Arxiv:2010.11120V2 [Physics.Flu-Dyn] 2 Dec 2020 Verma 2018)

Draft version December 3, 2020 Typeset using LATEX twocolumn style in AASTeX63

Non-Boussinesq low-Prandtl number convection with a temperature-dependent thermal diﬀusivity

Ambrish Pandey,1 Jorg¨ Schumacher,2, 3 and Katepalli R. Sreenivasan1, 3

1Center for Space Science, New York University Abu Dhabi, Abu Dhabi 129188, United Arab Emirates 2Institut f¨urThermo- und Fluiddynamik, Technische Universit¨atIlmenau, Postfach 100565, D-98684 Ilmenau, Germany 3Tandon School of Engineering, New York University, New York 11201, USA

(Dated: December 3, 2020) Submitted to AJ

ABSTRACT In an attempt to understand the role of the strong radial dependence of thermal diffusivity on the properties of convection in sun-like stars, we mimic that effect in non-Oberbeck-Boussinesq (NOB) convection in a horizontally-extended rectangular domain (aspect ratio 16), by allowing the thermal diffusivity κ to increase with the temperature (as in the case of stars). Direct numerical simulations (i.e., numerical solutions of the governing equations by resolving up to the smallest scales without requiring any modeling) show that, in comparison with Oberbeck-Boussinesq (OB) simulations (two of which we perform for comparison purposes), the symmetry of the temperature field about the mid- horizontal plane is broken, whereas the velocity and heat flux profiles remain essentially symmetric. Our choice of κ(T ), which resembles the variation in stars, results in the temperature field that loses its fine structures towards the hotter part of the computational domain, but the characteristic large scale of the turbulent thermal ‘superstructures’, which are structures whose size is typically larger than the depth of the convection domain, continue to be largely independent of the depth.

Keywords: editorials, notices — miscellaneous — catalogs — surveys

1. INTRODUCTION defined as the ratio of the kinematic viscosity and ther- Thermal convection is the primary driving mechanism mal diffusivity of the fluid. The aspect ratio Γ, which in many turbulent flows in nature. For instance, it is re- is the ratio of the horizontal to the vertical dimensions sponsible for the transport of heat and momentum in of the flow domain, also influences the flow. Flow do- the convection zone of the Sun, sun-like stars and many main in natural convective systems are usually horizon- of their planets (Miesch 2005; Nordlund et al. 2009; Rin- tally extended, which leads to the formation of regular con & Rieutord 2018; Hanasoge et al. 2016; Schumacher large-scale patterns which are denoted as turbulent su- & Sreenivasan 2020). The Rayleigh-Bénardconvection perstructures. These are, for example, granules and su- (RBC), which is a paradigm of thermal convection sys- pergranules in the Sun (Nordlund et al. 2009) and other tems, consists of a fluid layer confined between two stars (Michel et al. 2008) or cloud streets in the atmo- horizontal plates heated from below and cooled from sphere (Markson 1975). Moreover, convection in the above (Ahlers et al. 2009; Chillà& Schumacher 2012; Sun and other stellar objects are characterized by low arXiv:2010.11120v2 [physics.flu-dyn] 2 Dec 2020 Verma 2018). The main governing parameters of RBC P r and very high Ra (Nordlund et al. 2009; Hanasoge are the Rayleigh and Prandtl numbers. The Rayleigh et al. 2016; Schumacher & Sreenivasan 2020), accompa- number Ra signifies the strength of driving force by nied by strong spatial variations of fluid properties such buoyancy relative to dissipative forces due to viscosity as pressure, density and temperature. and thermal diffusivity, and the Prandtl number P r is Convection in low-P r fluids is dominated by inertia and the flow is highly turbulent compared to those for unity-P r fluids (Schumacher et al. 2015, 2016; Scheel Corresponding author: Ambrish Pandey & Schumacher 2016; Pandey et al. 2018; Schumacher & [email protected] Sreenivasan 2020). Direct numerical simulations (DNS) 2 A. Pandey, J. Schumacher, & K. R. Sreenivasan of very low-P r turbulent convection at high-Ra with- transport are largely unaffected, whereas the properties out parametrizations of turbulence are challenging due of the temperature field are different from those in the to very small length and time scales, which need to OB convection. Further, the variation of the small-scale be resolved properly in order to faithfully study such content of superstructures in the temperature field de- flows. Large Γ places even greater demands on re- creases with depth, although the underlying large-scale sources but this has been achieved recently (Pandey structure remains essentially unchanged. et al. 2018). However, these simulations were made Properties of turbulent superstructures, such as their under the Oberbeck-Boussinesq (OB) approximation, characteristic spatial and temporal scales and their con- where the density variation is neglected in all terms of tribution in the global heat transport, have been studied the governing equations except in the buoyancy term. for a range of Prandtl and Rayleigh numbers (Pandey Additionally, molecular transport coefficients, e.g., ther- et al. 2018; Schneide et al. 2018; Fonda et al. 2019; mal diffusivity, kinematic viscosity, and thermal expan- Stevens et al. 2018; Green et al. 2020; Krug et al. sion coefficient were assumed to be constant and inde- 2020). These numerical studies, however, adopted the pendent of temperature. OB model of convection, which, as we just implied, does The OB approximation is obviously not applica- not accurately yield the properties of turbulent super- ble to the stellar environment which is characterized structures in stellar convection. It is therefore important by a very strong stratification, which is in line with to systematically explore the non-Oberbeck-Boussinesq orders-of-magnitude variations of the material parame- effects on the properties of turbulent superstructures of ters and state variables such as temperature, pressure convection. This is one motivation of the current work. and mass density across the depth of the convection Another motivation is to study the statistical nature of zone (Christensen-Dalsgaard et al. 1996; Hanasoge et al. the velocity and temperature fields. 2016; Schumacher & Sreenivasan 2020). For the Sun, the In the past, NOB convection has been explored pri- thermal conductivity due to photons displays a cubic marily to study the properties of heat transport and dependence on temperature, k(T ) ∼ T 3, thus varying flow structures in confined flow domains of aspect ratio more strongly than the kinematic viscosity due to ions, Γ ≈ 1 (Wu & Libchaber 1991; Zhang et al. 1997; Ahlers ν(T ) ∼ T 5/2ρ−1, which causes the molecular Prandtl et al. 2006; Sameen et al. 2008, 2009; Horn et al. 2013; number to be of the order 10−6 throughout the convec- Valori et al. 2017, 2019). NOB effects manifest, among tion zone strongly even more towards the surface (Mi- other effects, by a mean temperature that is different esch 2005; Freytag et al. 2012; Schumacher & Sreeni- from the arithmetic mean of top and bottom tempera- vasan 2020). Therefore, we think that the effect of vary- tures, which leads to symmetry breaking characteristic ing thermal conductivity k(T ) and thus of the thermal of OB convection (Hurlburt et al. 1984; Wu & Libchaber diffusivity κ(T ) is the more dominant effect to include 1991; Zhang et al. 1997; Evonuk & Glatzmaier 2004; for purposes of generating in our simple model with con- Skrbek & Urban 2015). In NOB convection, the width stant ν the right qualitative trend of P r across the con- of the thermal and viscous boundary layers (BLs) as vection zone. We know from preliminary calculations well as the temperature drop within the thermal BLs that the additional inclusion of ν(T ) breaks the sym- at the two plates are different (Wu & Libchaber 1991; metry of the velocity field and heat flux, which is not Ahlers et al. 2006; Horn et al. 2013). In Sameen et al. observed by including only κ(T ). These results will be (2009), the NOB effects were systematically explored reported separately as the next logical step. by varying various temperature-dependent fluid prop- Specifically, we perform direct numerical simulations erties one at a time and keeping others temperature- of convection in an extended rectangular domain (aspect independent. Recently, Shcheritsa et al.(2018) incorpo- ratio 16) by letting the thermal diffusivity in the govern- rated temperature-dependent thermal diffusivity in their ing equations to depend on the temperature. This model DNS in a horizontally-extended domain to study the introduces a strong temperature stratification towards multiscale nature of the solar convective flows. They the top of the domain, resulting in a depth-variation of observed that the flow near the isothermal boundary is P r similar to that in the Sun (Schumacher & Sreeni- a superposition of cellular structures with different char- vasan 2020). To serve as a reference, we perform two acteristic scales, which qualitatively agree with those ob- simulations of RBC in the Boussinesq limit with Prandtl served in the Sun (Shcheritsa et al. 2018). and Rayleigh numbers corresponding to those at the top and bottom plates of our NOB convection run. Our 2. GOVERNING EQUATIONS AND FORMALISM major finding is that some of the flow properties in the In this work, we extend the OB model of thermal NOB convection, such as the profiles of velocity and heat convection by considering only κ as a function of tem- Non-Boussinesq low-Prandtl number convection 3 perature, while neglecting the temperature dependence where κtop is thermal diffusivity at the top plate. It of molecular viscosity and the thermal expansion coeffi- is clear that, for positive a, b, α, β, the thermal diffusiv- cient. We regard this approach as a first step in system- ity increases with depth due to increasing temperature. atically understanding the NOB effects in an extended As a result (see Busse(1967)), the conduction tem- flow domain that is relevant to natural convective flows. perature profile does not remain linear as in OB con- A considerable effort has already been devoted to the vection (see Figure1). In the following, we will de- study of the solar and astrophysical convection by simu- scribe the consequences of temperature-dependent κ on lating the governing equations in fully compressible and the velocity and temperature profiles, as well as on the anelastic limits, along with other driving factors such relations between global heat transport and dissipation as Coriolis and Lorentz forces, differential rotation, etc. rates (Howard 1972; Shraiman & Siggia 1990). Global simulations in the anelastic limit of solar convection have been reported for example by Miesch et al. 2.1. Conduction temperature profile (2008); Käpyläet al.(2010); Nelson et al.(2014); Hotta In the conduction state of no fluid motion, tempera- et al.(2016) or Guerrero et al.(2016). Simulations of ture varies only in the vertical direction, i.e., u = 0 and the upper convection zone, which are fully compressible T = Tc(z). Therefore, Eq. (2) in the conduction state and include detailed radiative transfer, were conducted becomes by Stein & Nordlund(1998); Gudiksen et al.(2011); ∂ ∂Tc κc(Tc) = 0, (6) Freytag et al.(2012); Magic et al.(2013); Riethmüller ∂z ∂z et al.(2014) or Wray et al.(2015). All these simulation which yields models have to apply subgrid models for the unresolved ∂T κ (T ) c = c , (7) scales, rely on numerical diffusion, or are even run as c c ∂z 1 implicit (large-eddy) simulations. Our long-term goal, where c1 is a constant of integration. Integration of however, is to systematically and incrementally include this equation and the use of the explicit temperature the complexities present in realistic astrophysical flows dependence of κ from Eqs. (4) and (5) gives in direct numerical simulation models starting with their influence in the simplest case of RBC. They will provide a α+1 b β+1 κtop Tc + T + T = c1z + c2, (8) helpful insights on scaling of effective eddy viscosity and α + 1 c β + 1 c diffusivity that have to be used in full-size simulations. where c is another constant of integration. The con- We solve the following governing equations 2 stants c1 and c2 can be determined by specifying the ∂u ∇p + u · ∇u = − + γgT zˆ + ν∇2u, (1) temperature boundary conditions at the top and bot- ∂t ρ0 tom plates. For the isothermal boundary conditions, ∂T + u · ∇T = ∇ · [κ(T )∇T ], (2) i.e., T (z = 0) = Tbot = 1 and T (z = 1) = Ttop = 0, ∂t the constants c1 and c2 can be fixed. The boundary ∇ · u = 0, (3) condition at z = 0 gives where u (= u xˆ + u yˆ + u zˆ), T , and p are the velocity, x y z a b temperature and pressure fields, respectively. The co- c = κ 1 + + =: κ C, (9) 2 top α + 1 β + 1 top efficients γ and ν are the thermal expansion coefficient and the kinematic viscosity of the fluid, respectively, ρ0 with a b is the mean density of the fluid, and g is the acceleration C = 1 + + , (10) due to gravity. Note that the last term in the tempera- α + 1 β + 1 ture equation [Eq. (2)] becomes κ∇2T in OB convection. and with the condition at z = 1, one obtains The temperature drops very quickly over a relatively short distance near the photosphere in the convection c1 = −c2 = −κtopC. (11) zone of the Sun (Schumacher & Sreenivasan 2020), so we choose a particular form of κ(T ) to mimic such Using the above c1 and c2 in Eq. (8), the conduction a temperature drop in our flow domain near the top temperature profile can be determined by solving plate (Shcheritsa et al. 2018). We consider κ(T ) as a a b T (z)+ T α+1(z)+ T β+1(z) = C(1−z). (12) polynomial in T such that c α + 1 c β + 1 c κ(T ) = κ f(T ) (4) top Note that the shape of resulting nonlinear conduction with temperature profile depends on the functional form of f(T ) = 1 + aT α + bT β, (5) κ(T ), i.e., on the coefficients a, b, α, and β in Eq. (5). 4 A. Pandey, J. Schumacher, & K. R. Sreenivasan

