Salt polygons and porous media convection

Jana Lasser,1, 2, ∗ Joanna M. Nield,3 Marcel Ernst,2 Volker Karius,4 Giles F.S.
Wiggs,5 Matthew R. Threadgold,6 Cédric Beaume,6 and Lucas Goehring7, †
1
Graz University of Technology, Institute for Interactive Systems and Data Science, Inffeldgasse 16c, 8010 Graz, Austria
2
Max Planck Institute for Dynamics and Self-Organization Am Fassberg 17, 37077 Göttingen, Germany
3
Geography and Environmental Science, University of Southampton Highfield, Southampton SO17 1BJ, UK
4
Geowissenschaftliches Zentrum, Georg-August-University Goldschmidtstrasse 3, 37077 Göttingen, Germany
5
School of Geography and the Environment, University of Oxford, South Parks Road, Oxford OX1 3QY, UK
6
School of Mathematics, University of Leeds, Leeds, LS2 9JT, UK
7
School of Science and Technology, Nottingham Trent University Nottingham NG11 8NS, UK
(Dated: December 6, 2022)
From fairy circles to patterned ground and columnar joints, natural patterns spontaneously appear
in many complex geophysical settings. Here, we investigate the origins of polygonally patterned
arXiv:1902.03600v4 [nlin.PS] 4 Dec 2022

crusts of salt playa and salt pans. These beautifully regular features, approximately a meter in
diameter, are found worldwide and are fundamentally important to the transport of salt and dust
in arid regions. We show that they are consistent with the surface expression of buoyancy-driven
convection in the porous soil beneath a salt crust. By combining quantitative results from direct field
observations, analogue experiments and numerical simulations, we further determine the conditions
under which salt polygons should form, as well as how their characteristic size emerges.

I. INTRODUCTION

Salt deserts or dry salt lakes are amongst the most

inhospitable places of our planet. Their otherworldly
shapes inspire the imagination (e.g. Star Wars’ desert
planet Crait, or the million tourists annually visiting
Death Valley [1]), and are an important drive of climate (b)
processes [2–8]. Variously referred to as salt pans, playas,
or dry lakes, the immediately prominent feature of such
landscapes (Fig. 1) is a characteristic tiling of polygons,
formed by ridges in the salt-encrusted surface. These pat-
terns are seen around the world [9–16], including Salar de
Uyuni in Chile [9], Chott el Djerid in Tunisia [10], Badwa-
ter Basin in California [11], Dasht-e Kavir in Iran [12] and (a) (c)
Dalangtan Playa in China [13]. Although they have been
argued to share similarities with fracture [12] or buckling FIG. 1. Typical salt polygon patterns at (a), (c) Badwater
[17] patterns, a quantitative and predictive mechanism Basin in Death Valley and (b), Owens Lake, California (image
for the emergence of the patterns remains obscure. (a) courtesy Sarah Marino).
A challenge to any explanation is that the polygo-
nal patterns in salt crusts are remarkably similar wher-
ever they occur, despite local differences in geology, salt
chemistry, and environmental conditions. For example, other environmental processes. They form in areas with
while crusts can vary from sub-centimeter to meters in a negative water balance (i.e. where evaporation signif-
thickness [2, 12, 18] they predominantly express poly- icantly exceeds precipitation) and the dominant water
gons about 1–2 m across [2, 11, 12, 15–20]. The same balance is between a perennial groundwater flow seeping
patterns also appear for pure halite crusts such as those in from distant sources, and the evaporation of this water
of Badwater Basin [11] and pans in Iran [12] as well as from the surface of the dry lake [2, 4, 24]. The groundwa-
the sulfate-rich crusts found for example at Owens Lake ter brings dissolved salts with it, which accumulate at the
(California) [11], Sua Pan [21], the coastal sabkhas of surface in a crust of evaporites. Wind blowing over the
Abu Dhabi [14] and Salar de Uyuni [9]. crust entrains dust, which forms a significant proportion
The salt crusts of dry lakes are known to be dynamic of the global atmospheric dust production [3, 6, 7] and
over months to years [2, 15, 17, 18, 22, 23], and couple to of mineral transport to the oceans [25]. As one example,
dust from the dry Owens Lake has posed health problems
for people living nearby [26, 27], and the site is the center
of a decades-long, intense remediation effort [28]; a pre-
∗ jana.lasser@tugraz.at ferred method is the use of shallow flooding to encourage
† lucas.goehring@ntu.ac.uk formation of a continuous, active surface crust [29, 30].
 2

While playas are inherently associated with dust evaporation salt ridge
sources, these sources are temporally and spatially vari-
able [24, 31–33], and there is a lack of understanding of
the relevant processes of how crusts form, are sustained,
and degrade [7, 34]. The height of polygon ridges affects high salinity
the threshold wind speed for dust entrainment [7, 35] and
complete coverage of a crust reduces the sediment avail- Z low
able for dust emission compared to a patchy crust. In salinity
pore fluid
order to better account for these complexities we need to flows
understand the relationships between surface and sub- X
surface drivers of playa dust sources [7, 31].
FIG. 2. Proposed dynamics of patterned salt crusts. The
Salt crusts also modify evaporation and heat flux from
dominant fluid motions are shown by the black arrows, and
the playa surface [5], and hence the water and energy bal- the water salinity is indicated by the color contours.
ances of these fragile environments. In particular, exper-
imental and field research consistently show that contin-
uous crusts reduce evaporation rates on playa surfaces by to evaluate any model that displays polygonal features.
an order of magnitude compared to climate-derived esti-
Based on the known phenomenology of salt playa, the
mates [8, 29, 35–37]. As a result, the observed groundwa-
most significant tests for a predictive model include: (a)
ter evaporation rates at sites with active and continuous
The driving mechanism should be specific to the geo-
salt crusts tend to lie in the relatively low range of 0.2–0.7
physical conditions encountered in salt lakes [9–12, 14–
mm/day, including seasonal variations [29, 37–41].
16] and (b) should spatially modulate salt transport,
Research on salt pans has typically focused on either leading to increased salt precipitation at ridges in the
the dynamics of their complex subsurface flows [42–46] crust [2, 12, 54]; (c) The response should be fast enough,
or their crusts [15, 17, 18, 35], without considering how and stable enough, to account for ridge growth on a scale
these features interact. Here we show that by treating of months [2, 15, 18, 22]; (d) The model dynamics should
the surface of a salt playa together with the fluid in the generate closed polygonal patterns of relatively narrow
porous media of the soil near its surface, the origins, dy- ridges spaced a few meters apart [11, 12, 15–18]; and
namics, length-scale and shape of the polygonal patterns (e) the characteristics and scale of the patterns should
in salt crusts can be apprehended. As we develop and be robust to environmental variations, including differ-
test such a model of salt crust formation, we will show ent types of soil ranging from silt to sand [11], salt com-
how it makes explicit predictions of crust growth rates, positions ranging from pure halite crusts to those with
the conditions necessary to form ridges and continuous significant amounts of sulphates and carbonates (e.g.
crusts, and other inputs into dust or groundwater mod- trona) [9, 11, 12, 14, 37] and any local differences in cli-
els of these extreme environments. In particular, by com- mate and variables like the evaporation rate [29, 37–39].
bining theoretical analysis, numerical simulations, exper-
Previous models of salt polygons have mainly consid-
iments and field observations, we will demonstrate how
ered mechanical effects. The earliest argument was made
the density-driven convective dynamics of groundwater
by Christiansen in 1963 [17]. He suggested that the
can lead to variations in the salt flux into the crust, with
shapes were the result of compressive stress, due to salt
faster precipitation along the boundaries between convec-
precipitation and temperature changes, which caused the
tive cells facilitating the growth of a polygonal network of
crust to first buckle and then crack as the growing ridges
salt ridges (see Fig. 2). Under measured field conditions
thrust upwards. He noted that the spacing of the buck-
the size of the expected convective cells matches the scale
ling features should be proportional to the thickness of
observed for the surface crust patterns and makes accu-
the stressed crust and also depend on its strength; these
rate predictions about the crust growth rates. Finally, we
predictions are consistent with the modern understand-
will present direct field-based evidence of convective dy-
ing of wrinkling phenomena [55–57]. However, wrinkles
namics that are correlated to the surface crust patterns.
due to isotropic compression typically form parallel fea-
tures, such as stripes, rather than closed shapes [55–57];
polygons can form by compression, but only by folds into
II. PHENOMENOLOGY OF SALT POLYGONS an elastic layer, rather than ridges projecting out of a
layer [57, 58]. A more fracture-focused model was pro-
Any model for the emergence of salt crust patterns posed by Krinsley [12] in his 1970 survey of Iranian salt
should be able to convincingly explain their commonly pans. He argued that desiccation places the crust into
observed features. Polygonal networks are one of the tension, rather than compression, which is relieved by
more common emergent patterns in geophysics, seen fracture. The cracks then allow for groundwater to more
for columnar jointing [47], ice-wedge polygons [48, 49], easily reach the surface to evaporate, so that ridges grow
polygonal terrain [50] some types of mud-cracks [51] as up along them. As the crust develops faster at the ridges,
well as convection [52, 53] for example. As such, a suite of these can buckle or thrust upwards (as in Fig. 1(b)). Like
more detailed or quantitative predictions are important wrinkling, for cracks the expected feature size is a few
 3

times the thickness of the cracking layer [59, 60]. precipitation of salt templates the growth of ridges at the
Both the wrinkling [2, 54] and fracture [18–20, 61] mod- locations of the downwellings. At upwelling plumes the
els can explain the qualitative appearance of cracks and larger salinity gradients strengthen the diffusive trans-
thrusting at the ridges in salt crusts. The main difficulty port away from the surface, rather than advective trans-
with any purely mechanical model, however, is length- port towards it, and crust growth is slower.
scale selection. The natural spacing between buckles or Our main field site, Owens Lake (Fig. 1(b)), has a
cracks is proportional to the thickness of the layer under typical crust pattern in a well-studied and controlled
stress [55–60]. As some geophysical examples, the scale landscape, and can introduce our modeling assumptions.
of ice-wedge polygons in permafrost is determined by the This dry, terminal saline lake has an aquifer extending
depth at which annual temperature changes diffuse into from the near-surface to over 150 m depth [69]. The
the ground [48, 62], whereas the size of columnar joints groundwater carries dissolved salts [27, 37, 70], which
in lava records how it cooled, and in particular the thick- collect in an evaporite pan of about 300 km2 [3, 69] and
ness of a cooling front invading into an initially molten which are particularly concentrated in the fluid within
lava formation [47]. For salt playa, crust thicknesses vary a meter or so of the surface [11, 37]. Average annual
widely, so cannot explain the robust polygonal scale seen groundwater evaporation rates of E = 0.4 ± 0.1 mm/day
in nature, and no other convincing explanation for the have been measured in areas with active salt crusts [37].
feature spacing has been proposed. The near-surface soil is a silt to fine sand with high poros-
As an important mechanism one may also consider ity of ϕ = 0.7 [37]. A tortuosity-corrected diffusion con-
the crust composition itself, which can include miner- stant, D = 1.00 ± 0.24 × 10−9 m2 s−1 is representative of
als in different hydration states. For example, thenardite the mix of dissolved salts present at the site (mainly NaCl
(Na2 SO4 ) transforms into mirabilite (Na2 SO4 · 10H2 O), and Na2 SO4 ). Efforts to control dust emission from the
and expands its volume by ∼320%, at temperatures eas- lake involve shallow flooding [30], vegetation [71], gravel
ily seen in salt deserts [63]. While phase changes can un- cover and encouraging crust growth [29]. As shown in
doubtably generate stresses, the difficulty of length-scale supplemental movie S1, after a shallow-flooding event the
selection remains, given similar patterns in crusts of dif- soil is saturated with water, which evaporates and leaves
ferent thickness. In addition, salt polygons are routinely behind salts that crystallize into a continuous crust, cov-
seen in crusts with almost pure halite (NaCl) composi- ered by a network of slowly growing ridges. Given the
tion [11, 12], where such effects are absent. high quality of the observational record at Owens Lake,
and in order to make quantitative predictions relevant to
the remediation of this important site, we will use repre-
III. BUOYANCY-DRIVEN CONVECTION AS A sentative values from it throughout this section.
MECHANISM FOR PATTERN FORMATION

