1. Introduction
Carbon capture and sequestration is one promising option to mitigate global warming (Orr Reference Orr2009) by reducing the anthropogenic greenhouse-gas emissions in the atmosphere. After directly capturing the carbon dioxide (CO
$_2$
) upon release from industrial plants, it is possible to re-inject it into geological formations such as deep saline aquifers and depleted oil reservoirs (Metz et al. Reference B., Davidson, de Coninck, Loos and Meyer2005). These aquifers are brine-saturated porous structures located between 800 m and 3 km beneath the Earth’s surface (Huppert & Neufeld Reference Huppert and Neufeld2014). At the pressure and temperature conditions of the aquifers, the
$\textrm {CO}_2$
becomes supercritical and positively buoyant with respect to the resident brine. As a consequence, the
$\textrm {CO}_2$
immediately rises upon injection until it encounters a cap rock, an impermeable geological layer which compels it to move sideways. This subsequently leads to a configuration where a large region of
$\textrm {CO}_2$
-free brine is surmounted by a thin layer of supercritical
$\textrm {CO}_2$
. Due to the solubility of carbon dioxide in aqueous media, a
$\textrm {CO}_2$
-enriched diffusive layer builds up at the interface, which becomes denser than the underlying pure brine. After some time, such a gravitationally unstable configuration, with a heavier fluid on top of a lighter one, destabilises. This leads to the formation of solutal plumes or fingers that pour down towards the bottom of the aquifer. During their downwards migration, the plumes continually interact and merge, thus leaving large spaces for the birth of new plumes and so on. Such a complex dynamics, where the
$\textrm {CO}_2$
-dissolution rate noticeably departs from that of the purely diffusive regime (enhancement followed by saturation), is commonly referred to as the nonlinear regime whereas the onset of the instability, where the initial diffusive nature of the dissolution rate is not affected yet, is simply called the linear regime. Note that, as soon as the fingers start feeling the influence of the bottom of the reservoir, the convective dissolution process progressively slows down and eventually stops (shut-down regime) once all of the brine is saturated by carbon dioxide. Convective dissolution has been extensively investigated in the Darcy regime, i.e. in the limit of vanishing permeability, through a comprehensive combination of theoretical, numerical and experimental studies. The characteristics (growth rate
$\sigma ^\star$
and most unstable wavelength
$\lambda ^\star$
) of the onset of the instability – i.e. in the linear regime – have been theoretically predicted by several authors Ennis-King, Preston & Paterson (Reference Ennis-King, Preston and Paterson2005), Riaz et al. (Reference Riaz, Hesse, Tchelepi and Orr2006), Hassanzadeh, Pooladi-Darvish & Keith (Reference Hassanzadeh, Pooladi-Darvish and Keith2007), Elenius & Johannsen (Reference Elenius and Johannsen2012) and then confirmed through numerical simulations (Slim Reference Slim2014; De Paoli, Zonta & Soldati Reference De Paoli, Zonta and Soldati2017) and bespoke Hele-Shaw experiments Slim et al. (Reference Slim, Bandi, Miller and Mahadevan2013), Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016). As for the nonlinear regime, it has been extensively depicted in various studies – based on numerical simulations and experimental observations Pau et al. (Reference Pau, Bell, Pruess, Almgren, Lijewski and Zhang2010), Hewitt, Neufeld & Lister (Reference Hewitt, Neufeld and Lister2013), Slim (Reference Slim2014), Vreme (Reference Vreme2015), De Paoli et al. (Reference De Paoli, Zonta and Soldati2017), De Paoli (Reference De Paoli2023) – where the features of the dynamics of the plumes (merging and births) depicted above have been consistently reported.
In reality, convective dissolution in aquifers appears to be a combination of mechanisms where geology, chemistry and fluid mechanics are deeply intertwined. Despite the large number of scientific questions and technical issues already addressed by the existing (mainly Darcy-based) models, it is crucial to include additional ingredients to reflect such complexity Hassanzadeh et al. (Reference Hassanzadeh, Pooladi-Darvish and Keith2007). In particular, the aquifer may be subject to a background flow, because of global natural currents (Emami-Meybodi Reference Emami-Meybodi2017b
) or due to the
$\textrm {CO}_2$
gravity current formed once the buoyant
$\textrm {CO}_2$
has reached the upper impermeable rock formation (MacMinn et al. Reference MacMinn, Neufeld, Hesse and Huppert2012; Hidalgo, MacMinn & Juanes Reference Hidalgo, MacMinn and Juanes2013; MacMinn & Juanes Reference MacMinn and Juanes2013; Huppert & Neufeld Reference Huppert and Neufeld2014). The presence of such a background flow in the aquifer can alter convective dissolution and this effect is twofold. First, by forcing the fluid to move, it can increase dissolution of CO
$_2$
by transporting it over large distances, thus bringing more fresh fluid in contact with it. Second, since a flow within a porous medium leads to hydrodynamic dispersion Bear (Reference Bear2018), a strong background flow may significantly increase the diffusion of CO
$_2$
in the brine with respect to molecular diffusion. Various works have shown that such additional diffusion smooths out the fingering pattern Wang et al. (Reference Wang, Nakanishi, Hyodo and Suekane2016), Emami-Meybodi (Reference Emami-Meybodi2017b
), Liang et al. (Reference Liang, Wen, Hesse and DiCarlo2018), Wen, Chang & Hesse (Reference Wen, Chang and Hesse2018), Tsinober, Shavit & Rosenzweig (Reference Tsinober, Shavit and Rosenzweig2023), which may provoke a slowdown of the instability and a loss in efficiency of the dissolution process, see Hidalgo & Carrera (Reference Hidalgo and Carrera2009), Dhar et al. (Reference Dhar, Meunier, Nadal and Méheust2022) and Tsinober et al. (Reference Tsinober, Shavit and Rosenzweig2023).
The effect of a transverse flow and hydrodynamic dispersion is generally quantified through the pore-scale-based Péclet number
$\textit{Pe}$
, which represents the ratio between the diffusion and advection time scales. Hassanzadeh, Pooladi-Darvish & Keith (Reference Hassanzadeh, Pooladi-Darvish and Keith2009) were the first to consider the effect of a background flow and hydrodynamic dispersion on the convective dissolution process. Using linear stability analysis, the authors showed that the effect of the background flow was (i) to delay the onset of the instability and (ii) to increase the value of the most unstable wavelength
$\lambda ^\star$
. Emami-Meybodi (Reference Emami-Meybodi2017a
) has confirmed these two findings, using a linear stability analysis able to decouple the respective effects of the background flow, of the longitudinal dispersion and of the anisotropy of the dispersion. However, note that in experiments or in realistic geological conditions, background flow and hydrodynamic dispersion cannot be disentangled from each other. Through experiments performed with glass beads and methanol and ethylene glycol analogue (Neufeld et al. Reference Neufeld, Hesse, Riaz, Hallworth, Tchelepi and Huppert2010), where a horizontal background velocity flow was forced through the cell, Michel-Meyer et al. (Reference Michel-Meyer, Shavit, Tsinober and Rosenzweig2020) obtained consistent results with the predictions of Hassanzadeh et al. (Reference Hassanzadeh, Pooladi-Darvish and Keith2009) and Emami-Meybodi (Reference Emami-Meybodi2017a
), but only qualitatively. Indeed, these predictions for the onset of the instability have not been tested quantitatively by the experiments of Michel-Meyer et al. (Reference Michel-Meyer, Shavit, Tsinober and Rosenzweig2020). In addition, some of these previous works and others have also focused on the nonlinear and constant-flux regimes, and particularly on the dissolution rate itself. Regarding this quantity, the different results obtained are less consistent between each other. Using a semi-analytical model and numerical simulations, Emami-Meybodi, Hassanzadeh & Ennis-King (Reference Emami-Meybodi, Hassanzadeh and Ennis-King2015) first showed that the dissolution rate initially starts to decrease when increasing the background velocity, but then resumes increasing when this velocity became of the order of the gravitational velocity. However, their initial model did not account for hydrodynamic dispersion. Using numerical simulations considering hydrodynamic dispersion, Emami-Meybodi (Reference Emami-Meybodi2017a
) later showed that the dissolution flux decreases when increasing either the background velocity or the dispersion, while keeping the other constant. On the contrary, the experiments of Michel-Meyer et al. (Reference Michel-Meyer, Shavit, Tsinober and Rosenzweig2020) did not show any significative variations of the flux with the background flow. Recent numerical simulations Tsinober et al. (Reference Tsinober, Rosenzweig, Class, Helmig and Shavit2022), designed to numerically replicate the experiments performed by Michel-Meyer et al, have also found that the flux is almost constant when the Péclet number increases, before finally being enhanced at large Péclet number.
