betaFeedback

WorksPhysics and Astronomy

Measuring Scale-dependent Shape Anisotropy by Coarse-Graining: Application to Inhomogeneous Rayleigh-Taylor Turbulence

Abstract

We generalize the `filtering spectrum' [1] to probe scales along different directions by spatial coarse-graining. This multi-dimensional filtering spectrum quantifies the spectral content of flows that are not necessarily homogeneous. From multi-dimensional spectral information, we propose a simple metric for shape anisotropy at various scales. The method is applied to simulations of 2D and 3D Rayleigh-Taylor (RT) turbulence, which is inhomogeneous and anisotropic. We show that 3D RT has clear shape anisotropy at large scales with approximately 4:34:3 vertical to horizontal aspect ratio, but tends toward isotropy at small scales as expected [2,3,4]. In sharp contrast, we find that RT in 2D simulations, which are still the main modeling framework for many applications, is isotropic at large scales and its shape anisotropy increases at smaller scales where structures tend to be horizontally elongated. While this may be surprising, it is consistent with recent results in [5]; large-scale isotropy in 2D RT is due to the generation of a large-scale overturning circulation via an upscale cascade, while small scale anisotropy is due to the stable stratification resultant from such overturning and the inefficient mixing in 2D.

Introduction

Flows encountered in nature or engineering are often anisotropic. Anisotropy can arise from the driving mechanisms, boundary conditions, or body forces that break rotational symmetry such as those due to gravity, rotation, or magnetic fields. Examples range from the motion of polymer fluids [6, 7] and quantum superfluids [8, 9], to flows in geophysics [10, 11, 12, 13, 14] and astrophysics [15, 16, 17, 18], including free shear flows such as jets and plumes [19], and bounded flows such as in a channel [20, 21].

In this work, we are concerned with shape anisotropy and not with anisotropy due to the vector components of a flow’s velocity. To distinguish the two notions of anisotropy, we shall call the latter vector anisotropy. In the turbulence literature, these two notions of anisotropy are related to so-called directional and polarization anisotropy, respectively [22, 23, 24, 25, 26]. We avoid this terminology since it is suggestive of wave phenomena. Here, we are motivated by more general flow structures that may be spatially localized. Shape anisotropy and vector anisotropy are often correlated, although the former can arise in scalar fields such as density or temperature. Shape anisotropy of the velocity field may be diagnosed for each velocity component separately or by analyzing the flow’s kinetic energy (KE) as we do below.

In canonical turbulence that may be described by Kolmogorov’s theory [27], any anisotropy present at the large scales is expected to decrease at smaller scales at sufficiently high Reynolds numbers, where the flow is statistically isotropic [28]. However, this is not always the case, such as in magnetohydrodynamic turbulence [29, 30, 31, 32] where anisotropy is expected to be more pronounced at smaller scales. For the objective testing of turbulence phenomenologies, and for understanding and modeling complex flows such as those mentioned above, the community has relied on a variety of metrics that quantify anisotropy at different scales.

The choice of anisotropy metrics depends on the scales of interest. Bulk turbulence (single-point) anisotropy is traditionally quantified using invariants of the Reynolds stress tensor [4, 14, 33, 34]. At the smallest scales, anisotropy is often measured from the components of gradients [35, 36, 37]. At intermediate (possibly inertial) scales, anisotropy is often characterized using two-point statistics such as autocorrelation functions, Fourier spectra, or 2nd-order structure functions [25, 36, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48]. The power-law scaling of wavelet spectra has also been used to gauge anisotropy [49, 50]. Our motivation here is to quantify anisotropy over all scales and in flows that may be inhomogeneous.

Fourier analysis is fraught with complications when applied to inhomogeneous fields [1]. Afterall, Fourier modes are not an eigenbasis for arbitrary domains and boundary conditions [51, 52]. Measuring the spectrum via a Fourier transform of the auto-correlation function, sometimes known as the Wiener-Khinchin theorem [51], is not justified in the presence of boundaries or if the field is statistically inhomogeneous such as with a spatially varying mean or autocorrelation. In practical applications, Fourier analysis of inhomogeneous fields (or non-stationary temporal signals) is often performed [53, 54] after removing the ensemble-mean [55], detrending [56, 57], and/or tapering (i.e. windowing) [58, 59]. Doing so removes potentially important components of the dynamics. An emblematic example is the global oceanic circulation, for which it had been asserted since the advent of global satellite altimetry in the 1990s that its wavenumber spectrum’s peak is at scales O(100)O(100) km based on detrended and windowed Fourier analysis (e.g. [60, 61, 62]). It was recently shown [63, 64] that this is untrue and that the spectral peak is in fact at O(104)O(10^{4}) km. The spectral peak and the existence of a power-law scaling over scales >103> 10^{3} km in the oceanic circulation could not have been detected from windowed Fourier analysis because all scales larger than the window size (typically taken to be a few hundred kilometers to avoid continental boundaries and curvature effects) are implicitly removed. These limitations of Fourier analysis exist for many realistic flows, including the Rayleigh-Taylor flows we consider here.