As discussed, explanations for the formation of salt A. Governing equations

polygons have so far focused on mechanical effects. Crust
growth is implied in these cases, but the salt source The transport of fluid beneath a dry salt lake can be
feeding this growth is not modeled explicitly. In other treated as a Darcy flow in a porous medium, with evapo-
contexts there is a well-developed literature regarding ration occurring at the surface and being fed from be-
buoyancy-driven convection in porous media, including low by water with some background salinity (see e.g.
salt lakes [44, 45, 64], sabkhas [46, 65], carbon seques- [42, 44, 64, 72]). Such a system will naturally develop
tration applications [52, 66, 67] and thermally-driven a salinity gradient below the surface, which can become
cases [42, 53, 68]; the model we develop here builds on unstable to convective overturning [42, 44, 64].
a recent study of the linear stability and approach to a As a prototype case we consider a large, flat salt lake
statistically steady state of convection in a dry salt lake with an average surface evaporation rate E, and where
[64]. However, the effects of such flows on a salt crust the groundwater is recharged from some deep, distant
have not yet been explored. Here, we will examine the reservoir. For the subsurface flows we model the volu-
implications of convection in the soil beneath a salt crust metric flux, or superficial fluid velocity, q, of a fluid of
and its predictions for the emergence of salt polygons. pressure p moving through a porous medium of constant
We consider a coupling of surface salt patterns to sub- uniform porosity ϕ and permeability κ. By mass conser-
surface flows, as visualized in Fig. 2. Briefly, surface evap- vation E is also the average of the upward component of
oration will leave the near-surface groundwater enriched the fluid flux, qz , at any depth as well as the recharge rate
in salt, and heavier than the fluid beneath it. This can from the reservoir. The fluid has a viscosity µ and carries
lead to precipitation of salt at the surface as well as con- dissolved salt, whose diffusivity D can be corrected for
vection in the soil, with narrow, regularly spaced down- the presence of different ions as well as tortuosity [73, 74].
wellings of high salinity [64]. We will show that convec- The dissolved salt contributes to the density of the
tion should result in a higher salt flux into the surface fluid, ρ, and hence to buoyancy forces. In salt lakes, ther-
above the downwellings, where salinity gradients in the mal buoyancy is several orders of magnitude weaker than
groundwater are weaker. We argue that this preferential solutal buoyancy, and can be neglected [44, 64]. Using
 4

the Boussinesq approximation, ρ = ρ0 +S∆ρ, where ρ0 is weight, VB = κ∆ρg/µ, to the upward drift due to surface
the density of the reservoir fluid (including its dissolved evaporation: Ra = VB /E. Thus, it is frequently used to
salts, see Section III C), and ρ1 = ρ0 + ∆ρ is the density characterize the vigor of porous medium convection with
of a saturated solution. The relative salinity S mediates a through-flow [42–44, 64, 72].
between these limits: S = 0 corresponds to the back- A more detailed derivation of this model, along with
ground salinity of fluid entering from a distant source discussion of the assumptions behind it, a linear stability
(z → −∞), whereas S = 1 represents a salt-saturated analysis of the onset of convection and a two-dimensional
brine in contact with the crust (at z = 0). numerical investigation into its statistically steady state,
The governing equations are then a continuity equa- was recently published by some of us [64]. Our aim here
tion for incompressible fluid flow, an advection-diffusion is to show how solutal convection beneath a salt lake
equation for the relative salinity, and Darcy’s law: will couple to the growth of patterns in the crust and to
make predictions which can be compared to field mea-
∇ · q = 0, (1) surements. As such, after a discussion of how the emer-
∂S gent dynamics of this model will appear in geophysically
ϕ = ϕD∇2 S − q · ∇S, (2)
∂t relevant conditions, we will develop predictions of the salt
κ flux into the crust, which can give rise to crust patterns.
q = − (∇p + ρg ẑ) , (3)
µ
where g is the acceleration due to gravity, and ẑ is
an upward-pointing unit vector. These mass and mo- B. Convective dynamics and scale selection
mentum balances can describe a variety of porous me-
dia flows, including solutal flows in playas and sabkhas The model above is designed to match the expected
[44, 45, 64, 72, 75], CO2 -rich flows in carbon sequestra- dynamics of groundwater in a dry salt lake with at least
tion applications [67, 76] and, with an appropriate trans- a thin salt crust developing on top of a fluid-saturated
formation of variables, flows driven by thermal buoy- soil. For the one-dimensional case, where salinity de-
ancy [42, 68, 77]. A review of these phenomena was re- pends only on depth, there is a unique stationary solu-
cently given in Ref. [78]. As boundary conditions for the tion to Eqs. (4)–(6), namely S = eZ = exp(zE/ϕD),
salt lake scenario we take: which represents a salt-rich layer of groundwater lying
just below the surface and a balance between advective
S = 1, qz = E on z = 0, and diffusive transport of salt [42]. In the absence of any
S → 0, qz → E as z → −∞. horizontal flows other plausible initial conditions, such as
an initially uniform lake (i.e. S = 0 everywhere), will re-
In order to non-dimensionalize the governing equa- lax towards the stationary solution on the characteristic
tions, one needs to specify a characteristic length scale L timescale T [44, 64]. Similarly, an initially crust-free lake
and velocity scale V, with a natural time scale following can be described at first by an unsaturated boundary at
as T = ϕL/V. Taking advantage of how the evaporation Z = 0, with no flux into the surface until the salinity
rate sets the average vertical fluid flux everywhere, we there reaches saturation: this scenario will also develop
use V = E. Then, in the absence of any geometric length a boundary layer that reaches saturation, and starts to
scale in the problem (e.g. the layer thickness often used precipitate, over the same timescale [44]. For represen-
for two-sided convection [78]), we set L = ϕD/E as the tative parameter values from Owens Lake, T is about
distance over which advective and diffusive effects will 6 months, consistent with the known timescale of crust
balance, for fluids moving at the characteristic speed E. development [15], and L = 15 ± 8 cm.
Rescaling Eqs. (1)–(3) then leads to For a wide variety of initial conditions, therefore, a dry
salt lake will naturally develop a heavy salt-rich bound-
∇ · U = 0, (4) ary layer of fluid near its surface. Above some critical
∂S Rayleigh number Rac this layer will be unstable to con-
= ∇2 S − U · ∇S, (5) vection [42–44, 64, 72]. Linear stability analysis shows
∂τ
that for constant, uniform evaporation at the surface,
U = −∇P − RaS Ẑ, (6)
the stationary solution has a finite-wavelength instability
with a rescaled velocity U = q/E, depth Z = zE/ϕD, above Rac ' 14.35 that leads to the growth of convec-
time τ = tE 2 /ϕ2 D and pressure P = ϕµD κ
(p + ρ0 gz). tive plumes [42, 43, 64]. At Rac the critical wavenumber
This system of equations is governed by a single dimen- ac ' 0.76 for conditions at Owens Lake, corresponding
sionless group, the Rayleigh number: to a wavelength of 1.3 m. Intriguingly, this characteristic
scale of the convective instability is of the same order of
κ∆ρg magnitude as the patterns typically seen in salt crusts.
Ra = . (7) This instability is also robust to different assumptions
µE
about the initial and boundary conditions of the lake. For
Here, the Rayleigh number also gives the ratio of the example, an initially uniform lake (where S = 0 at t = 0)
speed at which a large heavy plume will fall under its own has the same Rac as the stationary state, along with a
 5

very similar spectrum of unstable modes [64]. Alterna- which extend much deeper than the boundary layer.
tively, as would be appropriate for a salt lake with a thin Over long enough times a statistically steady state devel-
ponding of surface brine, the case of a constant pressure ops [64] where the rates of plume formation and merging
boundary condition at Z = 0 is unstable to convection balance, and a well-defined length-scale emerges.
above Rac ' 6.95, with ac ' 0.43. These dynamics can be compared with analogous re-
To predict the long-time dynamics of convection pat- sults from two-sided convection without throughflow [52,
tern in a dry salt lake, we simulated Eqs. (4)–(6) on rect- 53, 68, 79], which has become an important model of
angular domains (2D) or boxes (3D), where S = UZ = 1 carbon geosequestration. For this, some caution is re-
at the top boundary, S = 0 and UZ = 1 at the bottom, quired, as the characteristic scales and Rayleigh num-
and with periodic boundary conditions along the vertical ber are defined differently in the two cases. Compar-
walls. We conducted simulations from Ra = 20 to 1000. ing the two systems, the patterns near the walls are re-
As initial conditions, we started from the stationary state markably similar, with transient ribs or sheets appear-
for a bounded domain, S = (eZ − e−h )/(1 − e−h ), with ing between larger, more stable and regular polygonal
small random perturbations applied to the salinity distri- cells [52, 53, 68]. Interestingly, however, the wavelength
bution at all grid points. Most simulation domains had selection process is quite different. For two-sided con-
a horizontal extent of 12π (for 2D) or 24π (for 3D) and vection the dominant circulation cells, sampled at the
depth of h = 10. The domain depth introduces an ad- midplane, scale as a ∼ Ra0.5 . Near the walls the density
ditional length scale, and thus dimensionless group, into of the smaller, transient features increases more rapidly
the numerical problem. To confirm that the results are with Ra, with a dependence thought to approach a ∼ Ra
in a limit where they are essentially independent of h, we at high enough Ra [52, 53, 78]. However, averaging out
repeated selected simulations for domains with h = 20. these fast-moving features over time, or through spa-
Our numerical methods are summarized in Appendix A. tial filtering, shows a near-surface pattern that closely
A finite-wavelength instability sets in for all simula- matches the internal convection cells [53].
tions and rapidly grows beyond the linear regime. As For our one-sided salt lake model, quantitative details
shown for the 2D case in Fig. 3(a–c), perturbations of the scale selection process are presented in Fig. 4,
around the most unstable mode develop into a set of nar- which also includes the neutral stability curve and most
row plumes of high salinity, which fall from the upper sur- unstable mode of convection, as derived in Ref. [64]. Sim-
face. In 3D the situation is similar, as in Fig. 3(d,e), with ulations with Ra > Rac are unstable to convective over-
the convection near the surface instead organizing into a turning, which becomes more vigorous with increasing
closed polygonal network of downwelling sheets, which Ra. To begin with, pattern wavenumbers were calculated
surround patches of upwelling fluid. At lower depths the as in [68], as the first moments of the radially-averaged
sheets break up into isolated plumes. Both the growth power spectra of the salinity distributions at a depth of
rate of the linear instability and the speed of the ma- Z ≈ 1. The initial instability was characterized at the
ture downwelling features scale with  = (Ra−Rac )/Rac , rescaled time τb = 1. At these early times the spacing of
which can be understood as Ra ∼ VB . As such, we make downwellings in both the 2D and 3D simulations (red
use of a rescaled time, τb = τ , in order to compare fea- markers) agrees well with the most unstable mode of
tures at a similar stage of development. the linear stability analysis. When measured at τb = 30,
As they grow, the initial downwelling features strip however, many downwelling plumes have merged, lead-
the boundary layer of dissolved salt, leaving it depleted ing to smaller wavenumbers (blue markers). Confirm-
of the buoyancy forces that initiated the convection. The ing that the system is in a statistically steady state by
plumes also tend to merge together into larger structures. this time, simulations run to τb = 60 show no measur-
This coarsening process causes the pattern length-scale to able systematic drift in k (values match with τb = 30 to
increase over time; for further details, see Ref. [64]. How- ∆k = −0.004 ± 0.01, mean ± standard error). Similarly,
ever, as the distance between downwelling areas grows, additional simulations with h = 20 gave comparable re-
they become less efficient at draining salinity from the sults (values match with h = 10 to ∆k = −0.04 ± 0.04).
boundary layer, which can start to thicken again. This In the steady state, there is a gradual increase in
allows for the growth of additional instabilities. In 2D the average wavenumber with Ra, as is the case also
these take the form of small plumes that appear between, for 2-sided convection, although this trend is system-
and are attracted to, the larger established downwellings. atically weaker than the scalings evident in that sys-
In 3D, although new point-like plumes do also occur, the tem [52, 53, 68]. This increase appears to be driven
initiation of linear features is more typical. As shown by the proliferation of small, transient downwelling mi-
in Fig. 3(d,e), these are weaker downwelling sheets that croplumes as Ra increases. These intermittently appear,
connect two sides of polygonal convection cells. As they then are swept into the more established large-scale con-
grow, they also move towards the boundaries of the more vective features (see supplemental movies S2, S3). As the
established, larger convection cells, eventually merging surface crust is built up over time, we would expect it to
with an existing edge. These transient ‘microplume’ fea- reflect the more stable features, i.e. megaplumes. To
tures become more prevalent at higher Ra, overlaid on isolate the size of these features, we calculated the domi-
the more stable ‘megaplumes’ of convection cells [53], nant wavenumber, ā, associated with the maximum of the
 6

FIG. 3. The dynamics of porous media convection simulated below a salt lake are shown for Ra = 100. For two-dimensional
simulations, the development of high-salinity (S) downwelling plumes is shown for times (a) τ = 0.1, (b) τ = 0.3 and (c)
τ = 1.0. Supplemental movie S2 shows the full dynamics of the simulation. Panel (d) shows a snapshot of three-dimensional
salinity flux into the surface (Z = 0) for τ = 1.0. Panel (e) shows the same snapshot, now including a view of the downwellings
of high salinity along the vertical faces of the domain. Here, the downwellings arrange into narrow sheets, forming a network
around more diffuse upwelling regions. Supplemental movie S3 shows the full dynamics of the three-dimensional simulation.