These recent theoretical, numerical and experimental developments, where the combination of background flows and hydrodynamic dispersion are considered, have focused on both the linear and nonlinear regimes. In the linear regime, the main results are consistent with each other but the linear stability analyses have only been confirmed qualitatively by experiments. In the nonlinear regime, the results obtained on the flux evolution with the Péclet number do not converge towards a single behaviour and show some discrepancies. The present paper therefore proposes to fill these gaps by combining experimental, numerical and theoretical tools to investigate the growth rate and wavenumber (linear regime) as well as the dissolution flux (nonlinear regime). With this approach, we are therefore able to compare quantitatively the evolution of these quantities as a function of the Péclet number. We study convective dissolution in a Hele-Shaw cell in which a uniform background horizontal flow is applied. Since the medium does not contain any grain, Taylor–Aris dispersion – which is a crucial mechanism in microfluidics Stone, Strook & Ajdari (Reference Stone, Strook and Ajdari2004) – somehow plays the role of the hydrodynamic dispersion that occurs when using real granular media Taylor (Reference Taylor1953), Aris (Reference Aris1956). It is important to note that the Taylor–Aris dispersion scales as
$\textit{Pe}^2$
Taylor (Reference Taylor1953), Aris (Reference Aris1956), Wooding (Reference Wooding1960), while the hydrodynamic dispersion in a porous medium is essentially linear with respect to the Péclet number Souzy et al. (Reference Souzy, Lhuissier, Méheust, Le Borgne and Metzger2020). We also stress that the Taylor–Aris dispersion only affects the longitudinal diffusion coefficient, in the directional of the flow, while the hydrodynamic dispersion acts on both the longitudinal and transversal direction – with a coefficient for transverse dispersion approximately one order of magnitude smaller than the coefficient for longitudinal dispersion (Souzy et al. Reference Souzy, Lhuissier, Méheust, Le Borgne and Metzger2020). Despite these differences, our Hele-Shaw system provides a convenient way to get insight into the combined effects of a forced background flow and hydrodynamic dispersion on the onset of the convective dissolution process
This paper is organised as follows. In § 2, we present the experimental set-up, the numerical framework in which the simulations are performed and the system of equations that governs the problem. In § 3, we show the results obtained for the three main quantities which characterise the onset of the instability: growth rate, wavenumber and CO
$_2$
-dissolution rate. For these three quantities, the experimental results are compared with their numerical and theoretical counterparts. The main results concerning the effect of a transverse flow on the onset of the instability are summarised in § 4, together with the preliminary conclusions which can be drawn from our study regarding real systems.
2. Materials and methods
2.1. Experimental set-up and protocol
The effects of a transverse flow on the convective dissolution of carbon dioxide in an aqueous solution are studied by means of a large aspect ratio Hele-Shaw cell of average gap width
$E = 470\pm 50$
$\unicode {x03BC}$
m made out of polymethylmethaacrylate (PMMA), as shown in figure 1. The cell is filled up to a height
$H = 80$
mm with a
$5\times 10^{-4}$
mol l
$^{-1}$
aqueous solution of bromocresol green (BG) in its basic form (Bromocresol Green Sodium Salt, Sigma Aldrich). This concentration is ten times smaller than that of the dissolved
$\textrm {CO}_2$
so that it does not affect the dissolution process (Tanaka et al. (Reference Tanaka, Almarcha, Nagatsu, Méheust and Meunier2024). The contact line is pinned on the upper horizontal edge of the cell, as shown in the inset of figure 1. A transverse (horizontal) flow is forced through the cell parallel to the walls, from left to right, by means of two syringe pumps (NE-1000 SyringeONE, Darwin Microfluidics) – one acting as a pusher, the other (not represented) acting as a puller – in order to avoid any large spatial fluctuation of the level of liquid. The flow rate is initially fixed at a constant value (ranging from 1.25 to 25 ml h
$^{-1}$
) and then finely tuned manually during the experiment to prevent the creation of drops or bubbles at the interface. The aqueous solution is injected in the cell through four holes equally spaced over the height of the wet zone and located at 8 mm from the left vertical inner edge of the PMMA front plate. Likewise, after having travelled along the
$x$
-direction, the solution leaves the cell through four outlet holes, located at 8 mm from the right vertical edge of the same PMMA plate. The distance between the inlet and outlet holes is equal to 480 mm such that the useful length
$L$
– where the profile of the flow can be considered to be independent of the
$z$
coordinate (at least close to the liquid/gas interface) – is approximately equal to 450 mm. By means of the hydraulic system depicted above, the imposed flow rate can generate transverse mean velocities
$U_{0}$
up to 0.17 mm s
$^{-1}$
. Note that the value of
$U_{0}$
is double checked a posteriori (see § 2.2.2) and not only inferred from the value of the incoming flow rate. The exact physical meaning of
$U_{0}$
is clarified in § 2.4.
Sketch of the experimental set-up. A background transverse flow is forced through the Hele-Shaw cell using pushing and pulling syringe pumps. The carbon dioxide is first injected in the gas loop (red arrows) from a slightly overpressurised cylinder, and subsequently circulated at a constant rate by means of a gas pump. In order to avoid evaporation during the dissolution process the incoming (initially dry)
$\textrm {CO}_2$
passes through a water bubbling system in order to ensure a suitable level of humidity in the gas phase. See text for further explanation.

During each experimental run, the carbon dioxide is constantly circulated by means of a gas pump (22K Miniature Diaphragm Pump, Boxer Gmbh) through a closed-loop circuit (red arrows) consisting of a flexible plastic pipe and the
$\textrm {CO}_2$
cavity located above the liquid/gas interface (again see figure 1 and inset therein). At the very early stage of each experimental run, the pipe network and the cavity are filled with air instead of
$\textrm {CO}_2$
. In order to quickly replace the air by carbon dioxide, we proceed as follows. Initially, the purge tap is open and the regulator of the gas cylinder is set at a pressure that is slightly above the atmospheric pressure. The overpressure is large enough to force the carbon dioxide through the pipe network, and sufficiently small to avoid cell thickness expansion and thermal effects due to sudden gas compression. During the
$\textrm {CO}_2$
filling stage, the pump is left turned on so that the carbon dioxide coming from the cylinder preferentially goes into the cavity. After 15–20 s, there is virtually no more air left in the system and both the purge tap and the injection from the gas cylinder are closed. The carbon dioxide is now circulating in a closed-loop circuit until the end of the experiment. To account for the small amount of air left in the system, the partial pressure of carbon dioxide
$P_{\textrm{CO}_{2}}$
, which is equal to
$0.7\pm 0.5$
bar in all experiments, is monitored by means of a
$\textrm {CO}_2$
-sensor (SPRINTIR-W-F-100, Gas Sensing Solutions Ltd). Note the presence of a 1 l reservoir, the role of which is (i) to homogenise the air/carbon dioxide mixture and smooth out the possible fluctuations of the
$\textrm {CO}_2$
partial pressure and (ii) to mitigate the decrease of the
$\textrm {CO}_2$
partial pressure with time in the system (due to leaks and dissolution). The
$\textrm {CO}_2$
gas is also forced through water (bubbling Erlenmeyer flask) before injection into the closed-loop circuit, in order to avoid the evaporation of the BG/water mixture contained in the cell (that would generate a change in density of the solution due to thermal effects and BG concentration increase).