2.2. Conductive heat transport The boundary conditions and the conservation of inter- Heat flux in the conduction state is defined as nal energy implies that Nu(z) remains a constant across the fluid layer in the NOB convection, as in OB convec- ∂Tc Jc = −κc(Tc)∇Tc = −κc(Tc) , (13) tion. Integrating Eq. (20) in the vertical direction yields ∂z the volume-averaged Nusselt number where κc(Tc) is the thermal diffusivity profile in the con- huzT iV,t − hκ(T )∂T/∂ziV,t duction state. Note that Jc simply becomes κ∆T/H in Nu = , (21) OB convection due to a fixed temperature gradient over Jc the depth of the fluid layer. In the present case, how- which says that the total heat flux results from the ever, the conductive heat flux can be estimated using convective as well as the diffusive heat fluxes. Note Eq. (7) as that the volume-averaged diffusive heat transport Jd = Jc = −c1 = κtopC, (14) −hκ(T )∂T/∂ziV,t in the NOB convection is not equal where C = C(a, b, α, β) is given by Eq. (10). Thus Jc can to Jc, unlike in the OB convection. Thus, the volume- be estimated analytically once the coefficients a, b, α, β, averaged Nusselt number in the NOB convection with and κtop are specified. Moreover, Jc is a constant at κ(T ) is computed as each point of the flow, and is therefore the same as the hu T i + J volume-averaged conductive heat flux, i.e., Nu = z V,t d . (22) Jc hJciV = Jc = κtopC, (15) 2.4. Exact relations where h·i is the average over the entire domain. V Exact relations relating the volume-averaged dissipa- 2.3. Turbulent heat transport tion rates to the global heat transport in RBC can be derived by volume- and time-averaging the equations gov- To determine the heat transport in the convective erning the temporal evolutions for the kinetic and ther- state, we rewrite the temperature equation [Eq. (2)] us- mal energies (Howard 1972; Shraiman & Siggia 1990). ing the mass conservation as The viscous and thermal dissipation rates at each point ∂T in the flow are defined, respectively, as + ∇ · (uT − κ(T )∇T ) = 0. (16) ∂t 2 ν ∂ui ∂uj Horizontal-averaging of Eq (16) yields εu = + , (23) 2 ∂xj ∂xi ∂hT iA 2 + h∇ · (uT − κ(T )∇T )iA = 0, (17) ∂T ∂t εT = κ(T ) , (24) ∂xi where h·iA denotes a horizontally-averaged quantity. Us- ing the no-slip condition for the velocity field and the where ui and xi are the components of the velocity and adiabatic condition for the temperature field on the side- position vector. Following the similar steps as in Howard walls, and additionally averaging in time, we get (1972); Shraiman & Siggia(1990), one can obtain the exact relations in convection with temperature-dependent ∂ ∂T thermal diffusivity as uzT − κ(T ) = 0, (18) ∂z ∂z A,t hεuiV,t = γg[JcNu − Jd], (25) which yields ∆T hε i = J Nu, (26) ∂T T V,t c H hu T i − κ(T ) = J , (19) z A,t ∂z z A,t where ∆T and H are the temperature difference and distance between the top and bottom plates. In the where Jz is a constant. Note that the use of periodic conditions on the sidewalls also results in Eq. (19). A next section, we describe details of DNS of convection non-dimensional measure of the convective heat trans- in both OB and NOB cases. port is the Nusselt number Nu, defined as the ratio of the total heat transport to the conductive heat trans- 3. DETAILS OF DIRECT NUMERICAL port, which is computed for each horizontal plane as SIMULATIONS We investigate convection flow in a three-dimensional huzT iA,t − hκ(T )∂T/∂ziA,t Nu(z) = . (20) rectangular box of horizontal dimensions Lx = Ly = Jc Non-Boussinesq low-Prandtl number convection 5