power spectrum. As shown in Fig. 4 (turquoise markers), of Eq. (5) can be written in terms of a salinity flux
this predicts a pattern of polygonal cells, bounded by
salt-rich downwellings, and a characteristic wavelength JS = U S − ∇S (8)
that consistently settles to values near to Λ = 2π/ac , or
about 1.3 m for the conditions typical of Owens Lake. where ∂S/∂τ = −∇ · JS . At the surface, Z = 0, the
upwards water flux matches the evaporation rate, so
UZ = 1. The presence of a salt crust also sets S = 1
there. As such, the salinity flux into the crust is given by
C. Salt flux and ridge growth

∂S 
JZ = 1 − . (9)
At the surface of the soil water will evaporate and leave ∂Z Z=0
behind its dissolved burden of salt. In our model of this
process the rate at which salt is added to the surface This flux vanishes, JZ = 0, in the stationary solution
depends on its advection along with the water, but is S = exp(Z), although this scenario will still leave a con-
also influenced by diffusion down concentration gradients stant upward flux of salt into the crust at whatever con-
away from the crust. In other words, the mass balance centration is supplied by the reservoir, as we will describe
 7

τ , Z) = 2π/a
wavelength Λ(b salinity flux into the crust vanishes. This follows from
16 8 4 2 1 0.7 the divergence theorem. For a deep lake the descending
103 plumes will mix diffusively with upwelling fluid over a
timescale ∼ Λ2 , for plume spacing Λ. Since the plumes
Rayleigh number Ra

start their fall from the surface at a characteristic veloc-

ac
ity VB ∼ Ra [64], this mixing will occur over depths of
order RaΛ2 . As they mix and lose buoyancy the plumes
102 will slow, setting a maximum depth that they can pen-
etrate, of the same order. Below this depth the average
upwards salinity flux of the flows from the reservoir must
hai (b
τ = 1) (2d) hai (b
τ = 1) (3d)
hai (b
τ = 30) (2d) hai (b
τ = 30) (3d) approach zero. Since in a statistically steady state the
ā (2d) total salinity in the convecting part of the domain is con-
101 stant, or fluctuating around a well-defined average value,
0 2 4 6 8 10
wavenumber a(b
τ , Z) = 2π/Λ the salinity flux into the surface must also approach zero.
Probing this limit accurately, however, requires domain
sizes well beyond the scope of our current simulations.
FIG. 4. Pattern wavenumbers a, as measured in two-
dimensional (circles) and three-dimensional (squares) simu-
To make predictions of the crust dynamics at any par-
lations for different Ra. Average wavenumbers hai are mea- ticular field site, the salinity flux JZ needs to be con-
sured at two times: once early into the growth of the initial verted into estimates of the real salt flux into the crust.
instability, at τb = 1 (red), and later at τb = 30 (blue), after Assuming minimal volume change due to mixing, the
the pattern has had time to develop towards a statistically density of the groundwater can be written as ρ = ρw +βc,
steady state. We also report the wavenumber of the dominant where ρw is the density of pure water, c is the mass con-
mode ā (turquoise), at late times. For the 2D simulations, the centration of dissolved salts in the solution, and β is a
marker positions and error bars represent means and standard coefficient of expansion. Analogous to the salinity flux in
deviations as measured in ensembles of 6–10 realizations each. Eq. (8) the mass flux of salt, in dimensional variables, is

Jc = qc − Dϕ∇c, (10)
shortly. Variations in the near-surface salinity gradient
∆ρ
will then modulate crust growth around this reference = qc0 + EJS , (11)
case. Where vertical gradients are strong crust growth β
will be suppressed (JZ < 0), whereas over weaker gradi- where the second equation follows by changing variables
ents crust growth will be enhanced (JZ > 0). using ρ = ρ0 + S∆ρ. As can be seen, there is an off-
To demonstrate this modulation quantitatively, we cal- set between the two fluxes–representing the fact that
culated the surface salinity flux according to Eq. (9) the groundwater feeding into the system from a distant
in simulations at Ra = 100, with results for 2D shown source (S = 0) contains some initial level of dissolved
in Fig. 5, and for 3D shown in Fig. 3(d, e). Supple- salt, c0 . In dimensional terms, the predicted salt mass
mental movies S2 and S3 show the full dynamics of the flux into the crust is now
two- and three-dimensional systems, respectively. Above 
downwellings plumes (2D) or sheets (3D), the salinity ∂c 
Jc,z = Ecsat − Dϕ . (12)
flux into the surface is positive and shows marked peaks, ∂z z=0
whereas between downwellings the salinity flux into the
surface is negative, and roughly constant. As sketched in We will apply this to Owens Lake in Section IV C.
Fig. 2, this predicts the development of narrow regions
of faster salt precipitation above any convective down-
wellings, which we argue gives rise to ridges there. D. Model summary & predictions
As visible in the movies S2 and S3, the positions of
downwelling plumes slowly fluctuate as the simulations We have developed here a model of porous medium
progress, as does the pattern of flux into the crust. The convection beneath a salt lake, which can accurately pre-
motion of the more well-established plumes is gradual, dict much of the known phenomena of the crust pat-
especially in 3D, but the presence of large ridges at some terns in dry salt lakes, as summarized in Section II. The
field sites suggests that feedback mechanisms may be driving mechanism of salinity-driven convection and its
needed to help stabilize the patterns over long periods associated governing equations are drawn from existing
of time. Such feedback could act through a spatial mod- models of salt lake dynamics [44, 45, 64] and observations
ulation of the evaporation rate E based on salt coverage, of convection beneath tidal sabkhas [46, 65], but we ex-
which then acts to pin the position of the downwellings tend these to consider interactions with the surface crust.
in space. Reference [64] gives some discussion on how Evaporation at the surface causes a vertical salinity gra-
such a feedback mechanism could work. dient to emerge, which leads to salt precipitation and
We also note that the dynamics in a sufficiently deep crust growth. If a critical Rayleigh number is surpassed,
lake should approach the case where the average surface convective cells emerge and modulate the salinity flux
 8

0
JZ

-1

-2
0
Z

10 L
-5

FIG. 5. Salinity flux (JZ = 1 − ∂S/∂Z) into the modeled surface crust. The simulation snapshot is for Ra = 100 at time
τb = 150. Areas predicted to experience enhanced (JZ > 0) or suppressed (JZ < 0) salt flux are indicated as shaded red and
blue areas, respectively. Peaks, defined as local maxima with a flux value above zero, are indicated as dashed lines and lie
directly above downwelling plumes of high salinity; the lower panel gives the salinity S in the domain at the same time.

into the surface, with faster precipitation expected over in turn, focusing on empirical evidence from field work,
downwellings and slower crust growth over upwellings. supplemented by analogue experiments. Details of field
We will provide field observations in Section IV to quan- data collection are given in Appendix B, data processing
tify these predictions, and show that they are consistent in Appendices C–H and experimental methods in Ap-
with average growth rates on the order of millimeters per pendix I. The results point towards convective overturn-
month, and about a two-fold faster deposition of salt at ing as a suitable driving mechanism of salt crust patterns.
downwellings, relative to upwelling regions. Observing
the pattern structure in 3D, we found that over suffi-
cient time the downwelling features arrange into narrow A. Convective instability
sheets that delimit a closed polygonal pattern of convec-
tion cells. The scale of these polygons, in a steady state,
A driving mechanism should be specific to the geo-
does not have any strong dependence on the Rayleigh
physical conditions encountered in salt lakes. As our
number. In particular, from Ra = 20 to 1000 we showed
model is based on boundary conditions and governing
that the average wavenumber of the near-surface con-
equations of porous media transport that are appropri-
vection pattern changes by at most a factor of four; the
ate for playa and sabkhas (e.g. [44, 45, 64, 72, 75]), we
dominant wavenumber, which reflects the structure of
aim to prove here that typical dry salt lakes have con-
the more stable, deeper convection cells, does not evolve
ditions that would make them unstable to the onset of
significantly over this range and remains consistent with
convective motion. To this end, we visited and charac-
the critical wavenumber, ac = 0.76. This predicts that
terized field sites at Owens Lake (CA, USA), Badwater
the pattern wavelength will be relatively insensitive to
Basin (CA, USA) and Sua Pan (Botswana), to show that
soil composition (which determines e.g. permeability),
they can be described by a Rayleigh number above the
groundwater chemistry (e.g. sulphate rich, or pure halite,
critical value of Rac ' 14.35 [43, 64, 72]. All sites show
affecting ∆ρ and to a lesser extent D) and robust to fluc-
well-developed polygonal patterns, and details of the field
tuations in environmental driving parameters.
methodology are given in Appendix B. These results are
supplemented by experiments that visualize buoyancy-
driven convective dynamics in conditions similar to those
IV. EVIDENCE FOR CONVECTION AS of real salt lakes (see Supplemental movie S4).
DRIVING MECHANISM To determine the Rayleigh number appropriate for
each field site, we measured the relevant parameters
In Section II we outlined a set of criteria against which of Eq. (7). From grain-size distributions (data de-
to test mechanisms of pattern formation in salt playa. posited at [80], analysis in Appendix C) we calculated
Briefly, any mechanism should be active in salt lake set- ds , the Sauter diameter [81], of near-surface soil sam-
tings, modulate salt transport by directing more salt to- ples. The results, from 4.3 ± 0.6 µm to 138 ± 23 µm,
wards ridges, account for the timescales of crust growth, represent a silt to fine sand. A high soil porosity,
explain the consistent wavelength of closed polygonal ϕ = 0.70 ± 0.02, was previously measured at Owens
shapes in the crust and be robust enough to explain Lake [37]. From these, the relative permeability was esti-
all this across the variety of conditions associated with mated using the empirical relationship κ = 0.11 ϕ5.6 d2s ,
patterned crusts. Here, we will explore these criteria which fits a broad set of experimental and simulation
 9

data [82]. Across all sites κ = (3 ± 2) × 10−13 m2 to

(2.7 ± 1.2) × 10−10 m2 . At Owens Lake we also directly
measured density differences of ∆ρ = 210 ± 10 kg m−3
between pore water samples taken from close to the
surface and at approximately 1 m depth (data at [83]).
Annual average evaporation rates of groundwater are
taken as E = 0.4 ± 0.1 mm/day [29, 37] for Owens
Lake, 0.3 ± 0.1 mm/day [38] for Badwater Basin, and
E = 0.7 ± 0.5 mm/day for Sua pan [35, 39], with fur- 10 cm
ther details reviewed in Appendix D. These low levels
of groundwater evaporation are characteristic of salt pan 0.4 0.6 0.8 1.0
environments, and similar rates are seen in active playa salinity
of the Atacama desert (E = 0.5 ± 0.1 mm/day [40]) and
sabkha near Abu Dhabi (E = 0.2 ± 0.05 mm/day [41]).
Finally, we assume the dynamic viscosity of the ground- FIG. 6. Convective plumes in an experimental Hele-Shaw
water to be a constant µ = 10−3 Pa s. A detailed de- cell are highlighted by dye (the brighter upwelling fluid results
from dying the reservoir, well after convection has set in). The
scription of the data sets from Owens Lake and Badwater
motion is coupled to the salinity, which was measured at the
Basin is in Ref. [11]. colored squares by destructive sampling.
From these observations we calculated Ra at twenty-
one sites around Owens Lake, five in Badwater Basin
and seven at Sua Pan. The median values at these was seen in all experiments above Rac , but not in the
three regions were Ra = 3700, 36000 and 420, respec- finest-grained case, where Ra  Rac . These results are
tively. Values for all 33 sample locations were between consistent with those of Wooding et al. [84], who exper-
Ra = 120 ± 50 and (1.2 ± 0.5) × 105 , well above Rac . imentally confirmed theoretical predictions of the onset
The Sauter diameters, permeabilities and Rayleigh num- of convection for the slightly different case of a salt lake
bers for all sites are given in the Appendix, Table IV. with surface ponding, or constant pressure boundary con-
The conditions throughout these patterned salt playa are, ditions. Finally, we note that the range over which the
therefore, suitable to expect a convective overturning of experiments show convection, between Ra = 20 ± 10 and
their groundwater, with plumes of high salinity sinking (1.3 ± 0.3) × 103 , also coincides with much of the range
downwards from the surface. of Rayleigh numbers we estimated from the field sites.
It is interesting to note that convective plumes of salt-
rich water have also been observed by electrical resistiv-
ity measurements after a heavy rainfall on salt crusts B. Mapping salt heterogeneity
near Abu Dhabi [46] and beneath wind-tidal flats in
Texas [65]. While these are a slightly different phe- If salt polygon growth is driven by convective dynam-
nomenon, involving the sudden addition of salt-rich brine ics happening beneath the surface, then horizontal dif-
to the surface as the rain dissolves salt, they demonstrate ferences in salt concentration should be detectable in soil
convection in similar geophysical conditions to salt lakes. and pore fluid under typical field conditions, and also
The Rayleigh numbers calculated for these two cases are in laboratory-based analogue experiments. Salt varia-
up to about 40,000 and 90, respectively [46, 65]. tions of this nature have been seen in sabkhas after rain
We complemented these observations with analogue [46, 65]. We provide evidence here of salt-rich plumes be-
experiments in Hele-Shaw cells to demonstrate buoyancy- neath the crust of Owens Lake, and show that the plumes
driven convection in porous media under conditions com- are correlated with the positions of the ridges in the crust.
parable to the field. These experiments are inspired by First, to demonstrate the coupling of salt concentra-
Refs. [84–86] but instead of using a narrow and empty tion and convective motion under controlled conditions,
gap between the cell walls to create the porous medium, we sampled one of the Hele-Shaw experiments introduced
we used a 0.8 cm thick cell filled with glass beads. We ap- in Section IV A. For this we dissected an experiment that
plied a strong surface evaporation, driven by heat lamps was undergoing convection, and which had been evolving
and ventilation, and the base of the cells were connected under constant conditions for two months. Samples were
to fluid reservoirs with a fixed salt concentration of 50 extracted from locations along the downwelling and up-
kg m−3 . Further details of the experimental methods are welling plumes, as indicated by the motion of dye added
given in Appendix I and an example time-lapse movie into the reservoir fluid a few days before sampling. As
of an experiment in progress is given as supplemental shown in Fig. 6, the fluid flow in the analogue experi-
movie S4. To observe whether a convective instability ments is clearly driven by, and coupled to, variations in
arises for different conditions, we systematically varied salinity. Pore water samples taken from the upwelling
the grain sizes of the glass beads. This modifies the rela- regions (dyed yellow in the figure) show systematically
tive permeability of the system and allowed us to change lower salinity than those taken from downwellings.
the Rayleigh number of the experiments. Convection Next, from the field we collected samples of wet soil
 10