2.2. Measurement techniques
2.2.1. Carbon dioxide concentration
The Hele-Shaw cell is illuminated from the back with a white 900 cm
$^2$
square LED panel located 20 cm behind the cell and images are captured using a Nikon D200 camera equipped with a Nikkor 135 mm lens. Evaluation of the concentration of
$\textrm {CO}_2$
in the liquid is made possible from the local colour shade of the water–BG mixture by a suitable treatment of the colour map. Indeed, BG is a
$\mbox{pH}$
indicator with a logarithmic acidic constant
$\mbox{pK}_a = 4.9$
close to the pH of the 1 bar
$\textrm {CO}_2$
-saturated solution (which is of the order of 4) such that the full dynamics of the BG colour change is taken advantage of. This indicator is thus more suitable than
$\mbox{pH}$
indicators with higher
$\mbox{pK}_a$
s, the colour of which hardly changes in the range of
$\textrm {CO}_2$
concentration explored in our experimental conditions. The post processing method used to extract local
$\textrm {CO}_2$
-concentration fields is depicted in greater detail in Appendix A.
2.2.2. Measurement of the transverse background velocity
A reasonable value of the transverse velocity can be calculated from the volume flow rate, which can itself be inferred from the injection rate of the syringe pump. However, a more ‘local’ evaluation of the transverse velocity, especially in the region where solutal plumes develop, is needed to make the calculation of the Péclet number and more generally the whole quantitative analysis more reliable. Actually, the value of the transverse velocity is refined a posteriori by post-processing the spatio-temporal diagram obtained from a horizontal pixel line of concentration located at 5 mm beneath the liquid–gas interface. The initial spatio-temporal diagram obtained in the fixed (laboratory) frame of reference is straightened back by applying a time-dependent translation
$X \rightarrow X - U_0T$
, where
$U_0$
is a fitting parameter which must be adapted to get the straightest finger pattern possible. An example of the initial and rectified diagrams is given in figure 2. Here, the fitting value
$U_0$
of the velocity is found to be equal to
$0.09\pm 0.02$
mm s
$^{-1}$
. The approximate value for
$U_{\textit{SP}} = 0.074$
mm s
$^{-1}$
inferred from the syringe pump parameters thus falls within the acceptable interval associated with the a posteriori measurement. The fingers on the right are tilted toward positive
$X$
whereas the fingers on the left are tilted toward negative
$X$
. It indicates that there is a heterogenity (smaller than 10 % except for the smallest velocity) in the translation velocity between the left and the right of the cell, due to small fluctuations of the gap width. The values of the translation velocity are given in table 1.
(a) Initial spatio-temporal diagram obtained form a pixel line of concentration field located 5 mm below the interface. (b) Rectified diagram obtained after applying a back-in-time transformation on the spatial coordinate
$X\rightarrow X-U_{0}T$
.

Experimental data with the syringe pump flow rate, the translation velocity inferred from the flow rate
$U_{\textit{SP}}$
and measured from the plumes displacement
$U_0$
. The Péclet number is calculated using
$U_0$
. The measured partial pressure of
$\textrm {CO}_2$
is used to calculate the Rayleigh number and the characteristic velocity, length and time.

2.3. General formulation and dimensionless parameters
In the present study, it is assumed that the difference
$\Delta \rho$
between the density of the water-
$\textrm {CO}_2$
mixture and that of the pure water (denoted by
$\rho _{0}$
) varies linearly with respect to the concentration
$C$
of dissolved carbon dioxide, which can be written
$\Delta \rho = \rho _{0}\,\alpha \,C$
, where
$\alpha = (1/\rho _{0})\partial _{C}\rho$
refers to solutal coefficient of density change. We also consider that Henry’s law holds for low partial pressures of carbon dioxide (i.e. of the order of one bar), such that the saturation concentration at thermodynamic equilibrium is given by
$C^{{+}} = K_{{H}}P_{\mbox{CO$_{2}$}}$
, where
$K_{{H}}$
denotes Henry’s constant.
From there, we are able to propose typical scales for the velocity, length and time. A classical way to obtain a relevant velocity scale in porous media is to balance the gravity with the viscous forces, which yields
where
$\kappa = E^2/12$
is the permeability of the Hele-Shaw cell,
$\Delta \rho ^{{+}} = \rho _{0}\,\alpha K_{{H}}P_{\mbox{CO$_{2}$}}$
is the difference in density between the saturated
$\textrm {CO}_2$
-enriched solution and the pure water,
$g = 9.81$
m s
$^{-2}$
is the amplitude of the gravity field and
$\eta = 1.15 \times 10^{-3}$
Pa s is the viscosity of water, which depends very little on the amount of dissolved carbon dioxide (Vreme et al. Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016; De Reference De2021). In our experiments, this characteristic velocity ranges from 0.038 to 0.048 mm s−1 depending on
$\textrm {CO}_2$
partial pressure and temperature. As we are interested in the early stages of the gravitational instability, where the vertical dimension of the system plays no role, the relevant length scale
$\mathcal{L}$
is given by the ratio
where
$D_{0} = 1.3\times 10^{-9}$
m
$^{2}$
s
$^{-1}$
is the molecular diffusivity of carbon dioxide in water. By making the ratio between the previous length and velocity scales, one obtains the time scale of the problem, thus given by
The convective dissolution of
$\textrm {CO}_2$
is a solute transport process ruled by the Oberbeck–Boussinesq approximation of the Navier–Stokes equation (along with the continuity equation) and the classical convection-diffusion (solute transport) equation. By using
$\mathcal{U}$
,
$\mathcal{L}$
and
$\mathcal{T}$
as the characterictic velocity, length and time to make the problem dimensionless, the governing equations can be re-written in the following forms:
where
$\boldsymbol{\hat {z}}$
is the vertical unit vector oriented upwards (see figure 11 in Appendix A), and where
$\boldsymbol{u} = (u,v,w)$
,
$p$
and
$c$
respectively denote the velocity, pressure and concentration fields. In the following, incompressibility condition is assumed such that (2.4) and (2.5) are always implicitly complemented by the continuity equation
$\boldsymbol{\nabla }\boldsymbol{\cdot} \boldsymbol{u} = \boldsymbol{0}$
. Note that the pressure in (2.4) has been made dimensionless using the quantity
$\mathcal{P} = \eta D_{0}/\kappa$
. In (2.4) and in the rest of the paper, the Rayleigh, Darcy and Schmidt numbers are defined as follows:
The Rayleigh number ranges from 2400 to 3000 in the experiments; it is taken equal to 2700 in the numerical simulations. The Darcy and Schmidt numbers are respectively equal to
$3 \times 10^{-6}$
and 550 in both the experiments and the numerical simulations.
In the absence of translation speed, the problem is governed by these three non-dimensional parameters. The goal of our paper is to study the effect of the translation velocity
$U_{0}$
which is quantified by a non-dimensional parameter, the Péclet number based on the cell thickness given by
This Péclet number can also be written as a function of the non-dimensional translation velocity
$u_{0} = U_{0}/\mathcal{U}$
as
$\textit{Pe}=\sqrt {12}\,Ra\,Da^{1/2} u_{0}$
.
2.4. Thickness-averaged equations
With the aim of further simplifying the problem and deriving two-dimensional versions of (2.4) and (2.5), we make the so-called Brinkman assumption (Martin, Rakotomalala & Salin Reference Martin, Rakotomalala and Salin2002) that the flow is parabolic and has no component along the
$y$
-direction, which can be written
\begin{equation} \boldsymbol{u}(x,y,z) = \frac {3}{2}\langle \boldsymbol{u}(x,z)\rangle _y\left [1 - \left (\frac {2y}{e}\right )^2\right ] ,\end{equation}
with an
$y$
-averaged velocity defined as
With this definition, and assuming that the background flow does not depend on the vertical coordinate, the translation velocity
$u_{0}$
is exactly equal to the volume flow rate of the syringe pumps divided by the (dimensionless) cross section area
$e \times h$
of the Hele-Shaw cell.