16H, where the top and bottom plates are kept at (dif- is not the same as Racrit = 1708 for the OB convec- ferent) constant temperatures and satisfy no-slip bound- tion (Chandrasekhar 1981). We determined the criti- ary condition. Periodic boundary conditions are em- cal Rayleigh number, Racrit, at which temperature pro- ployed on the sidewalls (Bekki et al. 2017). Convective file starts to depart from the conduction profile Tc(z) flow properties in horizontally-extended domains with (shown in Figure1)(Shcheritsa et al. 2018). For κ(T ) = 6 5 periodic sidewalls have been observed to be nearly the κtop(1 + 49T + 450T ), we find Racrit ≈ 7 × 10 , more same as those with solid sidewalls (Pandey et al. 2018). than two orders of magnitude higher than the Racrit for In any case, periodic boundary conditions rather than OB convection. Note that critical Ra also increases in solid walls are more relevant to convection in the Sun compressible convection with increasing density stratifi- and other stars. We non-dimensionalize the governing cation (Hurlburt et al. 1984; Gastine & Wicht 2012). equations using H, ∆T , Uf , and Tf as the length, tem- For comparison purposes, we also perform two OB perature, velocity, and time scales, respectively. Here convection simulations with constant κ. One simula- √ Uf = γg∆TH is the free-fall velocity and Tf = H/Uf tion has the Rayleigh and Prandtl numbers correspond- is the free-fall time. The resulting non-dimensional gov- ing to the top plate of our NOB simulation, i.e., for erning equations are P r = 12.73, Ra = 1.708 × 108; another has Ra and P r corresponding to those at the bottom plate, i.e, for s ∂u P r P r = 0.02546, Ra = 3.416 × 105. These two OB simula- + u · ∇u = −∇p + T zˆ + top ∇2u, (27) ∂t Ratop tions will be referred to hereafter as OBtop and OBbot, " # respectively (or jointly as OB simulations), whereas ∂T f(T ) the NOB convection simulation with the temperature- + u · ∇T = ∇ · p ∇T , (28) ∂t P rtopRatop dependent thermal diffusivity will be referred to as NOB ∇ · u = 0, (29) simulation. The important details of the DNS are summarized in Table1. Some further comments on the where the Rayleigh and Prandtl numbers are speci- structure of the computational grid, imposed by the 3 fied at the top plate as Ratop = γg∆TH /(νκtop) and asymmetry of the temperature profiles, can be found P rtop = ν/κtop, and f(T ) is given by Eq. (5). Due in the next section. to varying κ(T ) with depth, the Rayleigh and Prandtl numbers, which are defined as Ra = γg∆TH3/νκ and 4. VELOCITY AND TEMPERATURE STATISTICS P r = ν/κ, also vary in the flow. Note that the temper- From the numerical data, we compute several quan- ature dependence of κ only appears in Eq. (28) as the tities to compare the NOB case with OB simulations. viscosity ν is kept constant everywhere. We first discuss the depth dependence of horizontally- The governing equations (27)–(29) are solved using averaged quantities. an exponentially fast converging spectral element solver Nek5000 (Fischer 1997). As in Scheel et al.(2013), 4.1. Vertical profiles the simulation domain is divided into a finite number The temperature profile in the conduction state is lin- of elements (N ), and on each spectral element the tur- e ear in OB convection, whereas that in NOB convection bulence fields are expanded with Lagrangian interpola- becomes nonlinear, as shown in Figure1. The tempera- tion polynomials of order N. For this work, we choose ture gradients are different at the two horizontal plates. f(T ) = 1 + 49T + 450T 6, and therefore, the thermal Using Eq. (7), we can estimate the ratio of the tempera- diffusivity varies with temperature as ture gradients at the two plates in the conduction state 6 as κ(T ) = κtop(1 + 49T + 450T ) . (30) (∂T /∂z) κ c top = bot = 500, (31) (∂Tc/∂z)bot κtop Thus, the thermal diffusivity at the bottom plate κbot = 500κtop. We perform a direct numerical simulation us- showing that, in the vicinity of the top plate, the tem- 8 ing this κ(T ) for P rtop = 12.73 and Ratop = 1.708×10 , perature drops very sharply to T = 0. Figure1 shows which correspond to P rbot = 0.02546 and Rabot = that ∂Tc/∂z is negative at all depths, that is, the flow 3.416×105. Thus, the Rayleigh and Prandtl numbers at is unstable everywhere. the bottom of the flow are 500 times smaller than those In the convective state shown in Figure2, enhanced at the top. mixing leads to approximately a constant temperature We mention here that the critical Rayleigh number, in the bulk region of the flow. We demonstrate this by Racrit, at which heat cannot be transported by molec- plotting the horizontally- and temporally-averaged tem- ular diffusion alone and convection comes into effect, perature field hT iA,t as a function of the depth z. The 6 A. Pandey, J. Schumacher, & K. R. Sreenivasan

Table 1. Important parameters of DNS in a three-dimensional (3D) rectangular box of dimensions Lx : Ly : H = 16 : 16 : 1; Ne is the total number of elements in the entire flow domain and N is the order of Lagrangian polynomials for the spectral interpolation within each element. For OB simulations, the flow domain is divided into 2202 × 32 = 1548800 elements. For the NOB simulation with temperature-dependent thermal diffusivity, the flow domain is divided into 3202 × 87 = 8908800 elements, where the width and length of the elements are the same everywhere but the height of the elements varies with the depth; as discussed in Section 4.1, the density of elements in the vicinity of the top plate is much larger than near the bottom. The P r and Ra values for the NOB case vary with the depth, and their values at the top and bottom plates are the same as those of the OB simulations, OBtop and OBbot. Nu, Re, and urms, are the volume- and time-averaged Nusselt number, Reynolds number, and the root-mean-square (rms) velocity, respectively. λU , λΘ, and τ are the characteristic spatial and temporal scales of turbulent superstructures (Pandey et al. 2018). ttotal is the total integration time in units of the free-fall time (in reference to the top wall conditions for the NOB case).

Run κ/κtop P r Ra Ne N Nu urms Re λU /H λΘ/H τ (Tf ) ttotal (Tf ) NOB 1 + 49T + 450T 6 variable variable 8908800 5 5.4 0.300 1100 4.0 4.0 47 24 8 OBtop 1 12.73 1.708 × 10 1548800 9 35.8 0.094 345 4.7 8.0 169 128 5 OBbot 1 0.025 3.416 × 10 1548800 9 3.65 0.469 1736 4.0 4.0 30 18

1984; Evonuk & Glatzmaier 2004). We compute the 1.0 NOB horizontally- and temporally averaged vertical temper- OB ature derivatives at the plates and ﬁnd that, in the con- 0.8 vective state,

0.6 ) h(∂T/∂z)topiA,t z

( ≈ 495, (32) c h(∂T/∂z)botiA,t T 0.4 which is roughly the same as for the conduction state. 0.2 Figure2 also shows that a boundary layer (BL) structure forms at the walls in the NOB case. The BL thick- 0.0 nesses are shown by vertical dashed lines. (Here and 0.0 0.2 0.4 0.6 0.8 1.0 elsewhere, the thermal BL thickness is defined as the z/H distance over which the temperature attains the center- plane value if it continued to vary linearly with the slope manifested at each wall.) Other characteristics of the Figure 1. Conduction temperature profiles for the OB and NOB cases. It is nonlinear for the latter, with strongly dis- NOB convection are the different thicknesses of the ther- similar structure at the bottom and top plates. mal BLs and different temperature drops within these BLs at the top and bottom plates (Wu & Libchaber Figure shows that the averaged temperature in the bulk, 1991). We compute the thickness of the thermal BL i.e., roughly between z = 0.2 and z = 0.8, does not vary δT , using the slope method (Scheel et al. 2012), and find that δT /H = 0.18 and 0.0017, respectively, at the appreciably, and that Tbulk ≈ 0.82, which is not equal to the arithmetic mean of the temperatures of the top and bottom and top plates. Therefore, δT at the bottom is roughly 100 times thicker than that at the top plate. bottom plates, i.e., Tbulk 6= ∆T/2. This is a characteristic of NOB convection, where the bulk temperature is A strong temperature stratification in our flow poses different from the arithmetic mean of the temperatures a great challenge to properly resolve the temperature of the top and bottom plates (Wu & Libchaber 1991; field near the top plate. Figure2 shows that the tem- Zhang et al. 1997; Ahlers et al. 2006; Horn et al. 2013; perature drop near the top boundary is similar for both Skrbek & Urban 2015). Higher bulk temperature in the NOB and OBtop simulations. This is demonstrated our simulation is consistent with the findings of Sameen more clearly by exhibiting the profiles in a magnified re- et al.(2009) who observed that the bulk temperature gion near the top plate in Figure3(a). We observe that rises if the thermal diffusivity increases with increas- hT iA,t(z) is indeed similar for the two cases in region ing temperature, and vice versa. Moreover, higher bulk z > 0.98H, and it appears that a similar resolution re- temperature observed here is similar to that observed quirement should be obtained for the two cases near the in fully compressible and anelastic models of convec- top boundary. However, hT iA,t for OBtop varies appre- tion with strong density stratification (Hurlburt et al. ciably only in region z > 0.98H, whereas the variation is discernible for the entire z > 0.8H in the NOB case. Non-Boussinesq low-Prandtl number convection 7