(a) (b) (c)

25

depth / cm
50
fit
2m 60 cm data 75

25

50
fit
2m 60 cm data 75
5 10 15 20 25
0 5 10 15 20 25 5 10 15 20
salt concentration / wt.%
elevation / cm salt concentration / wt.%

FIG. 7. Surface height maps and subsurface salinity profiles for two sites (site 1, top; site 2, bottom) at Owens Lake. (a)
TLS scans give the elevation of ridges on the surface crust. (b) Cross-sections of polygons at Owens Lake show the variation
of salt concentration with depth and relative to the ridge positions. Each square corresponds to the salt concentration in the
pore fluid of an individual field sample. (c) Exponential fits (see Appendix H) to the changing salt concentration with depth.
Red dots represent horizontally averaged salt concentrations from data in (b), and error bars show the standard deviation.

from fresh trenches dug at two unmanaged sites at Owens representative values for the boundary layer thickness
Lake. Surveys of the surface relief were made before from exponential fits to the horizontally-averaged salt
sampling, using a terrestrial laser scanner (TLS, see Ap- concentrations at both trench sites (see Appendix H).
pendix E and Ref. [22] for methodology; data deposited Interestingly, the observed values of 13.5 ± 5.3 cm and
at [87]). These show the presence of salt polygons of 17.7 ± 1.5 cm are comparable to the natural length scale
about 2 m in size delimited by high ridges, as in Fig. 7(a). of L = ϕD/E = 15.1 ± 8.0 cm estimated for Owens Lake.
In each case we sampled soil along a grid pattern in Thus, not only does direct field sampling of groundwa-
a cross-section below a polygon, including beneath two ter beneath a patterned salt crust show both horizontal
bounding ridges. Analysis of the salt concentration of and vertical variations in salt concentration, which sup-
the samples with respect to pore water content shows di- port the hypothesis that the system is unstable and con-
rect evidence of plumes of high-salinity fluid below the vecting, but it also demonstrates that the plumes of high
salt ridges (Fig. 7(b), methods in Appendix F and data salinity are co-localized with the surface ridges.
deposited at [88]). Specifically, we tested whether the
distribution of salt concentrations in an area below each
ridge was different to that below the flat pan of the poly- C. Surface salt flux in the field
gon; testing this hypothesis (two-sample KS test), shows
that the distributions below ridges and flat crust are sta-
tistically distinct (p < 0.02), at both sites. A driving mechanism for pattern formation should be
able to spatially modulate salt transport to the soil sur-
The measurements of the salt concentration in the face, leading to salt ridge growth on a scale of months.
pore water also shows an exponential decay in salinity We demonstrate this growth at Owens Lake through a
with depth (Fig. 7(c)), consistent with a salt-rich bound- time-lapse video (Supplemental movie S1) which shows
ary layer that is heavy enough to drive convection in the surface crust dynamics over a period of approxi-
a porous medium. As evidence for this, we recovered mately four months in spring 2018. At Sua Pan, crust
 11

dynamics have been measured by TLS scanning of sev- site area Jc,z 10−7 [kg m−2 s−1 ] r [mm/month]
eral sites over the course of several months [15]. Active 1 ridge 1 5.6 ± 2.0 (3.9) 1.5 ± 0.6 (1.1)
crusts there showed average growth rates of 0.7 to 1.5 1 ridge 2 6.8 ± 1.2 (3.2) 1.8 ± 0.4 (0.9)
1 center 3.3 ± 2.5 (4.9) 0.9 ± 0.7 (1.4)
mm/month. The fastest short-term growth rates of ridge
2 ridge 1 5.3 ± 1.3 (3.8) 1.4 ± 0.4 (1.1)
features, briefly reaching up to 10 mm/month, were asso- 2 ridge 2 6.8 ± 1.1 (3.1) 1.8 ± 0.3 (0.9)
ciated with ridge thrusting. In Section III C we explored 2 center 2.1 ± 0.9 (4.8) 0.6 ± 0.3 (1.3)
the corresponding predictions of our model. Specifically,
Eq. (12) describes the expected salt flux into the sur- TABLE I. Salt flux into the surface Jc,z and crust growth rate
face, Jc,z , depending on the salt concentration gradient r inferred at the polygon ridges and centers for the two sites
there. Here, we apply the framework of this model to the at Owens Lake depicted in Fig. 7. The error bounds consider
maps of the subsurface salt concentration measured at independent sources of uncertainty, while the values in brack-
Owens Lake (Fig. 7). In particular, we will show that the ets give ranges taking into account all sources of uncertainty
observed variations are sufficient to drive average crust (i.e. including lake-wide uncertainties in E, D and ϕ).
growth rates on the order of millimeters per month along
with a preferential growth of the ridges, over the rest of
the crust. the crust growth rates at polygon ridges when compared
To calculate the expected salt flux into the surface, we to polygon centers. As salt is being added faster to the
estimated the near-surface gradient in salt concentration ridge areas, we would expect differential growth, focused
from the field data from our trench sites. Dividing this on the ridges. Although the crust morphology can be-
data between regions beneath the ridges and centers of come quite complex [12, 19, 89], the faster expansion of
each polygon, we used an exponential fit to recover the the crust in these locations can plausibly contribute to
effective thickness of the salt-rich boundary layer L0 , the differential strains, and features like ridge thrusting or
background salt concentration cbkg and the salt concen- local buckling. This is consistent with the faster dynam-
tration at the surface csat , for each such region. For this, ics seen at ridges in Sua Pan [15] and that we report at
the measured weight fractions of salt were re-scaled by Owens Lake (see supplemental movie S1).
the densities of salt solutions, to give concentrations c Over long times the average crust growth rate will be
in terms of the mass of dissolved salt per unit volume essentially constrained by the evaporation rate and the
of fluid. Further details of the fitting are given in Ap- amount of salt in the groundwater feeding into the lake.
pendix H, including Figs. 11 and 12. However, it would not be surprising if these rates varied
Substituting the exponential form of the fit into with time, e.g seasonally, or via extreme events like sur-
Eq. (12) yields a prediction for the mass flux of salt, in face flooding. As we have not accounted for the excess
terms of measurable quantities, of mass of the hydrated states of any of the crust materials
(especially mirabilite, present as a significant minority
csat − cbkg species of salt at Owens Lake and Sua Pan, although
Jc,z = Ecsat − ϕD . (13)
L0 not noticeably present at other pans, such as Badwater
Basin), the rates we report in Table I also represent a
Values for L0 and cbkg below each ridge and polygon cen- lower bound on the anticipated crust growth rates.
ter are extracted from the fits. For csat , the fitted values
We have shown here how the heterogeneous salinity
are all consistent with each other, so a common value
distribution measured below salt polygons in the field
of a saturated solution, csat = 316 kg m−3 , was used
infers variations in the salt flux to the surface, if the sub-
to minimize uncertainties. As in Section IV A we use
surface fluids are undergoing porous medium convection.
values for the evaporation rate, E = 4.6 × 10−9 m s−1 ,
Within the framework of our convection model, the mea-
and porosity, ϕ = 0.7, taken from previous observa-
sured pattern of salt concentration near the surface of
tions at Owens Lake [29, 37]. For an effective diffu-
the crust imply about a twofold increase in crust growth
sion constant, we use the tortuosity-corrected diffusivity
rate at the ridges, rather than the centers of salt poly-
D = 1.00 × 10−9 m2 s−1 (see Appendix G). Finally, we
gons. The crust growth rates on the order of millimeters
measured the density of salt crust samples collected from
per month are also consistent with the time scale of crust
Owens Lake to be ρcrust = (960 ± 50) kg m−3 . This value
growth observed in the field, which show that noticeable
is used to convert mass flux rates into crust growth rates.
changes can occur over months.
The resulting salt flux and growth rates for the ridges
and centers of the two trench sites from Fig. 7 are shown
in Table I. The derived salt flux into the surface is about
twice as high at the ridges, as compared to the centers of D. Pattern length scale
polygons. In order to allow a clear statistical comparison
between these rates, only independent sources of errors If groundwater convection leads to preferential loca-
were propagated (i.e. excluding uncertainties in E, D tions for salt precipitation, and from thence to salt crust
and ϕ, as these will affect all rates in the same way) to patterning, then the convective cells and crust polygons
give the main uncertainties reported in Table I. Within should have similar length-scales and patterns. We pre-
these uncertainties, we see a consistent difference between sented model predictions for the length-scale selection of
 12

convection in Section III B. To test these predictions in at τb = 30. These covered 11–18% of the surface; nar-
the field, we measured the surface relief of the crusts rower features were seen at higher Ra, but the range of
at multiple sites from all three dry lakes using a terres- Ra explored is small. Although the salt ridges can be
trial laser scanner (see Appendix B). The crusts display relatively complex, involving breaking and thrusting for
the typical polygonal patterns of salt lakes, as shown in example, the convective model qualitatively captures the
Fig. 8(a–d), with further data in [11, 15]. From each essential asymmetry between the narrow ridge features
scan we extracted a characteristic wavelength for the pat- and the flat pans between them.
tern, based on the dominant frequency of the power spec- In summary, the pattern wavelength we measured for
trum, using the methods reported in [15]. Full results salt polygons at a range of sites across three different salt
are given in the Appendix, Table IV. In all cases the lakes is thus found to be consistent with the length scale
pattern wavelengths ranged from λ = 0.53 ± 0.20 m to expected for buoyancy driven convection going on in a
3.02 ± 1.40 m. These were converted into dimensionless statistically steady state. Simulations of the dynamics in
wavenumbers a = 2πL/λ, using the characteristic length- a three-dimensional domain (see Fig. 3(d)) also show that
scale L = ϕD/E as calculated for each lake (E from the convective mechanism can produce patterns of closed
Section IV A, ϕ and D from Section IV C). The results polygonal shapes at this lengthscale. These patterns are
are summarized in Fig. 8(e), which shows the Rayleigh qualitatively similar to the salt polygons observed in the
number and wavenumber for each site studied. field (Fig. 8(a–d)) with relatively narrow features delim-
We found that the wavenumbers characterizing the iting the larger polygonal shapes.
surface crust patterns in the field are clustered near
the critical wavenumber for convection, ac = 0.76, and
largely independent of Ra. In Section III B, we predicted E. Environmental variation & remaining challenges
how the statistically steady state of subsurface convec-
tion should look. In this state the average wavenumber To this point, we have focused on the effects of vary-
increased slowly with Ra, by about a factor of four from ing environmental conditions as they contribute to the
Ra = 20 to 1000. This was attributed to the proliferation Rayleigh number and natural lengthscale of the convec-
of small, transient features, superimposed on more sta- tion problem. We showed that a wide range of measured
ble convection cells. The dominant wavenumber, given field conditions should lead to convection, and to convec-
by the peak of the power spectrum, showed little-to-no tion cells with a size and pattern that is consistent with
dependence on Ra, and was argued to better represent the observed surface crust patterns; faster crust growth,
the long-term behavior of the convection cells. As shown and thicker accumulations in ridges, was shown to corre-
in Fig. 8(e), although there is large scatter in the data, late with the downwellings of the subsurface convection.
the wavelengths of the crust pattern broadly match the In this section, we will briefly look at and discuss other
wavelengths of the dominant features of the sub-surface factors, including the salt chemistry, crust thickness and
convection expected in a statistically steady state. features like cracks and crust wrinkling. We will also dis-
Furthermore, for the field sites there should be ample cuss some of the remaining assumptions and challenges
time for the convective dynamics to develop towards the of the convective model, in explaining crust patterning.
steady-state condition. In the model, the timescale most In addition to the soil, we characterized the composi-
appropriate to characterize the dynamics of convection tion of the salt crust at Owens Lake and Badwater Basin
is T /, where T = ϕ2 D/E 2 describes how the bound- (see Appendix G for methods; data deposited at [83]).
ary layer would develop in the absence of convection, Both sites are predominantly halite, although Owens
and  = (Ra − Rac )/Rac accounts for the increased vigor Lake additionally has significant concentrations of car-
of convective motion as the driving forces increase (see bonates and sulfates [37, 83]. Between all the field sites
also [64]). For the relevant parameter values of Owens we also observed salt crusts with thicknesses from below
Lake, T / is just over one day. Alternatively put, one a centimeter to approximately 30 cm. Across these dif-
year of evolution at Owens Lake, for a site with a repre- ferences in soil and salt crust composition there is a sim-
sentative value of Ra = 3700, would represent a rescaled ilar appearance of the salt polygons. This agrees with
time of τb = 350. This is long compared to the timescale a mechanism like subsurface convection, which is rela-
of τb = 30 that we used to characterize the steady-state tively insensitive to salt chemistry. It would, conversely,
conditions in Figs. 4 and 8. present difficulties to any mechanism that was reliant on
Finally, as shown in Fig. 8(a–d), polygon ridges are specific salts, such as mirabilite, or which varied strongly
relatively narrow features. To characterize their size, with crust thickness.
we thresholded our surface relief maps using Otsu’s Other surface features, including cracks and wrinkling
method [90]. The relative area occupied by ridges, mea- or buckling, are also associated with salt crusts [12, 17].
sured in this way, was 24 ± 8% (mean ± standard devi- While studying Owens Lake, we observed well-ordered
ation); this may be a slight overestimate, due to occlu- hexagonal desiccation cracks in crust-free mud with ap-
sion of the TLS behind larger ridges. In the 3D simu- proximately 10 cm spacing, as shown in Fig. 9(a). These
lations the comparable features are downwellings, which appeared at the crust-free edge of a pan that was oth-
we characterized as areas of positive surface salinity flux erwise covered in salt ridges (with a spacing of 1–2 m).
 13