Averaging (2.4) and (2.5) over the thickness of the cell and eventually dropping the average brackets
$\langle \,\boldsymbol{\cdot} \,\rangle _y$
, leads to new forms of the Navier–Stokes and convection-diffusion equations that no longer depend on the
$y$
-coordinate, as given by Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016). However, Letelier, Mujica & Ortega (Reference Letelier, Mujica and Ortega2019) showed that additional terms arise when doing an asymptotic expansion at small
$Da$
, leading for our dimensionless variables to
In the previous system, the two-dimensional gradient and Laplacian operators are defined as
$\boldsymbol{\nabla }_{\!{\Vert }}[\cdots ] = (\partial _x[\cdots ], \partial _z[\cdots ])$
and
$\boldsymbol{\nabla }_{\!{\Vert }}^2 = \partial _x^2[\cdots ]\,+\, \partial _z^2[\cdots ]$
,
$\gamma = 2Ra^2Da/35$
and the total diffusion tensor – which accounts for both molecular and Taylor–Aris dispersion – is given by
Here, the flow is two-dimensional rather than one-dimensional as in classical Taylor diffusion. Therefore, the diffusion coefficient is a tensor which, in the limit
$|\boldsymbol{u}| \ll u_{0}$
and for our experimental configuration, contains a contribution equal to
$\gamma u_{0}^2$
in the direction of the (horizontal) background flow (see Appendix B). In the above expression of
$\gamma$
, the
$2/35$
coefficient ensues from the bi-dimensional nature of the Taylor diffusion Wooding (Reference Wooding1960) while the factor
$Ra^2 Da$
comes from scaling parameter
$\mathcal{U}$
used to make the problem dimensionless.
For the sake of simplicity and to allow for the use of analytical results obtained in previous studies Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016), we eventually chose to work in the frame of reference moving at the
$y$
-averaged velocity of the horizontal forcing flow
${\boldsymbol{u}}_{0} = u_{0}\,\boldsymbol{\hat {x}}$
. Only the
$x$
-coordinate is affected by such a change, however, the time and space differential operators must be properly re-derived. Noting with a tilde the new fields, variables and operators and given that
$x = \tilde {x} + u_{0} t$
(or equivalently
$\boldsymbol{u} = \tilde {\boldsymbol{u}}+{\boldsymbol{u}}_{0}$
) and
$t = \tilde {t}$
, one obtains
$\boldsymbol{\nabla }_{\!{\Vert }} = \widetilde {\boldsymbol{\nabla }}_{\!{\Vert }}$
and
$\partial _{t} = \partial _{\tilde {t}} - u_{0} \partial _{\tilde {x}}$
. Introducing the previous expressions into (2.10) and (2.11) yields the final form of the governing equations that have been used for the numerical computations, and for the derivation of the wavenumber and growth rate (see section Appendix C)
\begin{align}& \frac {Ra^2\,Da}{S}\left [{\frac {6}{5}}\frac {\partial \tilde {\boldsymbol{u}}}{\partial \tilde {t}} + {\frac {12u_{0}}{35}}\frac {\partial \tilde {\boldsymbol{u}}}{\partial \tilde {x}} + {\frac {54}{35}}\left(\tilde {\boldsymbol{u}}\boldsymbol{\cdot} \widetilde {\boldsymbol{\nabla }}_{\!{\Vert }}\right)\,\tilde {\boldsymbol{u}}\right ] \notag\\&\quad= -\widetilde {\boldsymbol{\nabla }}_{\!{\Vert }}\,p + \left [\frac {{6} Ra^2Da}{{5}}\widetilde {\boldsymbol{\nabla }}_{\!{\Vert }}^2 - 1\right ] \tilde {\boldsymbol{u}} - \tilde {c}\,\boldsymbol{\hat {z}} {+ \gamma \tilde {\boldsymbol{u}}\boldsymbol{\cdot} \widetilde {\boldsymbol{\nabla }}_{\!{\Vert }} \tilde {c} \boldsymbol{\hat {z}},} \end{align}
Note that, in (2.14), due to the change of frame of reference, the Taylor–Aris dispersion tensor becomes
2.5. Numerical methods
We used the commercial code COMSOL Multiphysics© as a tool for consolidating our understanding of the destabilisation process. The full nonlinear equations (2.13) and (2.14) were solved in a two-dimensional fluid domain of dimensions
$l_{{num}} \times h_{{num}} = 5000 \times 2700$
using an implicit Backward Differentiation Formula solver with adaptative time step. The fluid domain is meshed with triangular elements of maximum (dimensionless) size
$\delta _{ max } = 75$
. The upper boundary layer (i.e. beneath the liquid–gas interface) extends over a vertical distance
$0.16\times h_{{ num}}$
and contains 80 elements (stretching factor 1.05) whereas the boundary layer at the bottom wall only contains 8 elements. The whole meshing is composed of 11 716 volume elements with 680 additional elements at the boundaries. We used a cubic discretisation for the concentration and a P2
$+$
P1 scheme for velocity and pressure fields, which finally led to a 1 59240 degrees of freedom model.
Periodic boundary conditions are imposed on the vertical edges of the domain for both velocity and concentration fields. A zero-flux boundary conditions is taken at the bottom of the cell and concentration at the liquidgas interface is set to the fixed value
$c^{{+}} = 1$
(dimensionless saturation concentration). A free-slip condition is considered at both the free surface and the bottom wall. The simulations are started with the fluid being motionless in the moving frame of reference (i.e.
$\tilde {\boldsymbol{u}} = \boldsymbol{0}$
) and the concentration profile being of the form
$\bar {c}(z,400) + \omega$
with
$\bar {c}(z,t) = 1 + \mbox{erf}(z/\sqrt {4 t})$
and with
$\omega$
a positive random initial noise of mean square value
$5 \times 10^{-5}$
. The equations are integrated for a time
$t$
varying from
$t=400$
to
$t=20\,000$
or
$50\,000$
depending on the growth of the instability. This choice of initial time allows for a flux being initially equal to the diffusive flux
$1/\sqrt {\pi t}$
.
As mentioned above, all the numerical simulations were performed in the moving frame of reference. Therefore, as a preliminary validation step, we checked on few cases that the resolution of (2.13) and (2.14) (the moving frame), and (2.10) and (2.11) (laboratory frame) led to similar results as far as the wave length and the growth rate of the instability were concerned.
3. Convective dissolution in the presence of a transverse flow
3.1. Concentration field visualisation
Three experimental concentration fields obtained at
$t \simeq 2000$
(
$T \simeq 20$
min) for increasing values of the translation velocity
$u_{0}$
are shown in figure 3. Only the right part of the Hele-Shaw cell, that is, the region of fluid domain which has been exposed to carbon dioxide from the very beginning of the experiment, is shown in the figure (see next paragraph for further explanations). At low transverse velocities (
$\textit{Pe}=2$
), the plumes grow exactly as they do when no background flow is present (Vreme et al. (2016), which indicates that the Taylor–Aris dispersion does not alter the instability. For intermediate Péclet number (
$\textit{Pe}=32$
), the observed (most unstable) wavelength
$\lambda ^\star$
is larger than expected in the absence of transverse flow whilst the amplitude of the plumes is slightly smaller (again at
$t \simeq 2000$
), which means that the growth rate of the instability is smaller. In other words, the time it takes for the instability to reach the nonlinear regime becomes larger than for the case
$\textit{Pe} = 0$
. At large values of the velocity at (
$\textit{Pe}=60$
), the growth rate becomes even smaller – while the most unstable wavelength continues to increase – so that eventually, no convective disturbance of the pure diffusive state can be observed. For such a large Péclet number, the plumes cannot be observed in the cell (no matter how long we wait) since the time it takes for the instability to grow becomes larger than the time it takes for the fluid to be translated from the left to the right of the cell – i.e. a longer cell would be required to acquire measurements at large translation velocity.
Experimental fields of concentration for different translation velocities at
$t=2000$
(
$T \simeq 20$
min). The thickness-based Péclet number
$\textit{Pe}=U_{0} E/D_{0}$
is equal to 2 in (a), 32 in (b) and 60 in (c). The corresponding movies are available for Péclet number equals 2, Péclet number equals 32 and Péclet number equals 60.

The observed trends regarding the wavelength and the growth rate are confirmed by numerical simulations, as shown in figure 4, where the numerical counterparts of the experimental field maps (figure 3) are displayed. For the smaller velocity (
$\textit{Pe} = 2$
), the system has entered the nonlinear regime and plumes have already experienced merging. For the intermediate value of the velocity (
$\textit{Pe} = 32$
) the instability is only starting and a weak undulation of the front is visible. For the largest velocity (
$\textit{Pe} = 60$
), the front has remained flat and no instability has appeared yet. Note that the numerical field maps are obtained for a larger dimensionless time
$t \simeq 10^4$
(
$T \simeq 1$
h 40 mn) than in the experimental figure. This apparent delay in the emergence of the instability is due to an initial noise amplitude deliberately set at a value (of approximately
$10^{-4}$
) lower than the amplitude in the experiments (initially equal to approximately
$0.02$
as extrapolated from figure 5
a). This allows for a longer linear stage and thus a better characterisation of the instability growth rate. However, there still is an excellent agreement for the growth rate and the wavelength between experiments and numerics, as will be shown in the following.