To quantify this variation, we estimate the temperature (a) scale height as 0.8 −1 d loghT iA,t hT iA,t HT = − = − (33) 0.6 dz dhT iA,t/dz t

and plot it as a function of z in Figure3(b). We observe , A 0.4 that HT for both the simulations are similar only in the T vicinity of the top plate, and, as one moves farther from the plate, HT becomes larger for OBtop than for NOB 0.2 case. This indicates that a finer resolution would be NOB needed in the entire top region for the NOB simulation. OBtop Furthermore, the NOB simulation is accompanied by 0.0 a strong variation of κ near the top plate, which causes 0.80 0.85 0.90 0.95 1.00 the quantities involving thermal diffusivity, such as the (b) thermal dissipation rate εT (x) and the diffusive heat flux 102 Jd, to become quite sensitive to the vertical resolution. To capture this sharp gradient in the temperature field and in the derived quantities near the top plate, we need 100 to generate a lot more grid points there compared to that 6 near the bottom plate. T H 10 2 T 4

1.0 2 10 4 0.8 0.9995 1.0000 z

t 0.6 0.80 0.85 0.90 0.95 1.00 , A z/H T 0.4 NOB Figure 3. Vertical profiles of (a) temperature and (b) tem- 0.2 OBtop perature scale height HT near the top plate for the NOB and OBtop simulations. Inset in panel (b) shows the profile of the OBbot 0.0 thermal dissipation rate for the NOB case in the vicinity of 0.0 0.2 0.4 0.6 0.8 1.0 the top plate, which reveals that the profile is smooth, even z/H though varying strongly, thus indicating the adequacy of the spatial resolution.

Figure 2. Vertical temperature profiles for all the simula- stratified regions. To show the sufficiency of our verti- tions. Averaged temperature in the bulk region for the NOB simulation is approximately 0.82, whereas it is 0.5 for the OB cal grid resolution, we display the vertical profile of the simulations. The thermal BL thicknesses δT are indicated by horizontally- and temporally-averaged thermal dissipa- vertical dashed lines with the same colors as the correspond- tion rate in the vicinity of the top plate in the inset of ing temperature profiles. It is clear that δT values for the Figure3(b). We can see that due to very large ∂T/∂z in NOB simulation are quite different at the top and bottom the vicinity of the top plate, the scalar dissipation rate, plates, unlike for OB simulations for which they are equal. εT , becomes very large as we approach the top boundary. Nonetheless, εT varies smoothly near the top plate Thus, we have to build a highly asymmetric spectral in our simulation indicating that our asymmetric grid is element mesh in the vertical direction with 18, 15, and able to capture the strong variation in the scalar dissi- 54 elements, respectively, in the bottom, bulk, and top pation rate, which is very sensitive to the spatial reso- regions. As a consequence, the number of spectral ele- lution (Scheel et al. 2013). Indeed, we have tested that ments for the NOB case grows by a factor of 6 at reduced if we use a symmetric grid as in the OB cases, strong polynomial order N in comparison to OB cases, which variations of εT and Jd cannot be captured adequately, demonstrates the flexibility of our numerical method while sustaining the spectral accuracy for the strongly 8 A. Pandey, J. Schumacher, & K. R. Sreenivasan leading to oscillations in the profiles of these quantities (a) near the top plate. 0.20 In Figure2, we also plot hT iA,t(z) from our OB simulations, which indicate the symmetry about the midplane in the OB convection. The temperature in the 0.15 bulk region is 0.5, which is the same as the arithmetic s

mean ∆T/2 of the top and bottom plate temperatures. m

r 0.10 The profile for the OBtop exhibits that the temperature field is well mixed in the bulk region, which lies for 0.05 ≤ z ≤ 0.95. This is due to large Prandtl 0.05 and Rayleigh numbers, which results in a thin ther- NOB OBtop OBbot mal BLs near the top and bottom plates. The profile 0.00 for the OB , however, exhibits a temperature gradient bot 0.0 0.2 0.4 0.6 0.8 1.0 even in the bulk region. The thermal BLs in this case are much thicker due to smaller Prandtl and Rayleigh (b) numbers (Scheel & Schumacher 2016, 2017; Schumacher 0.5 et al. 2016; Pandey 2020). In OB convection, the widths of the thermal BLs at the top and bottom plates are the 0.4

same, and related to the globally-averaged Nusselt num- s m r

, 0.3 ber as δT /H = 1/(2Nu)(Chill`a& Schumacher 2012). z u

We compute δT for the OB simulations (also indicated , s

m 0.2 in Figure2), and find very good agreement with those r u estimated from δT /H = 1/(2Nu). The δT for OBbot is roughly ten times thicker than that for OBtop, as the 0.1 Nusselt number in the former is roughly 10 times smaller than that in the latter (see Table1). 0.0 We compute the root-mean-square (rms) temperature 0.0 0.2 0.4 0.6 0.8 1.0 fluctuation θrms defined as z/H q θ (z) = h[T − hT i (z)]2i , (34) rms A,t A,t Figure 4. Vertical profiles of (a) the rms temperature fluc- tuations from the mean temperature profile and (b) the to- and show in Figure4(a) that θ (z) increases with in- rms tal (solid curves) and vertical (dashed curves) rms velocities. creasing distance from the horizontal plates. We observe Velocity profiles are symmetric about the midplane for all that θrms(z) attains a maximum near the edge of the cases, whereas the θrms(z) are symmetric only for the OB thermal BL (Scheel & Schumacher 2016; Pandey 2020), simulations. as a result of the generation and emission of thermal plumes from the thermal BL region. The thickness of side the BL region due to an enhanced turbulent mixing the plumes is similar to the thickness of the thermal BL. of the plumes. The plumes originating from the bottom are hotter than Figure4(a) further exhibits that θrms(z) is symmetric ambient fluid at some distance from the bottom plate. in OB convection due to the inherent up-down symmetry Thus, the decrease of hT iA,t(z) with increasing distance of the flow. For the NOB case, however, it is asymmet- from the bottom plate is caused primarily due the de- ric; the location of the maximum of θrms(z) is closer to creasing temperature of the ambient fluid as the tem- the top plate than to the bottom plate because of the perature of the plumes is nearly unchanging. Therefore, thinner thermal BL at the top. Furthermore, the peak of the disparity between the temperatures of the plumes θrms(z) is stronger near the top plate than near the bot- and the ambient fluid within the bottom BL region in- tom plate in the NOB case due to Tbulk > ∆T/2. As the creases with increasing distance from the plate, which is temperature of the cold plumes within the top thermal reflected as an increasing θrms(z) near the bottom plate. BL is very similar to the temperature of the top plate, A similarly increasing contrast between the tempera- the disparity between the temperatures of the plumes tures of the cold thermal plumes and the ambient fluid and the ambient fluid is larger, thus causing a stronger within the top thermal BL region causes an increase of peak in θrms(z) near the edge of the thermal BL in the θrms near the top plate (Pandey 2020). This tempera- top region. The asymmetric rms temperature profile in ture disparity, and consequently θrms(z), decreases out- Non-Boussinesq low-Prandtl number convection 9 our NOB simulation is similar to that observed in 2D anelastic convection with density stratification (Evonuk 5 & Glatzmaier 2004). We now turn to the properties of the velocity field. 4 We compute the total and vertical rms velocity profiles convective )