(a) (b) (e)

105

(c)
(d)

Rayleigh number Ra
104
2m 2m (a)

(c) (d) 103

(b)
neutral stability
Owens Lake
Badwater Basin 102
Sua Pan
simulation: ā
2m 2m simulation: hai
101
0 1 2 3 4
0.00 0.05 0.10 0.15 0.20 wavenumber a = 2π/Λ
elevation [m]

FIG. 8. Field observations of the length-scale selection of salt polygons. Panels (a) to (d) show the surface relief recorded
at several field sites. Panel (e) maps data from the field onto the stability diagram of porous media convection in a salt pan.
The neutral stability curve (black line) is the theoretical boundary above which an evaporating stratified pore fluid is unstable
to perturbations of wavenumber a. Various triangles show measurements at Owens Lake, Badwater Basin and Sua Pan, with
letters indicating data from the sites in (a–d). Turquoise and blue dots show the average and dominant wavenumber measured
in simulations, respectively (data from Fig. 4).

The hexagonal crack patterns can be explained as the our model. For example, dissolution will be favored at
result of the intermittent and episodic cracking of the the base of the crust, in contact with groundwater, and
soil, as water levels fluctuate [51]. However, they form precipitation at the top, where evaporation occurs. Sim-
at a very different scale to the salt ridges, and the cracks ilar observations have been made in playa [89] and in
develop as localized depressions in the ground, rather lab [91, 92] studies of efflorescence with cultural heritage
than upthrust ridges. Further, despite their presence applications. Dissolution-precipitation dynamics raises
near the salt crust, there were no signs of preferential porosity in crusts, leading to a ‘fluffy’ texture, trapped
precipitation of salt in the cracks. We also observed sur- air pockets, self-lifting effects [89, 91, 92], and a lower
face buckling of the salt crust at Owens Lake, as shown threshold for dust entrainment [24]. We accounted for
in Fig. 9(b) (see also supplementaal video S1 on dates the high crust porosity while calculating growth rates in
05/07 through 05/11). These features appeared when the section IV C, but a more nuanced approach may give fur-
crust was solidifying, and may result from the stresses ther insight, including into dust models.
induced by phase changes or the addition of salt mass
into the crust. However, the scale of their features is, There is also likely to be an element of positive feed-
again, much smaller than those of the dominant polyg- back between the growth of the salt crust, and the dy-
onal ridges in the crust, and the patterns formed are namics of the convective plumes. Our simulations showed
also different. Finally, we note that throughout our core that, without such a feedback, the positions of down-
samples and trench sites, as in Fig. 9(c), the soil horizons welling features slowly fluctuate in space, although this
remained normal and undisturbed by the presence of sur- tendency is weaker in 3D than in the 2D simulations.
face ridges. This would not be the case if deep fractures A feedback cycle, where evaporation rates are affected
were occurring along with the ridge formation (as in e.g. locally by the presence of the ridges, can act to stabi-
ice-wedge formation in permafrost terrain [48]). lize the locations of the plumes. This has been demon-
strated [64] by simulations in which downwellings tend to
More generally, we have treated the salt crust as two- move towards, and stay at, locations with lower evapora-
dimensional, of nominal thickness. This is a necessary tion rates. We have also measured temperature and rela-
simplification, given the novelty of the hypothesis ad- tive humidity records from Owens Lake [64], which show
vanced here. However, there will be some internal dy- that evaporation rates should be slower at the ridges,
namics within the salt crust itself, beyond the scope of compared to the centers of polygons, in a way consistent
 14

with such a feedback mechanism. Further model devel- (a) (b)

opment along these lines would be promising.
Incorporating feedback into crust evaporation may also
help refine predictions of wavelength selection. The scale
of salt polygons matches the dominant wavelength of
our simulations, which we argued to be representative
of a long-time integration of the more stable convection
pattern. The effect of the smaller-scale, transient ‘mi- 10 cm 10 cm
croplumes’ remains uncertain. As captured by the aver-
age wavenumber, these are more prevalent at higher Ra,
but their faster dynamics (changing visibly over simu- (c)
lation times of ∼ 0.1, or about two weeks in real terms)
15 cm
mean that they should only weakly contribute to the sur- polygon ridge
face pattern, which develops over months. This predic-
tion could be checked in simulations that modulate sur-
face evaporation, based on tracking of the total amount
of salt accumulating at the surface over time.
As a final point, one may also consider thermal con-
tributions to buoyancy, and phenomena like double-
diffusive convection. A typical 10◦ C diurnal varia-
tion [64] will change water density by about ∆ρT = 1 kg
m−3 , compared to the ∆ρS ' 200 kg m−3 density con-
FIG. 9. Additional surface and subsurface features at Owens
trast due to solutes at Owens Lake. The buoyancy ratio,
Lake. (a) Small-scale hexagonal mud cracks could be seen in
N = ∆ρS /∆ρT ' 200, means that the driving forces and the soil about 10 m away from one of our field sites. (b)
speeds of thermally-driven flows will be reduced by a fac- Wrinkles were also observed in the salt crust adjacent to one
tor of 1/N , compared to solute-driven flows (see e.g. [93] of the polygons that served as a field site. (c) At a trench
for review of double-diffusive flow in porous media). Such site the soil layering was exposed. The surface position of
effects can, therefore, be assumed to be negligible. a polygon ridge is indicated by the white arrow, and holes
mark the sampling locations. The horizons in the soil beneath
the polygon are visible and undisturbed (see for example the
V. DISCUSSION AND CONCLUSION reddish layer near the third sampling hole from the top).

Salt deserts, playa and pans are a common landform

important for climate balances such as dust, energy and Our findings help to elucidate the relationships be-
water, and express a rich repertoire of patterns and dy- tween subsurface drivers and the development of surface
namics. Here we have shown that, in order to model and crust patterns. Better knowledge of crust development
understand the surface expression of such deserts, insight can improve dust emission models [7, 24, 31], including
is gained by considering the crust together with the sub- quantifying the connection of salt pan hydrology to con-
surface dynamics. In particular, we have shown how the trols on dust emission [33] such as those used at Owens
emergence of regular salt polygons, which are a common Lake [29, 30]. As further context, dusts and aerosols re-
salt crust pattern, can result from their coupling to a con- main one of the larger uncertainties in modelling climate
vection process in the soil beneath them. The existence of sensitivity [94], with e.g. feedback from soil biocrust de-
salinity-driven convection in salt pans is itself already an gredation recently predicted to contribute half as much
anticipated finding (see e.g. [44]). Furthermore, a salinity as direct anthropogenic aerosol emissions by 2070 [95].
gradient below the surface, strong enough to drive such While the surface variability of crusted surfaces can be
convection, has been measured at Owen’s Lake [37] and probed using remote sensing [23], our model is a first
strong horizontal salinity gradients indicative of convec- step to improving predictions of how subsurface hydro-
tive motion have been shown to exist in sabkhas [46, 65]. logic variability links to surface crust variability. This
What we have demonstrated is that this convection pro- knowledge is increasingly important as saline lakes like
cess develops into a statistically steady state character- the Dead Sea or Great Salt Lake are shrinking [96] and
ized by large-scale convection cells with a wavelength of their newly exposed beaches are candidate areas for salt
a few meters, after a few weeks of evolution under typ- crust formation and dust emission.
ical field conditions. We further showed how this can To establish the connection between surface features
explain the remarkably consistent appearance, growth and subsurface flows, we demonstrated consistent results
rates, polygonal shape, narrow ridges, and representa- from theoretical and numerical modeling, analogue ex-
tive wavelengths of the salt crust patterns observed in periments and field studies. In contrast to previous ex-
different parts of the world, including our field sites at planations of crust patterns [12, 17], this model is able
Sua Pan, Badwater Basin and Owens Lake. to explain the robustness of the pattern length scale
 15

by considering the statistically steady state of porous ME developed the theory of convective dynamics and
media convection, based only on measured environmen- programmed the 2D simulation.
tal parameters. In fully three-dimensional simulations, VK performed the X-ray analysis, supervised JL for
the convective dynamics were also shown to give rise to the grain size measurements and provided feedback for
closed-form polygonal shapes. More importantly, it is the soil composition and salt species analysis.
able to predict a suite of quantitative details, such as GFSW participated in fieldwork at Sua Pan.
the timescales required for convection to set in, the rates MRT performed numerical experiments and analyzed
of crust and ridge growth, the positions of the ridges the data from the 3D simulations.
above salt-rich plumes, and the relative narrowness of CB supervised MRT, programmed the 3D simulation
the ridge features. At the downwellings the salinity is and provided feedback for the theoretical model.
higher and therefore the salinity gradient between the LG conceptualized the study, developed the theory of
crust and the underlying fluid is weaker (compare sketch convective dynamics, supervised JL and ME, performed
of Fig. 2 to measurements in Fig. 7). As salt transport the water density measurements and participated in field-
is a balance of advective and diffusive processes, this will work at Owens Lake and Badwater Basin.
lead to an increased rate of salt precipitation above down-
welling features, contributing to the growth of ridges at
the boundaries of convection cells. After the initial emer- ACKNOWLEDGMENTS
gence of ridges, the growth process might be bolstered by
feedback mechanisms such as a modulation of the evap- We thank Grace Holder (Great Basin Unified Air Pol-
oration rate by the presence of ridges, cracks or surface lution Control District) for support at Owens Lake, the
wicking phenomena. U.S. National Park Service for access to Death Valley
As such, our results show how salt polygons are part (Permit DEVA-2016-SCI-0034) and Antoine Fourrière
of a growing list of geophysical phenomena, such as fairy and Birte Thiede for their work on preliminary exper-
circles [97], ice wedges [49], polygonal terrain [50] and iments on convection. TLS processing used the Iridis
columnar joints [51], which can be successfully explained Southampton Computing Facility. MRT was supported
as the result of the instability of a dynamical process. by the Leeds-York-Hull Natural Environment Research
Council (NERC) Doctoral Training Partnership (DTP)
Panorama under grant NE/S007458/1. Sua Pan work
VI. DATA AND CODE AVAILABILITY was funded by Natural Environment Research Coun-
cil (NE/H021841/1), World University Network and
Field site labels and locations as well as site-by-site Southampton SIRDF (Strategic Interdisciplinary Re-
data for Fig. 8(e) are given in the appendix Table IV. search Development Funds) and enabled by Botswana
Full data sets for the soil particle sizes [80], surface height Ministry of Environment, Wildlife, and Tourism (Permit
profiles [87], salt chemistry [83], subsurface salinity dis- EWT 8/36/4 XIV) and Botswana Ash (Pty) Ltd.
tribution and pore water density [88] as well as images
describing the US field sites [98] are available under a
CC-BY-SA 4.0 license at PANGAEA. Appendix A: Numerical simulation
Code for the 2D simulation is available under an MIT
license on GitHub, doi:10.5281/zenodo.3969492. The three-dimensional (3D) simulations use a pseudo-
All scripts, data sets and images used to produce the spectral element method where the horizontal directions,
figures in this work are available at https://doi.org/ over which the numerical domain is periodic, are treated
10.17605/OSF.IO/KJDWF. using fast Fourier transforms, and the vertical (bounded)
direction is decomposed into elements, each of which is
treated using Gauss–Lobatto–Legendre quadratures (see
e.g. [99]). The time-derivative is approximated using a
VII. AUTHOR CONTRIBUTIONS
second-order backward differentiation formula and the
nonlinear term extrapolated using a Taylor expansion of
JL, ME and LG wrote the original draft of the second order. The coefficients of the resulting schemes
manuscript; all authors reviewed the manuscript. can be found in [100]. All simulations were carried out
JL participated in fieldwork at Owens Lake and Bad- on a domain of depth 10 and horizontal area of 24π×24π.
water basin, analyzed the soil salinity and grain size dis- To compute wavenumbers using two-dimensional sim-
tributions, performed numerical experiments of the 2D ulations (cf. Fig. 4), we used the same numerical solver
simulation, performed the Hele-Shaw experiments, ana- and imposed the solution invariance in one of the horizon-
lyzed all data except for the 3D simulations and TLS tal directions. For the two-dimensional (2D) simulations
measurements and created the data visualizations. of wavenumber selection (Fig. 4), we used the same nu-
JMN participated in fieldwork at Owens Lake, Badwa- merical solver and imposed the solution invariance in one
ter Basin and Sua Pan and performed and analyzed the of the horizontal directions. These simulations have hor-
TLS measurements. izontal domain size 12π and depth 10 or 20. Pattern
 16