Numerical fields of concentration obtained for different translation velocities at
$t=10^4$
. The thickness-based Péclet number
$\textit{Pe}=U_{0} E/D_{0}$
is equal to 2 in (a), 32 in (b) and 60 in (c).

3.2. Growth rate of the instability
In experiments, the amplitude
$\mathcal{A}(t)$
of the instability is evaluated as the root mean square of the average concentration perturbation
\begin{equation} \mathcal{A}(t) = \left \{ \frac {1}{\Delta x\,\Delta z}\!\int _{z_{0}}^{\,z_{{1}}}\!\!\!\int _{x_{0}(t)}^{l} \left [c\,(x,z,t) - {\bar {c}}\,(z,t)\right ]^{\,2}\,{\rm d}x\,{\rm d}z\right \}^{1/2}\!, \end{equation}
In the above expression, the mean concentration is defined as
and the quantities
$\Delta x = l - x_{0}(t)$
and
$\Delta z = z_{{1}} - z_{0}$
refer to the dimensions of the integration domain. The boundaries of the integration
$x_{0}(t)$
,
$l$
,
$z_{0}$
and
$z_{{1}}$
domain are chosen as follows. In the numerical simulations, the surface average can be taken over the whole domain because the periodic boundary conditions permit to (somehow artificially) expose the entire free surface to
$\textrm {CO}_2$
at any time. On the contrary, at a given time
$t$
, the left part of the experimental system – i.e. for
$0 \lt x \lt u_{0} t$
– contains particles which have entered the cell after the start of the experiment, and therefore cannot be accounted for in the analysis. The integral average (3.1) along the
$x$
-direction is thus calculated over the sliding interval
$[x_{0}(t) = u_{0} t, l = L/\mathcal{L}]$
, within which the liquid has ingested
$\textrm {CO}_2$
from the very beginning of the experiment. On the other hand, both the lower part of the images, which mainly contains noise, and the thin layer located in the vicinity of the interface, which is marred by the presence of reflections, should be excluded from the calculation of the integral average (3.1). Therefore, the boundaries of integration along the
$z$
-direction are fixed to
$z_{0} = Z_{0}/\mathcal{L}$
(with
$Z_{0} = -25$
mm) and
$z_{{1}} = Z_{{1}}/\mathcal{L}$
(with
$Z_{{1}} = -2$
mm) instead of
$-h$
and
$0$
. Note that, in principle, the experimental average profile (3.2) should be very close to the theoretical profile
$1 + \mbox{erf}(z/\sqrt {4 t})$
. Therefore, for the sake of simplicity, we used the same notation
$\bar {c}$
for the theoretical diffusive profile and its experimental counterpart.
Experimental (a) and numerical (b) amplitude
$\mathcal{A}(t)$
of the perturbation as a function of time for three different values of the Péclet number (lin–log scale).

The amplitude of the perturbation as a function of time is shown (lin–log scale) in figure 5 for three values of the Péclet number. As expected, we observe an initial increase, characteristic of the linear regime, where the amplitude grows nearly exponentially. As the Péclet number is increased, not only does the instability start later but it also has a weaker growth rate. Consequently, at a given time
$t=1000$
, the amplitude is much larger for
$\textit{Pe}=2$
(where (
$\mathcal{A}=0.02$
) than for
$\textit{Pe}=60$
(where
$\mathcal{A}=10^{-4}$
). This quantifies the effect of the Péclet number, that was observed qualitatively in figure 3. It can be noted that, as already mentioned, the linear stage lasts longer in the numerical simulation than in the experiments because the initial amplitude has been chosen much smaller to better characterise the instability. Beyond this first stage, the amplitude starts saturating, which is the signature of the nonlinear regime, where the solute plumes are fully formed and pour at a constant velocity down to the bottom of the cell (prior to merging and subsequent births).
Experimental (solid symbols) and numerical (open symbols) growth rate as a function of the Péclet number, for
$Da = 3\times 10^{-6}$
and
$Ra$
in the interval
$[2500 {-} 3000]$
. The theory (solid line) provides the relevant trends in the considered range of Péclet number, but systematically underestimates its experimental and numerical counterparts by a factor of up to two. The dashed line corresponds to the prediction at
$\textit{Pe}=0$
, i.e. without translation velocity (Vreme et al. Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016).

The experimental and numerical values of the growth rate
$\sigma ^\star$
are extracted from profiles such as those presented in figure 5. In the experiments, the method consists in visually identifying the linear regime on the lin–log graph
$\mathcal{A}(t)$
, and finding the average slope
$\sigma ^\star$
of the curve over the corresponding interval (dashed lines as in figure 5
a). It is important to note that the very choice of the linear region leads to a 50 % uncertainty in the measurement of the growth rate. In the numerical simulations, the method of extraction of the growth rate is more robust, yet somewhat arbitrary. The growth rate
$\sigma ^\star$
is calculated from the value of the slope of the graph at the end of the linear regime – i.e. when the mass flux departs from the diffusive ‘
$t^{-1/2}$
’ profile (see § 3.4, figure 9
b). Because the slope of the graph increases only by a few per cent between the beginning and the end of the linear stage, the uncertainty on the numerical value of the numerical growth rate is approximately 10 %.
The observed growth rates
$\sigma ^\star$
are plotted in figure 6 as a function of the Péclet number. Despite large uncertainties in the experimental data, there is a very good agreement between experimental and numerical values. The growth rate is insensitive to the translation velocity for small Péclet number, with a value very close to the usual value 0.002–0.003 (dashed line in figure 6) found in the literature Elenius & Johannsen (Reference Elenius and Johannsen2012), Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016). Above a Péclet number of approximately 20, the growth rate starts decaying algebraically: it is found to be approximately 10 times smaller for the largest velocity (
$\textit{Pe} = 90$
).
The experimental and numerical growth rates are compared with an analytical prediction adapted from Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016) for the case where a horizontal background flow is present (see Appendix C for further details). Assuming that the vertical velocity is equal to
$\check {w}\,e^{i(qx+\sigma t)}$
and that the perturbation of concentration is
$c-{\bar {c}} = \check {c}\,e^{iqx+\sigma t}$
, the equation for the concentration (2.14) yields after linearisation
\begin{equation} \left [\sigma + \frac {\textit{Pe}^2 q^2}{210} +q^2 - D_{{\tilde {z}}}^2\right ]\,\check {c} = - \check {w} \frac {\partial {\bar {c}}}{\partial {\tilde {z}}}. \end{equation}
This equation clearly shows that the effect of the background flow can be easily taken into account by considering an ‘altered’ growth rate
$\sigma + \textit{Pe}^2 q^2/210$
. Vreme’s analytical formula (A17), obtained in the absence of background flow, is simply modified into
with
$\chi ^2 = q^2+\sigma + \textit{Pe}^2 q^2/210$
. For
$\textit{Pe}=0$
, we recover their analytical formula in the Darcy regime (i.e. neglecting the Brinkman terms in
$Ra^2 Da$
). In this equation, they defined
$\eta =2\sqrt {\tilde {t}}$
as the thickness of the diffusive layer and used
$\zeta = 0.46$
. The maximum growth rate
$\sigma ^\star$
, obtained by solving numerically the modified dispersion relation (C6), is plotted as a solid line in figure 6. Although the theory underestimates the growth rate by a factor of up to two, it correctly predicts the transitional Péclet number (
$\textit{Pe} \sim 20$
) and the scaling law
$\sigma ^\star \sim \textit{Pe}^{-4}$
– which is derived analytically in the limit of large Péclet numbers in Appendix C. It should be noted that no additional fitting parameter has been introduced in the model compared with Vreme’s theory.
