2022, learning. arXiv preprint arXiv:2410.13228, 2024.
312: 118726. Tryggvason, G., Dabiri, S., Aboulhasanzadeh, B., et al. Mul-
Pau, G. S., Bell, J. B., Pruess, K., et al. High-resolution tiscale considerations in direct numerical simulations of
simulation and characterization of density-driven flow multiphase flows. Physics of Fluids, 2013, 25(3): 031302.
in CO2 storage in saline aquifers. Advances in Water Vasilyeva, M. Efficient decoupling schemes for multiscale
Resources, 2010, 33(4): 443–455. multicontinuum problems in fractured porous media.
Raeini, A. Q., Blunt, M. J., Bijeljic, B. Modelling two-phase Journal of Computational Physics, 2023, 487: 112134.
flow in porous media at the pore scale using the volume- Vialle, S., Druhan, J. L., Maher, K. Multi-phase flow sim-
of-fluid method. Journal of Computational Physics, 2012, ulation of CO2 leakage through a fractured caprock in
231(17): 5653–5668. response to mitigation strategies. International Journal of
Ramesh Kumar, K., Makhmutov, A., Spiers, C. J., et al. Greenhouse Gas Control, 2016, 44: 11–25.
Geomechanical simulation of energy storage in salt for- Voskov, D. Operator-based linearization approach for model-
mations. Scientific Reports, 2021, 11(1): 19640. ing of multiphase multi-component flow in porous media.
Rigas, S., Papachristou, M., Papadopoulos, T., et al. Journal of Computational Physics, 2017, 337: 275–288.
Adaptive training of grid-dependent physics-informed Wang, N., Chang, H., Zhang, D. Deep-learning-based inverse
kolmogorov-arnold networks. IEEE Access, 2024, 12: modeling approaches: A subsurface flow example. Jour-
176982–176998. nal of Geophysical Research: Solid Earth, 2021, 126(2):
Scheibe, T. D., Schuchardt, K., Agarwal, K., et al. Hybrid e2020JB020549.
multiscale simulation of a mixing-controlled reaction. Wang, W., Wen, J., Wang, C., et al. Current status and
Advances in Water Resources, 2015, 83: 228–239. development trends of CO2 storage with enhanced natural
Sheng, Q., Thompson, K. Dynamic coupling of pore-scale gas recovery (CS-EGR). Fuel, 2023, 349: 128555.
and reservoir-scale models for multiphase flow. Water Wang, W., Xie, Q., An, S., et al. Pore-scale simulation of
Resources Research, 2013, 49(9): 5973–5988. multiphase flow and reactive transport processes involved
Smith, E. R., Theodorakis, P. E. Multiscale simulation of in geologic carbon sequestration. Earth-Science Reviews,
fluids: Coupling molecular and continuum. Physical 2023, 247: 104602.
Chemistry Chemical Physics, 2024, 26(2): 724–744. Wang, Y., Hajibeygi, H., Tchelepi, H. A. Algebraic multiscale
Soulaine, C. Micro-continuum modeling: An hybrid-scale solver for flow in heterogeneous porous media. Journal
approach for solving coupled processes in porous of Computational Physics, 2014, 259: 284–303.
media. Water Resources Research, 2024, 60(2): Wang, Y., Lin, G. Efficient deep learning techniques for
e2023WR035908. multiphase flow simulation in heterogeneous porousc
Soulaine, C., Tchelepi, H. A. Micro-continuum approach for media. Journal of Computational Physics, 2020, 401:
pore-scale simulation of subsurface processes. Transport 108968.
260 Lyu, X., et al. Advances in Geo-Energy Research, 2025, 15(3): 245-260

Wang, Y., Voskov, D., Khait, M., et al. An efficient numerical for geological carbon storage. Water Resources Research,
simulator for geothermal simulation: A benchmark study. 2023, 59(10): e2023WR035402.
Applied Energy, 2020, 264: 114693. Yin, J., Medellín-Azuara, J., Escriva-Bou, A., et al. Bayesian
Wang, Y. D., Chung, T., Armstrong, R. T., et al. Ml-lbm: machine learning ensemble approach to quantify model
predicting and accelerating steady state flow simulation uncertainty in predicting groundwater storage change.
in porous media with convolutional neural networks. Science of The Total Environment, 2021, 769: 144715.
Transport in Porous Media, 2021, 138(1): 49–75. Zha, W., Zhang, W., Li, D., et al. Convolution-based model-
Wu, Y., An, S., Tahmasebi, P., et al. An end-to-end approach to solving method for three-dimensional, unsteady, partial
predict physical properties of heterogeneous porous me- differential equations. Neural Computation, 2022, 34(2):
dia: Coupling deep learning and physics-based features. 518–540.
Fuel, 2023, 352: 128753. Zhang, H., Diao, R., Mostofi, M., et al. Molecular simulation
Wu, Z., Wang, H., He, C., et al. The application of physics- of enhanced CH4 recovery and CO2 storage by CO2 –
informed machine learning in multiphysics modeling in N2 mixture injection in deformable organic micropores.
chemical engineering. Industrial & Engineering Chem- Journal of Natural Gas Science and Engineering, 2020,
istry Research, 2023, 62(44): 18178–18204. 84: 103658.
Xie, Q., Wang, W., Bakhshian, S., et al. Simulation of Zhang, X., Ma, F., Yin, S., et al. Application of upscaling
CO2 dissolution reactions in saline aquifers using lattice methods for fluid flow and mass transport in multi-scale
boltzmann method. Gas Science and Engineering, 2024, heterogeneous media: A critical review. Applied Energy,
125: 205284. 2021, 303: 117603.
Xie, Q., Wang, W., Su, Y., et al. Pore-scale study of Zhou, X.-H., McClure, J. E., Chen, C., et al. Neural network–
calcite dissolution during CO2 -saturated brine injection based pore flow field prediction in porous media using
for sequestration in carbonate aquifers. Gas Science and super resolution. Physical Review Fluids, 2022, 7(7):
Engineering, 2023, 114: 204978. 074302.
Xu, C., Fu, L., Lin, T., et al. Machine learning in petrophysics: Zhu, L.-T., Chen, X.-Z., Ouyang, B., et al. Review of machine
Advantages and limitations. Artificial Intelligence in learning for hydrodynamics, transport, and reactions in
Geosciences, 2022, 3: 157–161. multiphase flows and reactors. Industrial & Engineering
Yan, B., Harp, D. R., Chen, B., et al. A physics-constrained Chemistry Research, 2022, 61(28): 9901–9949.
deep learning model for simulating multiphase flow in 3D Zhuang, X., Wang, W., Su, Y., et al. Spatio-temporal sequence
heterogeneous porous media. Fuel, 2022, 313: 122693. prediction of CO2 flooding and sequestration potential
Yang, X., Kong, S.-C., Liu, Q. Smoothed particle hydrody- under geological and engineering uncertainties. Applied
namics with adaptive spatial resolution for multiphase Energy, 2024, 359: 122691.
flows with large density ratio. Physical Review E, 2021, Zhuang, X., Wang, W., Su, Y., et al. Multi-objective opti-
104(5): 055308. mization of reservoir development strategy with hybrid
Yang, Y., Wang, J., Wang, J., et al. Pore-scale modeling of artificial intelligence method. Expert Systems with Ap-
coupled CO2 flow and dissolution in 3D porous media plications, 2024, 241: 122707.