z 3

deﬁned respectively as ( diffusive u

q N Total 2 2 2 2 urms(z) = hux + uy + uziA,t(z) , (35) q 2 1 uz,rms(z) = huziA,t(z) , (36) and display urms(z) for all the runs in Figure4(b), 0 which reveals that the velocity profiles are fairly sym- 0.0 0.2 0.4 0.6 0.8 1.0 metric about the midplane for all runs. The magni- z/H tude of urms for NOB at all depths lies between those of the OB simulations. Due to a smaller Prandtl num- Figure 5. Vertical profiles of the convective and diffusive ber, the velocity magnitude is larger for OBbot than heat fluxes for the NOB case. The total heat flux remains for OBtop (Schumacher et al. 2015; Pandey & Verma the same in each horizontal plane, thus exhibiting similarity 2016; Scheel & Schumacher 2017). To further com- with OB convection. pare the velocity components, we also show vertical rms velocity profiles for all the simulations in Figure4(b), NOB run. Moreover, as discussed in Section 2.3, and the persistence of symmetry about the midplane; the volume-averaged diffusive heat transport Jd = the use of temperature-dependent thermal diffusivity −hκ(T )∂T/∂ziV,t in the convective state need not be alone does not break the symmetry of the velocity field. the same as the diffusive flux Jc in the conductive state. We compute the viscous BL widths using the slope Therefore, we compute Jd for each snapshot for the method (Scheel et al. 2012), and find that, unlike dif- NOB simulation and find, nevertheless, that Jd remains ferent δT , δu/H ≈ 0.018 at both the plates for the NOB the same as Jc. We also compute the volume- and simulation. time-averaged viscous and thermal dissipation rates and find that they satisfy the exact relations given by the 4.2. Turbulent transport of heat and momentum Eqs. (25) and (26). It is obvious that the heat entering from the bottom Equation (22) in OB convection simplifies to plate has to go through each horizontal plane and, there- √ Nu = 1 + RaP r hu T i , (37) fore, Nu(z), defined in Eq. (20), must be the same for z V,t each horizontal plane. To see this, we plot the convec- which yields Nu = 3.65 and Nu = 35.8 for the tive and diffusive components of Nu(z) for the NOB OBbot and OBtop simulations, respectively. Moreover, simulation in Figure5, which shows that the diffusive Nu computed using the exact relations in OB convec- heat flux, i.e., −hκ(T )∂T/∂ziA,t/Jc, dominates the to- tion (Shraiman & Siggia 1990) match very well with tal heat transport near the plates. Figure5 also shows those computed using uzT , which also indicates the ad- that the total heat flux is dominated completely by the equacy of the spatial resolution for the OB simulations. diffusive component in the vicinity of the plates, whereas The turbulent momentum transport is quantified us- by the convective component in the region around the ing the Reynolds number, which is computed for the midplane. Interestingly, although the thicknesses of the NOB run as s thermal BLs are different at the top and bottom plates, Ratop Re = urms , (38) the widths of the regions dominated by diffusive heat P rtop flux near the two plates are similar. This may be due with u as the root-mean-square velocity based on the to the similar thicknesses of viscous BLs near the two rms volume- and time-averaged kinetic energy, i.e., plates. For OB simulations too, we observe a similar q contribution of the convective and diffusive fluxes. Thus, 2 2 2 urms = hux + uy + uziV,t . (39) the heat transport profile for the case of temperature- dependent diffusivity remains similar to that in OB con- We also use Eq. (38) to compute the Reynolds num- vection. bers for the OB simulations as the Rayleigh and Prandtl We compute the volume- and time-averaged heat numbers remain the same throughout the simulation do- flux using Eq. (22) and find that Nu = 5.4 for the main. The Reynolds numbers from the above formula 10 A. Pandey, J. Schumacher, & K. R. Sreenivasan are Re = 1104, 345, and 1736 for the NOB, OBtop and 5.1. Characteristic scales of turbulent superstructures OBbot simulations, respectively. The values of Nu, Re, Turbulent superstructures, whose characteristic and urms for all the simulations are also listed in Table1. length scales are larger than the flow depth, are observed Rogers et al.(2003) performed simulations of high- Ra in convection with horizontally-extended domains (Cat- 2D convection using anelastic approximation and found taneo et al. 2001; Rincon et al. 2005; Bailon-Cuba et al. that the scalings of Nu and Re remain nearly similar to 2010; Emran & Schumacher 2015; Pandey et al. 2018; those observed in OB convection. Stevens et al. 2018; Schneide et al. 2018; Fonda et al. 2019; Green et al. 2020; Krug et al. 2020). Their charac- 1 10 [T(z)] teristic spatial (λ) and temporal (τ) scales are functions 108 Pr(z) of P r and Ra (Pandey et al. 2018). The superstructures 100 in the present flows are illustrated by the instantaneous

) 1 temperature ﬁeld in the midplane (Figure7). The Fig-

z 10 ( ) r 7

z ure exhibits alternating ridges of hot upwelling and cold

P 10 ( ,

2 a ) 10 downwelling fluid in the midplane. We observe that the R z ( width of the convection structures, which is half the 10 3 characteristic wavelength, λΘ/2, is nearly similar for 106 10 4 OBbot and NOB, but smaller than for OBtop, consistent with Pandey et al.(2018). Indeed, λ increases with increasing Ra (Pandey et al. 2018; Stevens et al. 2018), 0.0 0.2 0.4 0.6 0.8 1.0 z/H which results in larger λΘ for OBtop. The range of temperatures in Figure7(a) is smaller compared to those in other two panels. This shows that Figure 6. Vertical profiles of the temperature-dependent thermal diffusivity and the resulting Prandtl and Rayleigh the turbulent mixing is very efficient in the bulk region numbers exhibit that they vary sharply only in the region for large Prandtl and Rayleigh numbers. Temperature close to the top plate. distribution of the fluid parcels in Figure7(c) is large indicating the presence of very hot and very cold fluids in the midplane. Unlike those in panels (a) and (c) It is clear from Table1 that the transport of heat and of Figure7, the temperature in panel (b) varies asym- momentum as well as the rms velocity for the NOB sim- metrically about the arithmetic mean of 0.5 for top and ulation lie between the OB values, with some propen- bottom plate temperatures. The mean temperature in sity towards the OB values. To explore this, we show bot panel (b) is approximately 0.82, reaffirming the NOB the vertical profile of κ[T (z)] in Figure6, which varies nature of convection in this case. slowly in the region near the bottom plate and remains The turbulent superstructures in the velocity field, ex- almost unvarying in the bulk region. In the vicinity emplified by the vertical velocity in the midplane in Fig- of the top plate, however, it decreases very sharply to ure8, show similar features for all three simulations. reach κ , the value at the top plate. The resulting top The magnitude of the vertical velocity is largest for Prandtl number P r(z) = ν/κ(z) and Rayleigh num- OB and smallest for OB , which is also clear from ber Ra(z) = γg∆TH3/νκ(z) are also plotted in Fig- bot top the corresponding Reynolds numbers in Table1. ure6: the height-dependent Prandtl number P r(z) re- Following the earlier work of Pandey et al.(2018), mains less than 0.1 over the most of convection layer and we compute the characteristic length scales of the tem- increases sharply only in the vicinity of the top plate. perature and velocity fields from the two-dimensional Thus, the Prandtl number in the bulk region of power spectra on a polar wavevector grid of the verti- the NOB simulation remains approximately constant, cal velocity and temperature fluctuation fields in mid- P r ≈ 0.06. We compute the volume- and time- bulk plane, i.e., of U(x, y) = u (x, y, H/2) and Θ(x, y) = averaged thermal diffusivity hκ(T )i , which yields an z V,t θ(x, y, z = H/2). Here, θ is defined as θ(x, y, z) = effective Prandtl number P r = ν/hκ(T )i ≈ 0.06 eff V,t T (x, y, z) − hT i (z). The temperature and vertical ve- and an effective Rayleigh number Ra ≈ 8 × 105. In A,t eff locity power spectra are defined as (Pandey et al. 2018) most of the flow, these effective values are closer to the corresponding parameters of OBbot. Consequently, the Z 2π global heat and momentum transports, as well as urms, 1 2 EΘ(k) = [Θ(ˆ k, φk)] dφk , (40) are comparable for the NOB and OBbot simulations. 2π 0 Z 2π 5. FLOW STRUCTURES 1 2 EU (k) = [Uˆ(k, φk)] dφk , (41) 2π 0 Non-Boussinesq low-Prandtl number convection 11

Figure 7. Instantaneous temperature ﬁelds in midplane for (a) OBtop, (b) NOB, and (c) OBbot reveal turbulent superstructures. The sizes of convection structures are similar in panels (b) and (c), and are smaller than those in panel (a). The mean temperature in panel (b) is larger than 0.5, which reaﬃrms its essential NOB nature.