3.3. Wavenumber of the instability
In order to identify the most unstable wavenumber of the instability, the spectrum (norm of the Fourier-transform squared) of the experimental concentration field is calculated for each
$x$
-line, and then averaged in the
$z$
-direction (in order to increase the signal to noise ratio). This amounts to calculating the following quantity:
\begin{equation} \Gamma (q) = \int _{z_{0}}^{z_{{1}}} \left | \int _{x_{0}(t)}^l c(x,z) e^{iqx}{\rm d}x \right |^2 {\rm d}z ,\end{equation}
where the boundaries of the region of integration have been defined in the previous section.
Spectra of concentration obtained experimentally (a) and numerically (b) at the end of the linear stage for three values of the Péclet number.

Experimental (solid symbols) and numerical (open symbol) most unstable wavenumber as a function of the Péclet number for
$Da = 3\times 10^{-6}$
and
$Ra$
in the interval
$[2500 {-} 3000]$
. The dashed line corresponds to the prediction at
$\textit{Pe}=0$
, i.e. without translation velocity (Vreme et al. Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016).

The renormalised spectra are plotted in figure 7 at a time corresponding to the end of the linear regime (as defined in the previous paragraph). The wavenumber of the maximum peak strongly decreases when the translation velocity increases, going from approximately 0.05 at
$\textit{Pe} = 2$
to approximately 0.005 at
$\textit{Pe} = 60$
. This is consistent with the concentration fields displayed in figures 3 and 4, which show that the wavelength increases with the Péclet number. Although there is a large uncertainty in the experimental data, the agreement between the experimental and numerical maxima is satisfactory, with a maximum difference of the order of 20 %. It should be noted that the spectral resolution for the experiments has been increased by embedding the data in a larger domain and using a Hann window. However, the noise in the experimental spectra is still larger than in the numerical ones.
The value of the maximum peak extracted for each experimental and numerical spectra is plotted in figure 8 as a function of the Péclet number. For small translation velocity (
$\textit{Pe} \lt 15$
), the wavenumber is almost independent of the Péclet number, as expected. However, its value is lower than that found in the literature (horizontal dashed line in figure 8) by 20 %. For Péclet numbers larger than 15, the wavenumber starts decreasing algebraically, reaching, for
$\textit{Pe} = 90$
, values approximately one order of magnitude smaller than the small Péclet limit. Again, the experimental and numerical values of
$q^\star$
are compared with those obtained from the model (see Appendix C). As for the growth rate, the model very well predicts the transition Péclet number (
$\textit{Pe} \sim 20$
) and yields a scaling law
$q^\star \sim \textit{Pe}^{-5/2}$
which properly reproduces the decay of the experimental and numerical data. An excellent quantitative agreement is even obtained for large Péclet numbers.
3.4. Flux of carbon dioxide at the interface
In our experiments, the concentration is measured quantitatively rather than simply estimated from the light intensity. It is thus possible to measure the total quantity of
$\textrm {CO}_2$
which has been ingested and infer from its temporal derivation the flux of carbon dioxide passing through the interface
\begin{equation} J=\frac {d}{{\rm d}t} \left [ \frac {1}{\Delta x}\!\int _{z_{0}}^{\,z_{{1}}}\!\!\!\int _{x_{0}(t)}^{\,x_{{1}}(t)} c(x,z) \,{\rm d}x\,{\rm d}z \right ] ,\end{equation}
Experimental (a) and numerical (b) flux (rate of mass transfer) plotted with respect to the Péclet number for three values of the Péclet number. In (b), the minimum flux, which corresponds to the end of the linear regime – where the numerical growth rate is estimated – is marked with a dot (red for
$\textit{Pe} = 2$
, green for
$\textit{Pe} = 32$
and blue for
$\textit{Pe} = 60$
).

Flux at saturation as a function of the Péclet number. The experimental data (solid symbols) are marred by a very large uncertainty, whereas the numerical values (open symbols), inferred from time profiles such as those presented in figure 9, are more reliable.

where the bounds of the integral are defined in § 3.2. This flux, which is plotted in figure 9(a) as a function of time, is marred by a large dispersion, which make the data difficult to use for quantitative purposes (especially for
$\textit{Pe} = 32$
). However, some general observations can be made: for the largest Péclet number, the value of the flux is close to the diffusive flux during the whole experiment since the instability never appears. For the smallest Péclet number, the flux leaves the diffusive regime at
$t \simeq 400$
and saturates at a value of approximately 0.04, not too far from the value 0.02 found in the numerical simulation (see figure 9
b).
The flux obtained from the numerical simulations, is shown in figure 9(b). The time profile of the (numerical) flux is consistent with what is reported in the literature: in the linear regime, the flux remains equal to the diffusive flux
$J_{{D}} = 1/\sqrt {\pi t}$
, calculated from the purely diffusive profile diffusion of
$\bar {c}(z,t) = 1 + \mbox{erf}(z/\sqrt {4 t})$
, plotted as a dashed line in figures 9(a) and 9(b); at the end of the linear regime, the flux abruptly increases and then tends to saturate at a value
$J_{\textit{NL}}$
which is a decreasing function of the Péclet (the flux at saturation
$J_{\textit{NL}}$
decreases by a factor 3 when increasing the Péclet number from
$\textit{Pe} = 2$
to
$\textit{Pe} = 60$
). Note that the sudden flux increase – that is the transition towards the nonlinear regime – is also delayed by the presence of the background flow. Note also that the Rayleigh number is not high enough to reach the constant-flux regime (Slim Reference Slim2014). The value of the flux may thus be slightly overestimated by approximately 20 %. The saturation flux
$J_{\textit{NL}}$
measured in the nonlinear regime is plotted as a function of the Péclet number in figure 10. The effect of the background flow is less prominent than in the cases of the growth rate and wavenumber. It is only reduced by a factor approximately 5 between the small Péclet number case (very low background translation velocity) and the case with fastest background translation velocity (
$\textit{Pe} = 90$
). Note that the experimental values are larger than the numerical values by more than 50 %. However, an offset in the calibration of the experimental intensity can lead to a strong bias of the flux, up to 50 %, which could explain this overestimation.
4. Conclusion
In this paper, we have shown how the presence of a mean horizontal translation would strongly inhibit the convective dissolution of
$\textrm {CO}_2$
in a Hele-Shaw cell. The Taylor–Aris dispersion, which acts in the direction of the local flow, homogenises the transverse undulations of the
$\textrm {CO}_2$
solutal front during the onset of the gravitational instability, thus delaying the formation of convective plumes. Experimentally, the growth rate of the instability is found to be approximately 10 times smaller for a Péclet number
$\textit{Pe}=U_{0} E/ D_{0}$
(based on the translation velocity and the cell thickness) of the order of 60 than in the absence of mean background flow. Likewise, the most unstable wavelength is also strongly affected by the presence of the horizontal background flow: for
$\textit{Pe} \simeq 60$
, the wave length is about ten times larger than the ‘natural’ wavelength, when no flow is forced through the cell.
All the trends observed in experiments (and confirmed by two-dimensional numerical simulations) on the growth rate and wavenumber can be explained theoretically by adding in the diffusion equation (considered in its dimensional form) a Taylor–Aris dispersion term
$D_{0} \partial _{\scriptscriptstyle{\textit{XX}}} \textit{Pe}^2/210$
. By doing so, one introduces an additional damping term
$D_{0} Q^2 \textit{Pe}^2/210$
(where
$Q = q/\mathcal{L}$
refers to the physical wavenumber) in the initial expression of the growth rate of the instability. Using Vreme’s analytical model (Vreme et al. 2016), we can thus predict the most unstable wavenumber and growth rate with no additional fitting parameter. The obtained solution is in good agreement with the experimental and numerical results. For large translation velocities, that is large Péclet numbers, the most unstable wavenumber (in its physical dimensional form) becomes asymptotically equal to
while the corresponding growth rate asymptotically tends towards
In the experiments, the
$\textrm {CO}_2$
concentration has been measured quantitatively for the first time thanks to BG, which has
$pKa \simeq 4$
, well suited to the pH of the saturated solution. This technique allowed for the measurement of the flux of carbon dioxide in time. Although the experimental data were marred by considerable uncertainty, we observed the saturating flux – i.e. the flux measured at the beginning of the saturation period, in the nonlinear regime – was weakly influenced by the presence of the background horizontal flow. Indeed it is only reduced by a factor approximately 3 when changing from
$\textit{Pe} \simeq 0$
to
$\textit{Pe} = 60$
, as confirmed by numerical simulations.