wavenumbers were calculated from the salinity at the Owens River

gridpoint closest to Z = 1 as in 3D, but without radially- trench
averaging the power spectra. The dominant wavenum-
ber was identified as the mode with the highest power core T32-1-L1

spectral density. All other two-dimensional (2D) simula-

tions were carried out using a stream-function-vorticity T29-3

approach based on Refs. [101, 102] with a detailed imple- Keeler

T36-3 T27-A
mentation described in Ref. [64].

m
ai
N T27-S

n
Owens Lake

gr
T27-3
(dry)

av
W E

el
Appendix B: Field data collection

ro
ad
S unmanaged
Field work was conducted at Owens Lake and Bad- area
managed
water Basin in November 2016 and January 2018; see
Ca area
e.g. [103, 104] for geological descriptions of the area. The lif
orn 2 km
Owens Lake sampling sites are indicated in Fig. 10. At San
ia T16
Francisco
Badwater Basin five sites were visited ∼500 m south of Los T10-3
the main tourist entrance to the playa. For both loca- Angeles
T8-W
tions field methodology and data collection are fully doc-
umented in a separate data description publication [11]. T2-5
At Owens Lake we use field site labels referring to T2-4
surface management cells of the dust control project
there [105]. These labels link either to managed cells
or to unmanaged areas in the direct vicinity of a man-
aged cell. Labels typically start with “TX-Y”, where X
is a number and Y is a number or letter. The first num-
ber refers to the water outlet taps along the main water FIG. 10. Schematic map of Owens Lake (CA, USA) with
pipeline that crosses the lakebed south to north and is locations of soil core (red) and trench (orange) sites marked.
used to irrigate management cells. Numbers/letters after Labels give the names of local dust management areas.
the hyphen refer to subregions branching from the same
tap. For some sites we investigated more than one poly-
gon. This is indicated with brackets, e.g. T27-A (3) is which were sealed with parafilm.
the third polygon investigated at site T27-A, which cor- To evaluate the density difference, ∆ρ, we collected
responds to the Addition (A) region of the management pore water samples at Owens Lake from eight sites, in-
cell next to the 27th tap. Bracketed numbers are used for cluding liquid taken from directly below the surface (in
the sites visited at Badwater Basin. cases where the water table was at or within 10 cm of the
To evaluate the profile of salt concentration below surface), and at a depth of about 1 m.
polygons, at most sites soil cores (4 cm Dutch gouge Samples from Sua Pan were collected during a different
auger) were taken to a depth of up to 1 m. The soil field campaign, reported in Ref. [15]; site labelling follows
showed normal bedding (see also Fig. 9(c)), indicative that publication. Sediment samples at Sua Pan were col-
of sedimentation following flooding. Samples were col- lected 2 cm below the crust in August 2012. These were
lected from each visible soil horizon, or with a vertical double bagged and subjected to grain size analysis only.
resolution of ∆z =10–15 cm. At all three dry lakes, surface height maps were col-
Trenches were also dug at sites T27-S (Site 1 in Fig- lected with a Leica Terrestrial Laser Scanner (TLS); a
ures 7 and 11) and T32-1-L1 (3) (Site 2), in order to P20 model was used at Owens Lake and Badwater, and
take samples along cross-sections below salt polygons. a Scanstation at Sua Pan. The scanner head was posi-
The trenches were dug about 200 cm in length, 40 cm in tioned at least 2 m above the playa surface and scans were
width and down to a water table of ∼70 cm. Soil samples performed before the surface was disturbed by sampling.
were taken from a freshly cleaned trench wall in a grid
GPS locations of all sites are provided in Tab. IV.
pattern with spacings of ∆x = 15 cm and ∆z = 10 cm.
An example of such a trench along with sampling loca-
tions is shown in Fig. 9(c). The samples had an aver-
age volume of approximately 10 ml and were taken using Appendix C: Soil characterization
a metal spatula, which was cleaned with distilled water
and dried before each use. The samples were a mixture of Soil samples from Owens Lake, Badwater Basin and
soil with a grain size of medium sand to clay, water and Sua Pan were analyzed to determine their distribution
salt (both dissolved and precipitated). After collection, of grain sizes. This analysis was performed using the
samples were immediately stored in air-tight containers, soil remaining after a sample’s salinity was determined
 17

(Appendix F). Soil grain size distributions were measured Appendix F: Salinity and density measurements
by laser particle sizer (Coulter LS 13 320), from which the
Sauter diameter (the mean diameter, respecting the soil’s Owens Lake and Badwater Basin soil samples were an-
specific surface area [81]) was calculated. For each site alyzed to determine the amount of salt in their pore fluid.
a representative dS is calculated as the average Sauter These samples had been sealed immediately after collec-
diameter of all soil samples (for trenches, one sample per tion, to preserve their water content. After unsealing,
depth) from that site. For Sua Pan, only samples from each sample was first transferred to a crystallizing dish
sites B7 and L5 were available; dS for the other five sites and weighed, to give a combined mass of sand, salt and
is estimated as the mean of the measured values at these water. It was then dried at 80◦ C until all moisture had
two sites. All grain size data is deposited at Ref. [80]. visibly vanished, or for at least 24 h, and re-weighed to
Soil porosity has previously been measured to be determine the mass of the (evaporated) water. Next, it
around ϕ ≈ 0.70 ± 0.02 [37] at Owens Lake. Because was washed with 50 ml of deionized water to dissolve any
of lack of similar measurements at Badwater Basin and salt, allowed to sediment for 24 hours, and the super-
Sua Pan, we used the value measured at Owens for calcu- natant liquid was collected in another crystallizing dish.
lations of κ and Ra at these sites. For each site a perme- After two such washings the remaining soil and the re-
ability is then calculated based on the Sauter diameter covered salt solution were dried and weighed. Measure-
and porosity, as κ = 0.11ϕ5.6 d2S [82]. ment uncertainty is based on the difference between the
The Sauter diameter dS , permeability κ and Rayleigh initial sample mass and the sum of the separated wa-
number Ra for all sites investigated are provided in Ta- ter (mw ), salt (ms ) and soil masses. This gave a di-
ble IV. Uncertainty ranges for Ra were calculated as sys- rect measure of the mass fraction of salt in solution as
tematic errors based on standard errors of all the input C = ms /(mw + ms ), reported as a weight % (wt.%).
environmental parameters. Owens Lake pore water samples were analyzed to de-
termine their density using a vibrating-tube densitome-
ter (Anton Paar DMA4500). The near-surface (0–10
cm depth) groundwater density was consistently 1255 ±
8 kg m−3 , while water from 70–100 cm depth had a den-
Appendix D: Evaporation rate sity of 1050 ± 2 kg m−3 . These values are broadly consis-
tent with chloride concentration profiles previously mea-
Groundwater evaporation rates are taken from the lit- sured at Owens Lake [37]. We note that thermal effects
erature. At Owens Lake, Tyler et al. [37] combined on the groundwater density will be comparatively neg-
lysimeter, Bowen ratio and eddy correlation methods, ligible, as the mean annual variation in temperature at
with repeat seasonal measurements conducted well after Owens Lake will allow for a density change of, at most,
any rainfall. The value of E = 0.4 ± 0.1 mm/day cap- 5 kg m−3 . Similarly, the solubility of halite in water
tures the range of measurements at their NFIP site, near would change by less than 5 kg m−3 , seasonally. Data
our northern sites at the lake. At Badwater Basin, De- from salt concentration and pore water density measure-
Meo et al. [38] measured evaporation by Bowen ratio and ments are deposited at Ref. [88].
eddy correlation methods, with continuous monitoring
over several years. The value of E = 0.3 ± 0.1 mm/day is
taken from their two playa sites south of Furnace Creek, Appendix G: Salt chemistry and diffusivity
and the 2001 conditions most consistent with the wet-
ter season preceding our field work. At Sua Pan we use Pore water samples from selected sites at Owens Lake
E = 0.7 ± 0.5 mm/day, estimated by remote sensing and and Badwater Basin were analyzed to determine the
energy balances [39], and wind tunnel experiments [35]. dominant salt species present. This analysis was per-
formed on the dried salts remaining after the salinity
measurements (Appendix F). Mineral identification was
performed by quantitative X-ray powder diffraction anal-
ysis (Philips X’Pert MPD PW 3040). Samples from
Appendix E: TLS data processing
Owens Lake showed a mixture of salts with 53 ± 7 wt.%
sodium chloride and 30 ± 5 wt.% hydrated sodium sul-
TLS Scan data was processed following Ref. [22]. Data phate (mirabilite). Other minerals, such as natrite,
were first gridded into a digital elevation map (DEM) sylvite and burkeite, were variously present at less than
with a lateral resolution of 1 cm and a vertical resolution 10 wt.%, each. As such, in solution Na+ and Cl− ions
of 0.3 mm. Dominant frequencies of surface roughness, predominate (>70% by mass), with SO2− 2−
4 , CO3 and K
+

which we use here for the pattern wavenumber a (and ions present in descending order of significance. All data
wavelength λ = 2π/a as reported in Table IV), were then are deposited at Ref. [83].
quantified using the 90th percentiles determined with the Based on the ionic species found in the pore wa-
zero-upcrossing method from the DEMs [15]. TLS data ter, we estimate an average aqueous diffusivity of
from the US field sites are deposited at Ref. [87]. D∗ = 1.37 × 10−9 m2 s−1 from measurements of ternary
 18

site area L0 [cm] Cbkg [wt.%] Csat [wt.%] ridge 1 center ridge 2
site 1 ridge 1 15.1 ± 3.3 11.4 ± 0.8 25.8 ± 2.1
site 1 ridge 2 17.8 ± 2.6 11.1 ± 0.6 25.8 ± 1.5
site 1 center 12.1 ± 2.6 11.3 ± 0.7 26.2 ± 2.0
site 2 ridge 1 20.3 ± 2.6 4.6 ± 0.9 26.6 ± 2.0
site 2 ridge 2 25.3 ± 3.4 3.6 ± 1.1 26.1 ± 2.2
site 2 center 14.5 ± 1.0 5.7 ± 0.4 25.8 ± 0.9

TABLE II. Boundary layer thickness L0 , along with the salt

content (wt.%) of background (Cbkg ) and saturated (Csat )
pore fluid, from exponential fits to the salt concentration dis-
ridge 1 center ridge 2
tributions below the center and ridge areas of two trench sites.

mixtures of the two primary salts [106], using a weighted

average of the mole ratios of their main-term diffusion co-
efficients. Accounting for the tortuosity, θ, of the porous
medium, we then calculate an effective diffusion coeffi-
cient D = D∗ /θ2 = (1.00 ± 0.24) × 10−9 m2 s−1 following
[74], where we estimate θ2 = 1 − ln(ϕ), as in [73].