Figure 8. Instantaneous vertical velocity field in midplane for (a) OBtop, (b) NOB, and (c) OBbot reveal that the characteristic lengths are similar in all cases. where Uˆ(k, φk) and Θ(ˆ k, φk) are the Fourier trans- et al.(2019). We compute the characteristic temporal forms (Verma et al. 2017) of U(x, y) and Θ(x, y), respec- scale of these superstructures as (Pandey et al. 2018) tively. The characteristic length scales are determined Θ U π(λU /4 + H/2) by finding the wavenumbers kmax and kmax correspond- τ(Ra, P r) = 3 (42) ing to the maximum of EΘ(k) and EU (k), respectively, urms and then using λ = 2π/kΘ and λ = 2π/kU . Θ max U max for all the simulations and find that τ ≈ We compute λ and λ for each snapshot of our sim- U Θ 47T , 30T , 169T for NOB, OB , OB , respec- ulations and average them over many snapshots. We f f f bot top tively. The characteristic scales of the turbulent su- find that for both NOB and OB , λ ≈ λ ≈ 4H, bot U Θ perstructures are also summarized in Table1. whereas for OBtop, λU ≈ 4.7H and λΘ ≈ 8H. Thus, λU is smaller than λΘ for OBtop. This difference in the 5.2. Depth-dependence of flow structures characteristic length scales of the temperature and velocity fields for large Prandtl number RBC is consistent The characteristic length scales of the superstructures with the observations of Pandey et al.(2018) and Fonda in OB convection are similar in the bulk and in the BLs (Pandey et al. 2018), but the superstructures look 12 A. Pandey, J. Schumacher, & K. R. Sreenivasan

Figure 9. Instantaneous temperature field in various horizontal planes for the NOB simulation. Due to decreasing thermal diffusivity with increasing altitude, temperature structures become increasingly finer. The skeleton of the pattern is the same, however, at every altitude, thus yielding approximately the same characteristic length scale over the depth. different in the top and bottom halves for the NOB case. (a) We show this by plotting the temperature field in hori- 10 1 zontal cross-sections at various depths in Figure9. Fig- ) x ures9(a,b) show that the temperature patterns near the a 10 4 m k bottom plate are dominated by coarse (or large-scale) ( E structures but become increasingly finer with increas- / 7 ) 10 z = 0.02H k ing altitude. Because the thermal diffusivity decreases ( z = 0.20H

E z = 0.50H and the effective Prandtl number increases with altitude 10 10 z = 0.80H (given constant viscosity), the Batchelor length scale, z = 0.98H defined as 0 1 2 2 1/4 10 10 10 ηB = (νκ /εu) (43) becomes smaller, reflecting the increased content of (b) small structures in the temperature field as the altitude 10 3 )

(i.e., the height from the bottom plate) increases. It x a

m 7 can also be observed from Figure9 that, even within a k 10 ( OBtop, z = T horizontal plane, the colder (blue) structures are ﬁner E /

) OB , z = H/2 11 top k 10 than the hotter (red) ones. This is because the ther- ( OBtop, z = H T mal diffusivity in the NOB simulation is a function of E OBbot, z = T 15 temperature, which fluctuates within a horizontal plane. 10 OBbot, z = H/2 OB , z = H We compute the rms fluctuation of κ(T ) in each hori- bot T p 2 0 1 2 zontal plane as κrms(z) = h[κ(x) − hκiA,t(z)] iA,t and 10 10 10 k find that κrms(z) in the bulk region varies slowly and remains nearly one half of hκiA,t (not shown here). To characterize these features better, we compute the Figure 10. Thermal variance spectra normalized with their power spectrum Eθ(k) defined so that its integral over maximum value for (a) NOB and (b) OB simulations. Both the entire wavenumber range is the thermal variance on the panels show increasing small-scale variance when moving that plane, i.e., towards the plates. The dashed vertical line in both the panels indicate the cutoff wavenumber kc chosen to separate Z kl small and large scales. 2 Eθ(k)dk = hθ (x, y)iA , (44) 0 acteristic temperature length scale remains unchanged. where kl is the largest wavenumber in the Fourier trans- Figure 10(a) confirms that the total temperature vari- form. We compute Eθ(k) for an instantaneous snap- ance contained in the larger wavenumbers is not a sym- shot of the NOB simulation in each horizontal plane metric function of the distance from the midplane. In and plot after scaling with the peak value of Eθ(k) (Fig- contrast, for OB simulations, the total temperature vari- ure 10(a)). The peak value occurs at exactly the same ance in the larger wavenumbers is smallest in the mid- wavelength at each altitude, confirming that the char- Non-Boussinesq low-Prandtl number convection 13 plane and increases symmetrically towards either plate plate, with asymmetric increase in Rθ as the boundaries (Figure 10(b)). Moreover, due to larger Ra and P r, are approached; Rθ increases by more than an order of the relative spectral content at small scales is larger in magnitude from its minimum as the top boundary is ap- OBtop. proached, whereas this increase is only four-fold towards the bottom boundary.

(a) Figure 11(a) also reveals that Rθ is larger for OBtop at NOB all altitudes than those for NOB and OBbot. This occurs because thermal plumes are thinner for the former case OBtop (larger P r and Ra), and the presence of thinner plumes 10 1 OBbot everywhere contributes to the enhanced variance in the small-scale structures. Perhaps not surprisingly, the Rθ curve for NOB is closer to that for OBbot (OBtop) near 2 10 the bottom (top) plate. We have verified that slightly different choices of kc do not qualitatively alter the conclusions of Figure 11. 0.0 0.2 0.4 0.6 0.8 1.0 The structure of the thermal BL can be explored via (b) the vertical temperature gradient at the top and bot- 0.08 tom plates, as it provides a blueprint of the prominent plume scales (Shishkina & Wagner 2006; Pandey et al. 0.06 2018). The contours of ∂T/∂z at the top and bottom plates for the NOB case, shown in Figure 12, re- u 0.04 veal that the patterns in (∂T/∂z)z=0 (Figure 12(a)) are quite similar to those in the temperature field of Fig- 0.03 ure9(a), but those in ( ∂T/∂z)z=H are different from the patterns in the temperature field even at z = 0.98H 0.02 (Figure9(d)), because of sharp variation in the tem- 0.0 0.2 0.4 0.6 0.8 1.0 perature field near the top plate. Moreover, ∂T/∂z at z/H the top plate contains finer structures than that at the bottom plate, despite the similarity in the skeleton of Figure 11. Ratio of (a) the thermal variances and (b) the their large structures. This is in contrast to OB convec- kinetic energies at the small and large scales. For the OB tion, where both fields appear very similar at the two simulations both the Rθ and Ru are symmetric with respect to midplane, whereas this symmetry is only observed in Ru plates (Pandey et al. 2018). Also, (∂T/∂z)z=H is much for the NOB simulation. larger than (∂T/∂z)z=0, consistent with Eq. (32). We compute the spectra of ∂T/∂z at both To quantify the increase of small-scale structures in the plates, and show them in normalized form, the temperature field with varying altitude, we compute E∂T/∂z(k)/E∂T/∂z(kmax), in Figure 12(c). The charac- the ratio of the total thermal variance contained in the teristic length scale, which corresponds to kmax, is the smaller and larger scales as same for both spectra, so the Figure clearly reveals the dominance of small-scale thermal structures at the top R kl Eθ(k)dk kc plate, consistent with the smaller thermal diffusivity to- Rθ = , (45) R kc wards the top. 0 Eθ(k)dk Finally, we compute the kinetic energy spectrum where k is some cutoff wavenumber chosen to differ- c E (k) in each horizontal plane as entiate the small-scale structures from the larger ones. u Z 2π We chose kc = 10, corresponding to a length scale 1 2 2 2 Eu(k) = [ûx(k, φk) +u ˆy(k, φk) +u ˆz(k, φk)] dφk. `c/H = 2π/kc = 0.62, which is roughly half the width 2π 0 of the convection layer. We compute the ratio Rθ in (46) each horizontal plane and plot it in Figure 11(a). For Hereu î(k, φk) is the two-dimensional power spectrum of all runs, Rθ is smaller in the bulk and increases as the ui(x, y) on a polar wavevector grid. We exhibit the nor- boundaries are approached. For OB simulations, the malized kinetic energy spectra Eu(k)/Eu(kmax) for a few minimum of Rθ occurs near the midplane, with sym- horizontal planes in Figure 13. Figure 13(a) reveals that metric increase in Rθ towards the plates. For the NOB the kinetic energy spectra for the NOB case are sym- simulation, the minimum occurs towards the bottom metric about the midplane, in contrast to the thermal 14 A. Pandey, J. Schumacher, & K. R. Sreenivasan