It is interesting to try to extrapolate these results to the case of an actual porous medium. In carbon sequestration, subcritical
$\textrm {CO}_2$
, once injected in the aquifer, typically encounters an impermeable rock layer on its ascent. The impermeable nature of this so-called cap rock forces the carbon dioxide to move horizontally, naturally creating the background flow analysed in this paper. However, as mentioned in the introduction, hydrodynamic dispersion in porous media is proportional to the Péclet number (rather than to the square of the Péclet number, as in the case of Taylor–Aris dispersion). In a porous medium, the most unstable wavenumber is thus expected to scale as
$\textit{Pe}^{-5/4}$
and the corresponding growth rate as
$\textit{Pe}^{-2}$
.
This extrapolation, however, holds for two-dimensional configurations and with longitudinal dispersion only (i.e. without transverse dispersion). In addition, Hassanzadeh et al. (Reference Hassanzadeh, Pooladi-Darvish and Keith2009) showed that the transverse dispersion also delays the onset of the instability with a characteristic time proportional to the quantity
$1+\alpha _{{\perp }}\,U_{0} D_{0}$
, (where
$\alpha _{{\perp }}$
the transverse dispersivity). In a two-dimensional porous medium, the effect of the transverse dispersion may thus be superimposed to the effect of the longitudinal dispersion described in this paper.
Finally, real porous media are three-dimensional, allowing for two types of instabilities. In three dimensions, gravitational instability may develop as plumes; which are expected to be damped by longitudinal dispersion in the presence of a horizontal background flow. However, gravitational instability may also take the form of two-dimensional rolls. If the wavevector of the rolls is oriented perpendicular to the background flow, the concentration field of the perturbation remain invariant in the background flow direction so that longitudinal dispersion is expected to have no effect on the linear growth of the instability (Hassanzadeh et al. Reference Hassanzadeh, Pooladi-Darvish and Keith2009). However, it may modify the nonlinear regime of the instability.
The problem is therefore more complex in real-world conditions. Further research is necessary to improve predictions of convective dissolution in the context of
$\textrm {CO}_2$
sequestration.
Supplementary movies
Supplementary movies are available at https://doi.org/10.1017/jfm.2025.10411.
Declaration of interests
The authors report no conflict of interest.
Appendix A. Measurement of the concentration
We choose to use the red intensity since it has the maximum rate of change during the experiment (by a factor up to 3). According to Beer–Lambert’s law, the red component of each image at location
$(X,Z)$
is equal to the intensity
transmitted through the cell. In the above expression,
$I_i(X,Z)$
refers to the intensity of the red component of the incoming light, which can be measured by using pure (clear) water instead of the BG aqueous solution. Equation (A1) can first be used to measure the thickness of the cell in the absence of
$\textrm {CO}_2$
(i.e. for a uniform
$\mbox{pH}$
close to 7) as
The result is plotted in figure 11, where the fitting parameter
$\ell _{{B}L}(7)$
is taken equal to
$590$
$\unicode {x03BC}$
m in order to get
$E = 470\pm 50$
$\unicode {x03BC}$
m at the top of the cell (where the thickness can be directly measured from above by optical methods). From the figure, it can be inferred that the thickness is uniform within 10 % over the whole Hele-Shaw cell.
Thickness of the Hele-Shaw cell measured from the intensity of the images at
$T=0$
using (A2) with
$\ell _{{B}L} = 590$
$\unicode {x03BC}$
m.

Calibration curve for the measurement of the
$\textrm {CO}_2$
-concentration. The theoretical concentration corresponding to a diffusive profile (A4) is plotted as a function of the ratio of intensity
$I/I(T=0)$
averaged over
$X$
at
$T=300s$
(red),
$T=600s$
(green) and
$T=1200s$
(blue). The experiment corresponds to a large translation velocity
$U_{0} = 0.167$
mm s−1 (
$\textit{Pe} = 60$
). The solid line corresponds to the fitting function (A5).

In principle, the Beer–Lambert law could also be used to obtain the
$\textrm {CO}_2$
concentration provided
$\ell _{{B}L}(\mbox{pH})$
is known, by using a theoretical relationship between the
$\mbox{pH}$
and the concentration
$C$
in carbon dioxide. However, the law
$\ell _{{B}L}(\mbox{pH})$
is not known with enough accuracy and would have to be calibrated as well. It is thus simpler to empirically determine the dependence of
$\ell _{{B}L}$
with respect to the
$\textrm {CO}_2$
concentration or, equivalently, the image (red) intensity as a function of the
$\textrm {CO}_2$
concentration. To avoid having to take an image with pure water for each experiment, we rather use the ratio between the images taken at a time
$T \gt 0$
and initial time
$T = 0$
. From (A1), such a ‘renormalised’ intensity is equal to
where
$f$
is an unknown function to be determined empirically. To do so, we used a set of images obtained for a large background transverse velocity (
$U_{0}=0.167$
mm s−1), such that no instability was observed and the concentration profile was a purely diffusive one at any time
In the experiment, the intensity is nearly independent of the
$X$
position since the concentration front is nearly flat. We can thus plot the ratio
$I(Z,T)/I(Z,T=0)$
(averaged over
$X$
) with respect to the estimated (theoretical) concentration
$\bar {C}(Z,T)$
at a given time
$T \gt 0$
(thus eliminating the
$Z$
space variable). The calibration curves obtained for three different times
$T = 5$
, 10 and 20 min are shown in figure 12. As expected, the relationship
$C = f[I/I(T=0)]$
which links the concentration
$C$
to the renormalised intensity is nearly independent of the time, and a function of the form
allows for a good fit of the empirical law to our experimental data. This (dimensional) calibration abacus has been used throughout the entire experimental campaign to measure the local
$\textrm {CO}_2$
-concentration.
Appendix B. Derivation of the effective diffusion tensor
We derive in the present section the Taylor–Aris dispersion term in a Hele-Shaw cell as done by Wooding (Reference Wooding1960). We assume that the mean flow in the
$x$
and
$z$
directions has a parabolic profile along
$y$
and is uniform along
$x$
and
$z$
. In a rotated frame
$(x',z')$
where
$x'$
-axis is aligned with the direction of the flow, the parabolic profile is thus given by
where
$\langle u'\rangle _y$
is the
$y$
-averaged velocity amplitude and
$\boldsymbol{\hat {x}}'$
the unit vector along the
$x'$
-direction. Following Taylor (Reference Taylor1953), we then write the advection-diffusion equation (2.5) in the frame moving at
$\langle u' \rangle _y$
(i.e. using the change of coordinate
$x'=x''+\langle u' \rangle _y\,t$
)
Note that, in the above equation, the molecular diffusive term
$\partial _{xx} c$
has been (temporarily) neglected. The solution at late times (i.e. neglecting the time derivative) to (B2) is
which satisfies the no-flux boundary condition
$\partial _y c=0$
at
$y=\pm e/2$
. The average flux of solute in the
$x'$
-direction
is thus equal to
where
$\partial _{x''}[\cdots ]$
has been replaced by
$\partial _{x'}[\cdots ]$
. The above expression is already derived in Wooding (Reference Wooding1960) (page 503) with the notation
$2h$
instead of
$e$
for the dimensionless thickness. As the
$x'$
-axis of the rotated frame has been chosen parallel to the direction of the flow
$\boldsymbol{u} = (u,w)$
such that the change of variable takes the simple form
where
$\cos \theta =\langle u \rangle _y/\langle u' \rangle _y$
,
$\sin \theta =\langle w \rangle _y/\langle u' \rangle _y$
, and where we remind the reader that
$\langle u' \rangle _y = [\langle u \rangle _y^2+\langle w \rangle _y^2]^{1/2}$
.