FIG. 11. Division of data from the two trench sites into two
Appendix H: Fitting the salt-rich boundary layer
ridge areas and one center area each. The horizontal sample
spacing is 15 cm. White areas indicate missing samples.
We used the distributions of salt at the two trench sites
to estimate the near-surface gradients in salt concentra-
tion, and from this the surface salt flux, Jc,z . For this, iments. The base of each cell was connected to a reser-
we divide the grid of measurements at each site into ridge voir containing a 50 kg m−3 solution of NaCl, such that
and center areas–as shown in Fig. 11. Ridge areas cover ∆ρ = 162 kg m−3 (compared to a saturated salt solution
about 40 cm on either side of a polygon ridge. The centers [107]). This reservoir maintained a fluid-saturated pore
include samples from the areas between the ridges, and space between the beads. Evaporation at the top of the
were approximately 70 cm wide. An exponential
0
of the cells was controlled and enhanced by overhead heating
form C(z) = Cbkg + (Csat − Cbkg )ez/L was then fit to and air circulation, and could be controlled from E ≈
the horizontally-averaged salt concentrations (C, wt.%) 1 mm/day to 10 cm/day. Assuming a kinematic viscos-
of each respective area, using an error-weighted fitting of ity of µ = 10−3 Pa s, these conditions allowed for experi-
the data. Here, L0 is the effective thickness of the salt- ments from Ra = (7 ± 2) × 10−3 to Ra = (1.3 ± 0.3) × 103 .
rich boundary layer, Cbkg is a background salt content of The experiments with Ra < Rac did not show any con-
the pore fluid at depth and Csat is the salt content of fluid vective dynamics over a period of three months, while all
at the surface. The fits are demonstrated in Fig. 12 and experiments with Ra > Rac showed convection.
output values for the fit parameters are given in Table II, Visualization of the convective dynamics in the cells
with associated errors. As a consistency check, we note (see Supplemental movie S4) was accomplished by inter-
that all fitted Csat values agree with the mass fraction mittently injecting 2–4 ml of dyed saline solution through
of a saturated NaCl solution (26.3 wt.% at 10◦ C [107]), a thin tube embedded in the cell. The dye was then ad-
which is by far the most predominant dissolved salt in vected by the flows inside the cell over time. Dye move-
the pore water at Owens Lake (see Appendix G). Follow- ment was recorded using time-lapse photography with a
ing fitting the salt concentrations, in terms of a weight digital SLR camera (Nikon D5000 series).
per unit volume, were calculated as c = ρC, assuming To determine the experimental concentration profile
densities ρ based on NaCl solutions at 10◦ C [108]. shown in Fig. 6, one Hele-Shaw experiment at Ra =
120 ± 13 was destructively sampled. After two months
of operation, the reservoir fluid was first dyed with
Appendix I: Experimental methods rhodamine and then fluorescein, to visualize the down-
welling (dark, rhodamine) and upwelling (light, fluores-
Experiments were performed in 40 × 20 × 0.8 cm cein) plumes. Once the dynamics of the plumes became
(width×height×gap spacing) Hele-Shaw cells. As sum- apparent, the wet bead packing in the cell was removed
marized in Table III, these were filled with glass beads in layers, while sampling every 2 cm in depth along the
(Sigmund Lindner GmbH) with diameters of 0–20 µm to centers of the plumes. The resulting ∼1 ml samples were
200–300 µm, dS = 1.9 ± 0.2 µm to 264 ± 25 µm, and analyzed for their salt concentration using the same pro-
ϕ = 0.37 × 10−2 to ϕ = 0.21 × 10−2 . The permeability tocol as described for the field samples in Appendix F.
of the bead packs was evaluated by flow-through exper-
 19

site 1: T27-S (1) site 2: T32-1-L1 (3)

ridge 1 ridge 1
25.0 25
center center
22.5 ridge 2 ridge 2
20
20.0
C [wt. %]

C [wt. %]
17.5 15

15.0
10
12.5

10.0 5

0 20 40 60 80 0 20 40 60 80
depth [cm] depth [cm]

FIG. 12. Horizontally-averaged salt concentration, C(z), at the ridge (blue) and center regions (red) of the two trench sites
(dots) along with exponential fits (lines). Error bars indicate standard deviations of the concentrations, along with a sample
position uncertainty of ±2 cm.

grain size ds [µm] κ [m2 ] ϕ E [ms−1 ] Ra convect

0 – 20 1.9 ± 0.2 (1.5 ± 0.3) × 10−15 0.37 (3.6 ± 1.0) × 10−7 0.007 ± 0.002 no
70 – 110 86 ± 9 (0.8 ± 0.1) × 10−11 0.32 (6.1 ± 2.4) × 10−7 20 ± 9 yes
90 – 150 123 ± 12 (1.4 ± 0.1) × 10−11 0.30 (3.6 ± 1.0) × 10−7 64 ± 19 yes
100 – 200 150 ± 15 (1.7 ± 0.1) × 10−11 0.28 (3.6 ± 1.0) × 10−7 74 ± 22 yes
150 – 250 214 ± 20 (4.7 ± 0.2) × 10−11 0.24 (3.6 ± 1.0) × 10−7 207 ± 59 yes
200 – 300 264 ± 25 (8.3 ± 0.3) × 10−11 0.21 (1.0 ± 0.3) × 10−7 1298 ± 300 yes

TABLE III. Details of experimental tests of the onset of convection, including: grain sizes (from manufacturer) of the glass
beads used, along with the Sauter diameter ds , permeability κ, porosity ϕ, evaporation rage E and Rayleigh number Ra.
 20

Site dS [µm] κ × 10−11 [m2 ] Ra [RaL ; RaU ] λ [m] Latitude Longitude Year
Death Valley
Badwater Basin (1) 59 ± 15 4.99 ± 2.95 28643 [8374; 93961] 1.42 ± 0.58 36◦ 13.6510 -116◦ 46.7230 2016
Badwater Basin (2) 66 ± 12 6.24 ± 2.91 35793 [12736; 103415] 1.42 ± 0.58 36◦ 13.6740 -116◦ 46.7350 2016
Badwater Basin (3) 67 ± 9 6.52 ± 2.56 37444 [15096; 98635] 1.27 ± 0.55 36◦ 13.6650 -116◦ 46.8200 2016
Badwater Basin (4) 72 ± 3 7.39 ± 2.29 42419 [20975; 93764] 0.58 ± 0.32 36◦ 13.6600 -116◦ 46.9030 2016
Badwater Basin (5) 50 ± 15 3.10 ± 2.35 21064 [5692; 72316] 0.55 ± 0.28 36◦ 13.6540 -116◦ 47.0360 2016
Owens Lake
T10-3 4.3 ± 0.6 0.03 ± 0.02 117 [48; 286] 1.79 ± 0.86 36◦ 23.1470 -117◦ 56.7720 2018
T16 6.8 ± 1.6 0.07 ± 0.04 294 [99; 828] 1.19 ± 0.51 36◦ 23.9530 -117◦ 56.4540 2018
T2-4 29 ± 14 1.23 ± 1.12 5457 [918; 21627] 1.13 ± 0.54 36◦ 20.8030 -117◦ 58.6420 2016
T2-5 (1) 24 ± 4 0.81 ± 0.35 3594 [1443; 8953] 1.04 ± 0.41 36◦ 21.0550 -117◦ 58.8240 2016
T2-5 (2) 20 ± 3 0.55 ± 0.22 2436 [1052; 5742] 0.94 ± 0.50 36◦ 20.8950 -117◦ 58.7400 2016
T2-5 (3) 18 ± 3 0.46 ± 0.18 2033 [886; 4762] 1.62 ± 0.65 36◦ 20.8770 -117◦ 58.711 2018
T25-3 (1) 11 ± 2 0.17 ± 0.08 771 [300; 1967] 2.25 ± 0.89 36◦ 27.0390 -117◦ 54.5100 2018
T25-3 (2) 24 ± 4 0.83 ± 0.36 3697 [1484; 9211] 1.18 ± 0.56 36◦ 28.3830 -117◦ 54.9570 2018
T27-A (1) 31 ± 8 1.35 ± 0.81 5994 [1843; 17825] 1.70 ± 0.65 36◦ 29.3020 -117◦ 55.9530 2016
T27-A (2) 33 ± 4 1.55 ± 0.56 6916 [3149; 15614] 2.72 ± 0.98 36◦ 29.0610 -117◦ 55.6020 2016
T27-A (3) 22 ± 3 0.71 ± 0.26 3172 [1453; 7126] 1.44 ± 0.55 36◦ 29.1120 -117◦ 55.8040 2018
T27S (∗) 21 ± 6 0.65 ± 0.42 2892 [836; 8913] 1.51 ± 0.64 36◦ 28.5490 -117◦ 54.9940 2018
T29-3 (1) 138 ± 23 27.42 ± 12.20 121991 [47476; 310939] 3.02 ± 1.40 36◦ 29.9550 -117◦ 55.9990 2016
T29-3 (2) 116 ± 22 19.49 ± 9.29 86704 [32068; 229077] 2.80 ± 1.34 36◦ 29.9600 -117◦ 55.9620 2016
T32-1-L1 (1) (∗) 36 ± 12 1.84 ± 1.35 8188 [2018; 27469] 1.56 ± 0.66 36◦ 53.8970 -117◦ 57.2090 2016
T32-1-L1 (2) (∗) 20 ± 4 0.56 ± 0.26 2492 [946; 6467] 2.65 ± 0.98 36◦ 32.3540 -117◦ 57.2180 2018
T32-1-L1 (3) (∗) 37 ± 10 2.01 ± 1.18 8923 [2784; 26303] 2.43 ± 0.92 36◦ 32.3370 -117◦ 57.2040 2018
T36-3 (1) 46 ± 4 3.04 ± 1.01 13538 [6483; 29282] 1.17 ± 0.91 36◦ 29.9530 -117◦ 58.5050 2016
T36-3 (2) 35 ± 8 1.82 ± 0.99 8086 [2689; 22918] 2.27 ± 1.03 36◦ 30.0500 -117◦ 58.5180 2016
T36-3 (3) 75 ± 25 8.08 ± 5.83 35924 [9027; 119336] 1.43 ± 0.62 36◦ 29.7240 -117◦ 57.9160 2016
T8-W 8.7 ± 2.6 0.11 ± 0.08 485 [135; 1527] 0.87 ± 0.41 36◦ 22.5220 -117◦ 57.2560 2018
Sua Pan
B7 9.6 ± 0.9 0.13 ± 0.05 341 [117; 1683] 0.92 ± 0.48 -20◦ 35.0460 25◦ 54.6540 2012
L5 11.6 ± 1.1 0.20 ± 0.08 497 [170; 2456] 0.59 ± 0.22 -20◦ 33.9960 26◦ 0.4200 2012
D10 10.6 ± 1.4 0.16 ± 0.07 415 [131; 2198] 0.41 ± 0.16 -20◦ 36.6780 25◦ 55.7940 2012
J11 10.6 ± 1.4 0.16 ± 0.07 415 [131; 2198] 0.69 ± 0.27 -20◦ 37.2420 25◦ 59.2500 2012
B3 10.6 ± 1.4 0.16 ± 0.07 415 [131; 2198] 0.58 ± 0.22 -20◦ 32.8800 25◦ 54.6720 2012
D5 10.6 ± 1.4 0.16 ± 0.07 415 [131; 2198] 0.95 ± 0.33 -20◦ 33.9720 25◦ 55.8180 2012
I4 10.6 ± 1.4 0.16 ± 0.07 415 [131; 2198] 0.53 ± 0.20 -20◦ 33.4470 25◦ 58.6990 2012

TABLE IV. Sauter diameter dS , permeability κ, Rayleigh number Ra, lower Ra limit RaL , upper Ra limit RaU and pattern
wavelength λ measured or calculated for each of the field sites at Badwater Basin, Owens Lake and Sua Pan. For each site,
longitude, latitude and year of the field campaign are also given. Samples stemming from a trench are indicated as (∗).
 21