Figure 12. Contours of the vertical temperature derivative ∂T/∂z at (a) the bottom and (b) the top plates show the enhanced small-scale thermal structures in the top region for the NOB simulation. Panel (c) shows the normalized power spectra of ∂T/∂z at the two plates. respect to the midplane, similar to the thermal variance (a) 0 10 spectra. We also compute the ratio of kinetic energy contained ) 3 x 10 a in small- and large-scale structures for all the simula- m k

( tions as u kl E 6 R / 10 Eu(k)dk ) z = 0.02H kc k R = , (47) ( u z = 0.20H kc u R

E E (k)dk z = 0.50H 0 u 10 9 z = 0.80H and plot them in Figure 11(b), which exhibits that R z = 0.98H u is symmetric about the midplane for all the simulations. 100 101 102 6. DISCUSSIONS AND CONCLUSIONS (b) 10 1 In this paper, we studied non-Oberbeck-Boussinesq convection (NOB case) by incorporating a temperature- ) x a 4 dependent thermal diﬀusivity κ(T ) in the governing

m 10 k

( equations. This is important because the natural con- u OBtop, z = T E / ) 10 7 OBtop, z = H/2 vection, such as that in the Sun, usually occur with k

( OBtop, z = H T u varying molecular transport properties and accompa- E OBbot, z = T 10 nied by a strong stratification, and are thus non- 10 OBbot, z = H/2 OBbot, z = H T Boussinesq (Schumacher & Sreenivasan 2020). Clearly, our model does not capture these strong variations of 100 101 102 k density and pressure. Indeed, the effects of density stratification have been investigated earlier in two- and three-dimensional convection models: The flow struc- Figure 13. Kinetic energy spectra normalized with their ture expands (compresses) the hot rising (cold sink- maximum value for (a) NOB and (b) OB simulations. Eu(k) in both panels exhibit essential symmetry about the mid- ing) plumes (Rogers et al. 2003), thus yielding stronger downflows and weaker upflows (Brandenburg 2016; An- plane, unlike asymmetric Eθ(k) for NOB. The dashed vertical line in both panels indicates the cutoff wavenumber kc ders et al. 2019) and significant heat due to viscous dis- chosen to separate small and large scales. sipation (Currie & Browning 2017). We note that esti- mates in Freytag et al.(2012) or Schumacher & Sreeni- variance spectra which exhibit increasingly finer struc- vasan(2020) show also that a free-fall velocity Uf re- tures when one moves upwards. Energy spectra for OB mains by orders of magnitude below the speed of sound simulations in Figure 13(b) also exhibit symmetry with cs across the solar convection zone and reaches that magnitude only at the surface which provides a further motivation to start on this significantly simpler setup. Non-Boussinesq low-Prandtl number convection 15

We chose κ(T ) as a polynomial in T such that the those in OB convection. This feature may have an im- diffusivity increases with increasing depth, whereas the portant implication for modeling NOB convection. kinematic viscosity was maintained constant. As a re- We found that the characteristic scales of the turbu- sult, the Prandtl and Rayleigh numbers are different in lent superstructures as well as the Nusselt and Reynolds the top and bottom regions. We studied this NOB con- numbers lie between those of the OB simulations, and vection using DNS in a rectangular domain of dimen- are closer to the corresponding quantities for the OBbot sions Lx = Ly = 16H for a κ(T ) such that Ra and P r simulation. This may be because the volume-averaged at the bottom of the flow are 500 times smaller than Prandtl and Rayleigh numbers in the NOB simulation those at the top. For reference, we also performed two remain closer to those of the OBbot simulation. More- simulations of OB convection with Ra and P r fixed at over, the characteristic length scale of the temperature the top and bottom plates of the NOB case. field in the NOB simulation remains the same through- The use of κ(T ) in the governing equations breaks the out the depth, despite the growing small-scale content symmetry of the temperature field about the midplane. with increasing altitude. We found that the bulk temperature is approximately Finally, we stress once more that the present study 0.82, which is much higher than the arithmetic mean 0.5 should be considered as a first step towards more realis- of the top and bottom plate temperatures (Wu & Libch- tic simulations of convection processes in sun-like stars. aber 1991; Ahlers et al. 2006; Skrbek & Urban 2015). Our emphasis here has been on breaking the top-down The thermal BL at the top plate is much thinner than symmetry at low Prandtl numbers that is characteris- that at the bottom plate, which makes it challenging to tic of stellar cases. A further increase of the typically properly resolve the temperature field close to the top strong stratification should be possible by addition of plate. Due to decreasing κ(T ), the flow structures in a temperature dependence of the kinematic viscosity, the temperature field become increasingly finer with in- ν(T ). An additional switch to flux boundary conditions creasing altitude, displaying smaller convective granules could be a further important step to understand the hi- inside bigger ones. We quantified this increase by com- erarchy of convection structures which was emphasized puting the ratio Rθ of the total thermal variances in the e.g. by Cossette & Rast(2016). Our direct numeri- small and large structures, and found that Rθ increases cal simulation data can also be used as a test bed for more towards the top plate than towards the bottom. small-scale eddy viscosity and diffusivity models which We remark that the outcomes of the present work are will allow us to discuss the dependence of a turbulent qualitatively insensitive to the specific form of κ(T ) cho- effective Prandtl number on the molecular Prandtl num- sen here. The conclusions from this work are generally ber. These studies are currently underway and will be valid even if slightly different polynomial orders of κ(T ) reported elsewhere. are considered. The symmetry of the velocity field about midplane, We thank Mark Miesch for many helpful discussions however, is nearly preserved in the NOB case. The rms on the dynamics of solar convection. A.P. acknowl- velocity profiles and the kinetic energy spectra change edges support from the Deutsche Forschungsgemein- symmetrically in the top and bottom region as we move schaft within the Priority Programme “Turbulent Su- away from the midplane. The profiles of convective and perstructures” under Grant DFG-SPP 1881. Comput- diffusive heat fluxes also exhibit symmetry about the ing resources at the Leibniz Rechenzentrum Garching midplane. Thus, the properties of heat transport and (Germany) within the Large Scale Project pn68ni of velocity field in the NOB simulation remain similar to the Gauss Centre for Supercomputing are acknowledged. This work was also supported by NYUAD Institute Grant G1502 “NYUAD Center for Space Science.”

REFERENCES

Ahlers, G., Brown, E., Araujo, F. F., et al. 2006, J. Fluid Bailon-Cuba, J., Emran, M. S., & Schumacher, J. 2010, J. Mech., 569, 409, doi: 10.1017/S0022112006002916 Fluid Mech., 655, 152, doi: 10.1017/S0022112010000820 Ahlers, G., Grossmann, S., & Lohse, D. 2009, Rev. Mod. Bekki, Y., Hotta, H., & Yokoyama, T. 2017, The Phys., 81, 503, doi: 10.1103/RevModPhys.81.503 Astrophysical Journal, 851, 74, Anders, E. H., Lecoanet, D., & Brown, B. P. 2019, The doi: 10.3847/1538-4357/aa9b7f Astrophysical Journal, 884, 65, Brandenburg, A. 2016, The Astrophysical Journal, 832, 6, doi: 10.3847/1538-4357/ab3644 doi: 10.3847/0004-637x/832/1/6 16 A. Pandey, J. Schumacher, & K. R. Sreenivasan