Actually, the flux of solute is
$F_x'$
is the
$x'$
-component of a vector flux aligned with
$x'$
-direction, that is
\begin{equation} \displaystyle \textbf{F} = -\frac {e^2}{210}\langle u' \rangle _y^2 \frac {\partial c}{\partial x'} \left (\!\! \begin{array}{cc} \cos \theta \\ \sin \theta \end{array} \!\!\right ) = -\frac {e^2}{210}\langle u' \rangle _y^2 \left ( \!\!\begin{array}{cc} \displaystyle \cos ^2 \theta \frac {\partial c}{\partial x} + \frac {1}{2}\sin 2\theta \,\frac {\partial c}{\partial z} \\[3mm] \displaystyle \frac {1}{2}\sin 2\theta \, \frac {\partial c}{\partial x} + \sin ^2 \theta \frac {\partial c}{\partial z} \end{array} \!\!\right )\!, \end{equation}
since
$\partial _{x'} = \cos \theta \,\partial _x[\cdots ] + \sin \theta \,\partial _z[\cdots ]$
. Therefore, the Taylor–Aris flux is finally equal to
$\textbf{F}=-{\textbf{D}}\,\boldsymbol{\nabla }_{\!{\Vert }}\,c$
with a diffusion tensor given by
given that
$e^2=E^2/\mathcal{L}^2=12 Ra^2 Da$
. Adding the molecular diffusion dimensionless tensor – i.e. the identity matrix – to the above expression and dropping the brackets leads to the expression of the complete diffusion tensor (2.12), equal to the term derived by Letelier et al. (Reference Letelier, Mujica and Ortega2019).
Appendix C. Dispersion equation in the Darcy limit
We aim at performing a linear stability analysis of the diffusive base state in the presence of a forced transverse flow. To do so, and for the sake of simplicity, we work in the moving frame, that translates along the
$x$
-direction at velocity
$u_{0}$
. We first write the concentration as the sum of a base leading-order term – which is nothing more than the pure diffusion profile – and a first-order harmonic correction of wavenumber
$q = 2\pi /\lambda$
and a growth rate
$\sigma$
As for the velocity
$\tilde {\boldsymbol{u}}$
, it is a first-order quantity (no flow in the purely diffusive regime) which can be written as
We now introduce these expansions in (2.14), in order to derive the leading- and first-order governing equation. Since the advection term is at first order, the base concentration profile
$\bar {c}$
is governed by a simple diffusion equation so that one simply gets
It should be noted that the diffusion tensor
$\tilde {{\textbf{D}}}$
in (2.14) introduces a term
$D (1+\gamma u_{0}^2)\partial _{\tilde {x} \tilde {x}}$
which, fortunately, has no effect on the base concentration profile since it is invariant with respect to
$\tilde {x}$
.
At next order, the equation for the concentration (2.14) yields the first following linearised equation
\begin{equation} \left [\sigma + \frac {\textit{Pe}^2 q^2}{210} +q^2 - D_{{\tilde {z}}}^2\right ]\,\check {c} = - \check {w} \frac {\partial {\bar {c}}}{\partial {\tilde {z}}}. \end{equation}
Inside the brackets of the right-hand side of the above equation, the term
$\textit{Pe}^2 q^2/210$
comes from the second-order derivative
$\gamma u_{0}^2\partial _{\tilde {x} \tilde {x}}[\cdots ]$
(since
$\gamma u_{0}^2=\textit{Pe}^2/210$
), which is the key term regarding the effect of the translation.
A second linearised equation relating
$\check {c}$
and
$\check {w}$
is derived by taking twice the curl of (2.13) and neglecting the left-hand term (since
$S=550$
is large compared with unity), and one obtains for the
$\tilde {z}$
-component
Here, the Brinkman terms which scale as
$Ra^2 Da$
and
$Ra Da^{1/2}$
can be neglected since we focus here on the Darcy regime exclusively. In principle, the growth rate of the instability could be found by solving analytically equations (C4) and (C5) along with a suitable set of boundary conditions. However, one can notice that the previous equations, which take into account the presence of a translation background velocity, are formally identical to the equations derived by Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016) in the case
$\textit{Pe} = 0$
, provided an ‘amended’ growth rate
$\sigma + \textit{Pe}^2 q^2/210$
is considered instead of
$\sigma$
(in the Darcy regime). We can thus formally use the dispersion equation derived by Vreme et al. (Reference Vreme, Nadal, Pouligny, Jeandet, Liger-Belair and Meunier2016) (see their equation (A17)), and then plug in it the expression of the corrected growth, which yields
where
$\eta =2\sqrt {\tilde {t}}$
is the thickness of the diffusive layer,
$\chi ^2 = q^2+\sigma + \textit{Pe}^2 q^2/210$
and
$\zeta = 0.46$
. As a reminder, two key assumptions have been made to derive the above version of the dispersion equation. First, a quasi-steady state assumption has been formulated: the base state
$\bar {c}({\tilde {z}},\tilde {t})$
is considered as ‘frozen’ at any time so that the thickness
$\eta$
is not considered explicitly as a function of time. Second, the problem has been solved by initially considering a step-like base concentration profile.
The dispersion equation (C6) has three unknowns
$\sigma$
,
$\eta$
and
$q$
. In order to close the system and reduce the number of unknowns, we further need to assume that the growth occurs on a time scale which is of the same order of magnitude as the time spent by the experiment to build the thickness, which can be written as follows:
From (C6) and (C7), it can be shown that the function
$\sigma (q)$
has a single maximum
$\sigma ^\star$
located at the most unstable wavenumber
$q^\star$
. The latter is searched numerically for each Péclet number and plotted in figure 13(a). In the small Péclet number limit, we recover the classical value
$q^\star = 0.0067$
while above a Péclet number of approximately 20, the wavenumber decays algebraically with an exponent
$-5/2$
. The corresponding (maximum) growth rate
$\sigma ^\star$
is plotted in figure 13(b). Likewise, it is equal to the classical value 0.0024 for small Péclet numbers and then decreases as
$\textit{Pe}^{-4}$
. The corresponding thickness (at the onset of the instability)
$\eta ^\star$
and quantity
$\chi ^\star$
are also plotted in figure 13(c,d) as an additional information.
Analytical solutions as a function of the Péclet number for the most unstable wavenumber (a), the maximal growth rate (b), and the corresponding layer thickness (c) and quantity
$\chi$
(d). The values are obtained by maximising
$\sigma$
in (C6) while assuming
$\eta =\sigma ^{-1/2}$
. The thin dashed lines are the asymptotic values at small and large Péclet number.

The large Péclet number case can be recovered asymptotically by expanding the unknowns with respect to the small parameter
$\varepsilon =210^{1/4}/\sqrt {\textit{Pe}}$
. The scalings obtained numerically suggest that we should take
The definition of
$\chi$
induces at dominant order that
$\chi ^2=\varepsilon ^{-4} q^2$
so that
$\chi _0=q_0$
. The dispersion relation (C6) contains two terms which are asymptotically equal to
and
since
$1/[1+\coth (\chi \eta )] = (1/2)(1 - \exp [-2 \chi _{0}\eta _{0}/\varepsilon + \mathcal{O}(1)])\sim 1/2$
. At order
$\varepsilon ^6$
, the dispersion relation is simply
which provides the value of the layer thickness at leading order
At next order (
$\mathcal{O}(\varepsilon ^7$
) terms), the dispersion relation takes the following form:
Replacing
$\eta _{0}$
by
$\zeta ^{-1}$
yields the value of the layer thickness at first order as a function of the wavenumber
Since
$\eta =1/\sqrt {\sigma }$
, maximising
$\sigma$
as a function of
$q$
amounts to minimising
$\eta _{{1}}$
, which is obtained for
The asymptotic value of the most unstable wavenumber is thus
$210^{5/4} \zeta \textit{Pe}^{-5/2}/\sqrt {2}$
, which is plotted as a thin dashed line in figure 13(a). It is in excellent agreement with the numerical treatment of the full equation (C6), plotted with a solid line.
For large Péclet, the growth rate is simply obtained at first order as
$\sigma ^\star =\eta ^{-2}=\zeta ^2 \textit{Pe}^4/210^2$
, which is again in excellent agreement with the numerical value (thin dashed line in figure 13). The asymptotic values for the thickness and the wavenumber
$\chi$
are also very close to their numerical counterparts.

















