[1] “NPS Stats: National Park Ser- Press, Washington, DC, 2020).
vice Visitor Use Statistics,” [29] D. Groeneveld, J. Huntington, and D. Barz, J. Hydrol.
https://irma.nps.gov/Stats/Reports/Park/DEVA 392, 211 (2010).
(2022]), acessed on 2022-01-30. [30] D. P. Groeneveld and D. D. Barz, Open J. Mod. Hydro.
[2] T. K. Lowenstein and L. A. Hardie, Sedimentology 32, 03, 241 (2013).
627 (1985). [31] Heather J. Tollerud and Matthew S. Fantle, Remote
[3] T. E. Gill, Geomorphology 17, 207 (1996). Sens. Environ. 154, 285 (2014).
[4] P. R. Briere, J. Arid Environ. 45, 1 (2000). [32] M. C. Baddock, P. Ginoux, J. E. Bullard, and T. E.
[5] R. Bryant and M. Rainey, Remote Sens. Environ. 82, Gill, Geophys. Res. Lett. 43, 2606 (2016).
360 (2002). [33] M.R. Sweeney and V.A. Zlotnik and R.M. Joeckel and
[6] J. M. Prospero, Rev. Geophys. 40 (2002). J.E. Stout, J. Arid Environ. 133, 37 (2016).
[7] M. Klose, T. E. G. abnd Vicken Etyemezian, [34] R. G. Bryant, Prog. Phys. Geogr. 37, 397 (2013).
G. Nikolich, Z. G. Zadeh, N. P. Webb, and R. S. V. [35] J. M. Nield, C. M. Neuman, P. O’Brien, R. G. Bryant,
Pelt, Aeolian Res. 40, 1 (2019). and G. F. Wiggs, Aeolian Res. 23, 51 (2016).
[8] X. Li and F. Shi, Vadose Zone J. 20, e20122 (2021). [36] H. Eloukabi, N. Sghaier, S. B. Nasrallah, and M. Prat,
[9] G. E. Ericksen, J. D. Vine, and R. Ballon, in Lithium Int. J. Heat Mass Trans. 56, 80 (2013).
Needs and Resources (Elsevier, 1978) pp. 355–363. [37] S. Tyler, S. Kranz, M. Parlange, J. Albertson, G. Katul,
[10] G. Wadge, D. Archer, and A. Millington, Terra Nova G. Cochran, B. Lyles, and G. Holder, J. Hydrol. 200,
6, 391 (1994). 110 (1997).
[11] J. Lasser, J. M. Nield, and L. Goehring, Earth System [38] G. A. DeMeo, R. J. Laczniak, R. A. Boyd, J. L. Smith,
Science Data 12, 2881 (2020). and W. E. Nylund, U.S. Geol. Survey 3 (2003).
[12] D. Krinsley, U.S. Geol. Survey CP 70-800 (1970). [39] P. Brunner, P. Bauer, M. Eugster, and W. Kinzelbach,
[13] Y. Dang, L. Xiao, Y. Xu, F. Zhang, J. Huang, J. Wang, J. Hydrol. 294, 241 (2004).
J. Zhao, G. Komatsu, and Z. Yue, Journal of Geophys- [40] S. K. Kampf, S. W. Tyler, C. A. Ortiz, J. F. Munoz,
ical Research: Planets 123, 1910 (2018). and P. L. Adkins, J. Hydrol. 310, 236 (2005).
[14] W. E. Sanford and W. W. Wood, Hydrogeol. J. 9, 358 [41] W. W. Wood, W. E. Sanford, and A. R. S. A. Habshi,
(2001). JSA Bull. 114, 259 (2002).
[15] J. M. Nield, R. G. Bryant, G. F. Wiggs, J. King, D. S. [42] R. Wooding, J. Fluid Mech. 9, 183 (1960).
Thomas, F. D. Eckardt, and R. Washington, Geology [43] G. M. Homsy and A. E. Sherwood, AIChE J. 22, 168
43, 31 (2015). (1976).
[16] V. R. Martı́nez, C. I. Galli, and R. N. Alonso, Revista [44] R. A. Wooding, S. W. Tyler, I. White, and P. A. An-
de la Asociación Geológica Argentina 77, 163 (2020). derson, Water Resour. Res. 33, 1219 (1997).
[17] F. W. Christiansen, Science 139, 607 (1963). [45] R. A. Wooding, Water Resour. Res. 43 (2007).
[18] S. Lokier, J. Arid Environ. 79, 32 (2012). [46] R. L. Van Dam, C. T. Simmons, D. W. Hyndman, and
[19] J. C. Dixon, “Aridic Soils, Patterned Ground, and W. W. Wood, Geophysical Research Letters 36 (2009),
Desert Pavements,” in Geomorphology of Desert Envi- 10.1029/2008GL036906.
ronments (Springer Netherlands, Dordrecht, 2009) pp. [47] L. Goehring, L. Mahadevan, and S. W. Morris, Proc.
101–122. Nat. Acad. Sci. 106, 387 (2009).
[20] P. D. Deckker, Palaeogeogr. Palaeocl. 62, 237 (1988). [48] A. H. Lachenbruch, U. S. Geol. Surv. Spec. Paper 70,
[21] F. D. Eckardt, R. G. Bryant, G. McCulloch, B. Spiro, 69 (1962).
and W. W. Wood, Appl. Geochem. 23, 1563 (2008). [49] R. S. Sletten, B. Hallet, and R. C. Fletcher, J. Geophys.
[22] J. M. Nield, J. King, G. F. S. Wiggs, J. Leyland, R. G. Res. 108, 8044 (2003).
Bryant, R. C. Chiverrell, S. E. Darby, F. D. Eckardt, [50] M. A. Kessler and B. T. Werner, Science 299, 380
D. S. G. Thomas, L. H. Vircavs, and R. Washington, (2003).
J. Geophys. Res.: Atmos. 118, 12,948 (2013). [51] L. Goehring, Philos. T. R. Soc. A 371, 20120353 (2013).
[23] R. Milewski, S. Chabrillat, and B. Bookhagen, Remote [52] X. Fu, L. Cueto-Felgueroso, and R. Juanes, Phil. Trans.
Sens. 12, 474 (2020). R. Soc. A 371, 20120355 (2013).
[24] R. L. Reynolds, J. C. Yount, M. Reheis, H. Goldstein, [53] Marco De Paoli and Sergio Pirozzoli and Francesco
J. Pat Chavez, R. Fulton, J. Whitney, C. Fuller, and Zonta and Alfredo Soldati, J. Fluid Mech. 943, A51
R. M. Forester, Earth Surf. Process. Landforms 32, 1811 (2022).
(2007). [54] S. G. Fryberger, A. M. Al-Sari, and T. J. Clisham,
[25] I. Y. Fung, S. K. Meyn, I. Tegen, S. C. Doney, J. G. AAPG Bulletin 67, 280 (1983).
John, and J. K. B. Bishop, Global Biogeochem. Cy. [55] N. Bowden, S. Brittain, A. G. Evans, J. W. Hutchinson,
14, 281 (2000). and G. M. Whitesides, Nature 393, 146 (1998).
[26] T. A. Cahill, T. E. Gill, J. S. Reid, E. A. Gearhart, and [56] L. Mahadevan and S. Rica, Science 307, 1740 (2005).
D. A. Gillette, Earth Surf. Proc. Land. 21, 621 (1996). [57] B. Li, Y.-P. Cao, X.-Q. Feng, and H. Gao, Soft Matter
[27] T. E. Gill, D. A. Gillette, T. Niemeyer, and R. T. Winn, 8, 5728 (2012).
Nucl. Instrum. Meth. B 189, 209 (2002). [58] J. Dervaux and M. B. Amar, Annu. Rev. Condens. Mat-
[28] National Academies of Sciences, Engineering, and ter Phys. 3, 311 (2012).
Medicine, Effectiveness and Impacts of Dust Control [59] A. H. Lachenbruch, Journal of Geophysical Research 66,
Measures for Owens Lake (The National Academies 4273 (1961).
 22

[60] K. A. Shorlin, J. R. de Bruyn, M. Graham, and S. W. [86] C. Thomas, S. Dehaeck, and A. D. Wit, Int. J. Greenh.
Morris, Phys. Rev. E 61, 6950 (2000). Gas Con. 72, 105 (2018).
[61] R. M. Tucker, J. of Sediment. Res. 51, 779 (1981). [87] J. M. Nield, J. Lasser, and L. Goehring, “TLS surface
[62] M. T. Mellon, C. P. Mckay, and J. L. Heldmann, scans from Owens Lake and Badwater Basin, central
Antarctic Sci. 26, 413 (2014). California, measured in 2016 and 2018,” (2020).
[63] M. Steiger and S. Asmussen, Geochimica et Cosmochim- [88] J. Lasser and L. Goehring, “Subsurface salt concentra-
ica Acta 72, 4291 (2008). tion profiles and pore water density measurements from
[64] J. Lasser, M. Ernst, and L. Goehring, J. Fluid Mech. Owens Lake, central California, measured in 2018,”
917 (2021). (2020).
[65] J. D. Stevens, J. M. S. Jr., C. T. Simmons, and T. Fen- [89] Uri Nachshon and Noam Weisbrod and Roee Katzir and
stemaker, J. Hydrol. 375, 394 (2009). Ahmed Nasser, Geophys. Res. Lett. 45, 6100 (2018).
[66] J. A. Neufeld, M. A. Hesse, A. Riaz, M. A. Hallworth, [90] N. Otsu, Systems, Man and Cybernetics, IEEE Trans-
H. A. Tchelepi, and H. E. Huppert, Geophys. Res. Lett. actions on 9, 62 (1979).
37 (2010). [91] Sheng Dai and Hosung Shin and J. Carlos Santamarina,
[67] A. C. Slim, J. Fluid Mech. 741, 461–491 (2014). Acta Geotech. 11, 1103 (2016).
[68] D. R. Hewitt, J. A. Neufeld, and J. R. Lister, J. Fluid [92] G. Licsandru and C. Noiriel and P. Duru and S. Geoffroy
Mech. 756, 844 (2014). and A. Abou Chakra and M. Prat, Phys. Rev. E 100,
[69] C. Gueler and G. D. Thyne, J. Hydrol. 285, 177 (2004). 032802 (2019).
[70] J. Ryu, R. A. Zierenberg, R. A. Dahlgren, and S. Gao, [93] A. Mojtabi and M.-C. Charrier-Mojtabi, in Handbook
Chem. Geol. 229, 257 (2006). of Porous Media, edited by K. Vafai (CRC Press, 2005)
[71] L. Nicholas and B. Andy, Earth Surf. Proc. Land. 23, pp. 269–320.
69 (1997). [94] IPCC, “Climate Change 2021: The Physical Science Ba-
[72] C. van Duijn, G. Pieters, R. Wooding, and A. van der sis.” (Cambridge University Press, 2021) Chap. 6.
Ploeg, in Environmental Mechanics: Water, Mass and [95] E. Rodriguez-Caballero, T. Stanelle, S. Egerer,
Energy Transfer in the Biosphere, edited by P. A. Raats, Y. Cheng, H. Su, Y. Canton, J. Belnap, M. O. Andreae,
D. Smiles, and A. W. Warrick (American Geophysical I. Tegen, C. H. Reick, U. Pöschl, , and B. Weber, Na-
Union, 2002) pp. 155–169. ture Geosci. 15, 458 (2022).
[73] B. P. Boudreau, Geochim. Cosmochim. Ac. 60, 3139 [96] W. A. Wurtsbaugh, C. Miller, S. E. Null, R. J. DeRose,
(1996). P. Wilcock, M. Hahnenberger, F. Howe, and J. Moore,
[74] B. P. Boudreau, Diagenetic models and their implemen- Nature Geosci. 10, 816 (2017).
tation: Modelling transport and reactions in aquatic sed- [97] N. Juergens, Science 339, 1618 (2013).
iments (Springer, 2011). [98] J. Lasser, L. Goehring, and J. M. Nield, “Images and
[75] M. Boufadel, M. Suidan, and A. Venosa, J. Hydrol. Videos from Owens Lake and Badwater Basin in central
221, 55 (1999). California, taken in 2016 and 2018,” (2020).
[76] V. Loodts, L. Rongy, and A. De Wit, Chaos 24, 043120 [99] C. Canuto, M. Y. Hussaini, A. Quarteroni, and T. A. Z.
(2014). Jr., Spectral Methods in Fluid Dynamics (Springer,
[77] J. Elder, J. Fluid Mech. 27, 29 (1967). 1988).
[78] D. Hewitt, Proc. R. Soc. A 476, 20200111 (2020). [100] G. E. Karniadakis, M. Israeli, and S. A. Orszag, J.
[79] D. R. Hewitt, J. A. Neufeld, and J. R. Lister, Phys. Comp. Phys. 97, 414 (1991).
Rev. Lett. 108, 224503 (2012). [101] A. Riaz and E. Meiburg, J. Fluid Mech. 494, 95 (2003).
[80] J. Lasser and L. Goehring, “Grain size distributions of [102] C.-Y. Chen, Topics in miscible porous media flows (Uni-
sand samples from Owens Lake and Badwater Basin in versity of Southern California, 1998).
central California, collected in 2016 and 2018,” (2020). [103] K. J. Hollet, W. R. Danskin, W. F. McCaffrey, and
[81] J. Sauter, VDI-Forschungsheft 279 (1928). C. L. Waiti, U.S. Geol. Survey (1991).
[82] X. Garcia, L. T. Akanji, M. J. Blunt, S. K. Matthai, [104] C. B. Hunt, T. Robinson, W. Bowles, and A. Washburn,
and J. P. Latham, Phys. Rev. E 80, 021304 (2009). U.S. Geol. Survey (1966).
[83] J. Lasser and V. Karius, “Chemical characterization of [105] LADWP, “Owens Lake habitat management plan,”
salt samples from Owens Lake and Badwater Basin, cen- (2010).
tral California, collected in 2016 and 2018,” (2020). [106] O. Annunziata, J. A. Rard, J. G. Albright, L. Paduano,
[84] R. A. Wooding, S. W. Tyler, I. White, and P. A. An- and D. G. Miller, J. Chem. Eng. Data 45, 936 (2000).
derson, Water Resour. Res. 33, 1199 (1997). [107] R. Cohen-Adad and J. W. Lorimer, eds., Solubility Data
[85] A. C. Slim, M. M. Bandi, J. C. Miller, and L. Mahade- Series, Vol. 47 (Pergamon, 1991).
van, Phys. Fluids 25, 024101 (2013). [108] R. H. Perry, ed., Perry’s Chemical Engineers’ Handbook,
7th ed. (McGraw-Hill, 1997).