Busse, F. H. 1967, J. Fluid Mech., 30, 625–649, Käpylä,P., Korpi, M., Brandenburg, A., Mitra, D., & doi: 10.1017/S0022112067001661 Tavakol, R. 2010, Astronomische Nachrichten, 331, 73, Cattaneo, F., Lenz, D., & Weiss, N. 2001, The doi: https://doi.org/10.1002/asna.200911252 Astrophysical Journal, 563, L91, doi: 10.1086/338355 Krug, D., Lohse, D., & Stevens, R. J. A. M. 2020, J. Fluid Chandrasekhar, S. 1981, Hydrodynamic and Mech., 887, A2, doi: 10.1017/jfm.2019.1054 Hydromagnetic Stability (New York: Dover) Magic, Z., Collet, R., Asplund, M., et al. 2013, A&A, 557, Chillà,F., & Schumacher, J. 2012, Eur. Phys. J. E, 35, 58, A26, doi: 10.1051/0004-6361/201321274 doi: 10.1140/epje/i2012-12058-1 Markson, R. 1975, Science, 188, 1171, Christensen-Dalsgaard, J., Däppen, W., Ajukov, S. V., doi: 10.1126/science.188.4194.1171 et al. 1996, Science, 272, 1286, Michel, E., Baglin, A., Auvergne, M., et al. 2008, Science, doi: 10.1126/science.272.5266.1286 322, 558, doi: 10.1126/science.1163004 Cossette, J.-F., & Rast, M. P. 2016, Astrophys. J. Lett., Miesch, M. S. 2005, Living Rev. Solar Phys., 2, 1, 829, L17, doi: 10.3847/2041-8205/829/1/L17 doi: 10.12942/lrsp-2005-1 Currie, L. K., & Browning, M. K. 2017, The Astrophysical Miesch, M. S., Brun, A. S., DeRosa, M. L., & Toomre, J. Journal, 845, L17, doi: 10.3847/2041-8213/aa8301 2008, The Astrophysical Journal, 673, 557, Emran, M. S., & Schumacher, J. 2015, J. Fluid Mech., 776, doi: 10.1086/523838 96, doi: 10.1017/jfm.2015.316 Evonuk, M., & Glatzmaier, G. A. 2004, Geophysical & Nelson, N. J., Brown, B. P., Sacha Brun, A., Miesch, M. S., Astrophysical Fluid Dynamics, 98, 241, & Toomre, J. 2014, Solar Physics, 289, 441, doi: 10.1080/03091920410001696126 doi: 10.1007/s11207-012-0221-4 Fischer, P. F. 1997, J. Comp. Phys., 133, 84, Nordlund, A., Stein, R., & Asplund, M. 2009, Living Rev. doi: 10.1006/jcph.1997.5651 Sol. Phys., 6, doi: 10.12942/lrsp-2009-2 Fonda, E., Pandey, A., Schumacher, J., & Sreenivasan, Pandey, A. 2020, J. Fluid Mech. (in press) K. R. 2019, Proc. Natl. Acad. Sci. USA, Pandey, A., Scheel, J. D., & Schumacher, J. 2018, Nat. doi: 10.1073/pnas.1900358116 Commun., 9, 2118, doi: 10.1038/s41467-018-04478-0 Freytag, B., Steffen, M., Ludwig, H.-G., et al. 2012, Journal Pandey, A., & Verma, M. K. 2016, Phys. Fluids, 28, of Computational Physics, 231, 919 , 095105, doi: 10.1063/1.4962307 doi: https://doi.org/10.1016/j.jcp.2011.09.026 Riethmüller,T. L., Solanki, S. K., Berdyugina, S. V., et al. Gastine, T., & Wicht, J. 2012, Icarus, 219, 428 , 2014, A&A, 568, A13, doi: 10.1051/0004-6361/201423892 doi: https://doi.org/10.1016/j.icarus.2012.03.018 Rincon, F., Lignières,F., & Rieutord, M. 2005, A&A, 430, Green, G., Vlaykov, D. G., Mellado, J. P., & Wilczek, M. L57, doi: 10.1051/0004-6361:200400130 2020, J. Fluid Mech., 887, A21, Rincon, F., & Rieutord, M. 2018, Living Rev. Sol. Phys., doi: 10.1017/jfm.2019.1008 15, doi: 10.1007/s41116-018-0013-5 Gudiksen, B. V., Carlsson, M., Hansteen, V. H., et al. 2011, Rogers, T. M., Glatzmaier, G. A., & Woosley, S. E. 2003, A&A, 531, A154, doi: 10.1051/0004-6361/201116520 Phys. Rev. E, 67, 026315, Guerrero, G., Smolarkiewicz, P. K., de Gouveia Dal Pino, doi: 10.1103/PhysRevE.67.026315 E. M., Kosovichev, A. G., & Mansour, N. N. 2016, The Sameen, A., Verzicco, R., & Sreenivasan, K. R. 2008, Astrophysical Journal, 819, 104, Physica Scripta, T132, 014053, doi: 10.3847/0004-637x/819/2/104 doi: 10.1088/0031-8949/2008/t132/014053 Hanasoge, S., Gizon, L., & Sreenivasan, K. R. 2016, Annu. —. 2009, EPL (Europhys. Lett.), 86, 14006, Rev. Fluid Mech., 48, 191, doi: 10.1209/0295-5075/86/14006 doi: 10.1146/annurev-fluid-122414-034534 Horn, S., Shishkina, O., & Wagner, C. 2013, J. Fluid Scheel, J. D., Emran, M. S., & Schumacher, J. 2013, New J. Mech., 724, 175, doi: 10.1017/jfm.2013.151 Phys., 15, 113063, doi: 10.1088/1367-2630/15/11/113063 Hotta, H., Rempel, M., & Yokoyama, T. 2016, Science, 351, Scheel, J. D., Kim, E., & White, K. R. 2012, Journal of 1427, doi: 10.1126/science.aad1893 Fluid Mechanics, 711, 281–305, doi: 10.1017/jfm.2012.392 Howard, L. N. 1972, Annu. Rev. Fluid Mech., 4, 473, Scheel, J. D., & Schumacher, J. 2016, Journal of Fluid doi: 10.1146/annurev.fl.04.010172.002353 Mechanics, 802, 147–173, doi: 10.1017/jfm.2016.457 Hurlburt, N. E., Toomre, J., & Massaguer, J. M. 1984, The —. 2017, Phys. Rev. Fluids, 2, 123501, Astrophysical Journal, 282, 557, doi: 10.1086/162235 doi: 10.1103/PhysRevFluids.2.123501 Non-Boussinesq low-Prandtl number convection 17

Schneide, C., Pandey, A., Padberg-Gehle, K., & Stein, R. F., & Nordlund, A. 1998, The Astrophysical Schumacher, J. 2018, Phys. Rev. Fluids, 3, 113501, Journal, 499, 914, doi: 10.1086/305678 doi: 10.1103/PhysRevFluids.3.113501 Stevens, R. J. A. M., Blass, A., Zhu, X., Verzicco, R., & Lohse, D. 2018, Phys. Rev. Fluids, 3, 041501, Schumacher, J., Bandaru, V., Pandey, A., & Scheel, J. D. doi: 10.1103/PhysRevFluids.3.041501 2016, Phys. Rev. Fluids, 1, 084402, Valori, V., Elsinga, G., Rohde, M., et al. 2017, Phys. Rev. doi: 10.1103/PhysRevFluids.1.084402 E, 95, 053113, doi: 10.1103/PhysRevE.95.053113 Schumacher, J., G¨otzfried,P., & Scheel, J. D. 2015, Proc. Valori, V., Elsinga, G., Rohde, M., Westerweel, J., & Natl. Acad. Sci. USA, 112, 9530, van der Hagen, T. 2019, Exp. Fluids, 60, 143, doi: 10.1073/pnas.1505111112 doi: 10.1007/s00348-019-2789-z Verma, M. K. 2018, Physics of Buoyant Flows (Singapore: Schumacher, J., & Sreenivasan, K. R. 2020, Rev. Mod. World Scientiﬁc) Phys., 92, 041001, doi: 10.1103/RevModPhys.92.041001 Verma, M. K., Kumar, A., & Pandey, A. 2017, New J. Shcheritsa, O., Getling, A., & Mazhorova, O. 2018, Physics Phys., 19, 025012. Letters A, 382, 639 , http://stacks.iop.org/1367-2630/19/i=2/a=025012 doi: https://doi.org/10.1016/j.physleta.2018.01.009 Wray, A. A., Bensassi, K., Kitiashvili, I. N., Mansour, N. N., & Kosovichev, A. G. 2015, Simulations of Stellar Shishkina, O., & Wagner, C. 2006, J. Fluid Mech., 546, 51, Magnetoconvection using the Radiative MHD Code doi: 10.1017/S0022112005007408 ‘StellarBox’. https://arxiv.org/abs/1507.07999 Shraiman, B. I., & Siggia, E. D. 1990, Phys. Rev. A, 42, Wu, X.-Z., & Libchaber, A. 1991, Phys. Rev. A, 43, 2833, 3650, doi: 10.1103/PhysRevA.42.3650 doi: 10.1103/PhysRevA.43.2833 Skrbek, L., & Urban, P. 2015, J. Fluid Mech., 785, 270, Zhang, J., Childress, S., & Libchaber, A. 1997, Phys. doi: 10.1017/jfm.2015.638 Fluids, 9, 1034, doi: 10.1063/1.869198