When removing the mean flow before using Fourier analysis, a starting assumption is that identifying the statistical ensemble in an application is well-posed. However, the choice of an ensemble is seldom unique and can have important implications on the conclusions as discussed in the case of thermal convection by Kraichnan [65]. For Earth systems applications, such in the case of oceanic or atmospheric observations, we only have a single realization. Appealing to ergodicity with respect to time to treat the temporal record as an ensemble is formally unjustified (although it is often employed) due to a lack of statistical stationarity: the system is variable over a wide range of time-scales. Even if we overlook the difficulties in choosing an ensemble, and if fluctuations about the ensemble-mean flow were assumed to be statistically homogeneous, scales of the ensemble-mean flow itself would remain inaccessible [66]. Often, there is a lack of scale separation between the mean and fluctuating flow components. In such systems, the ensemble-mean flow has significant spatial (or temporal) variations at the same scales populated by the fluctuating field [64]. This necessitates a self-consistent scale analysis of both the mean and fluctuating contributions to the full non-linear dynamics.

Another tool for analyzing scales is the 2nd-order structure function. It has been a valuable phenomenological tool in turbulence theory, but it requires statistical averaging and is not a formal scale decomposition of a field [67]. Unlike a spectrum, which when integrated yields total energy (Parseval’s relation), a 2nd-order structure function, S2(r)S_2(r), yields total energy in the limit1 of length-scales r→0r \to0 (e.g. [69]). At any scale rr, S2(r)S_2(r) can have significant contributions from all scales larger or smaller than rr. This is unsurprising since for a field such as velocity u(x)u(x), S2(r)=⟨∣δu(x;r)∣2⟩S_2(r) = \langle|\delta u(x;r)|^2\rangle at scale rr is constructed from increments δu(x;r)=u(x+r)−u(x)\delta u(x;r) = u(x+r)-u(x) of separation rr before spatial averaging, ⟨…⟩\langle\ldots\rangle. Increments δu(r)\delta u(r) can have contributions from all scales larger or smaller than rr depending on the regularity (or smoothness) of the field u(x)u(x) [70, 71, 72] (see discussion following eq. (4) in [73]). It is known that the power-law scaling of a 2nd-order structure function, S2(r)∼rαS_2(r) \sim r^\alpha, is related to that of the Fourier spectrum, E(k)∼k−α−1E(k) \sim k^{-\alpha-1}, but only if α<2\alpha< 2, i.e. the scaling relation breaks down if E(k)E(k) is steeper than k−3k^{-3} as a function of wavenumber kk (e.g. [68, 74]). Perhaps less well-known is that the scaling relation between S2(r)S_2(r) and E(k)E(k) also breaks down when α<0\alpha< 0, i.e. E(k)E(k) is shallower than k−1k^{-1} [70, 73, 75]. Figs. A10-A11 in the Appendix provide a demonstration of these facts. Another obvious limitation, shared with Fourier analysis, is that structure functions do not provide spatial information about various scales.

Of relevance to results herein are KE spectra of variable density (VD) flows [76, 77], characterized by significant variations in mass density ρ\rho. When using Fourier analysis, structure functions, or wavelets to analyze KE, ρ∣u∣2/2\rho|\mathbf{u}|^2/2, as a function of scale, KE is often treated as quadratic to ensure the spectrum (i.e. power spectral density) is positive. For example, the Fourier transform of an auxiliary field w≡ρu\mathbf{w} \equiv\sqrt{\rho}\mathbf{u} is performed and then squared such that the spectrum is positive and integrates to ρ∣u∣2/2\rho|\mathbf{u}|^2/2 [78, 79]. This approach, while being mathematically sound on its own, creates difficulties when analyzing the scale-dependent equations governing such KE spectra as shown in [80].

Recently, the so-called ‘filtering spectrum’ was proposed [1] to measure spectral content using straightforward coarse-graining in physical space, which is closely related to the continuous wavelet transform [81, 82]. This permits its application to inhomogeneous flows with complex boundaries and allows us to probe scales of both the mean and fluctuating fields concurrently [64]. The approach has been recently adopted to measure the first global energy spectrum of the oceanic general circulation [63].

The filtering spectrum can be regarded as a generalization of the Fourier spectrum to inhomogeneous fields. The filtering spectrum is an energy-preserving scale decomposition [1] and can represent the non-quadratic KE content at different scales of VD flows as shown in [5]. If the filtering kernel has a sufficient number of vanishing moments, the filtering spectrum follows any power-law scaling that the Fourier spectrum may have (assuming Fourier analysis is possible). In fact, the filtering spectrum converges to the Fourier spectrum when using a kernel with an infinite number of vanishing moments (e.g. the Dirichlet kernel), which is justified only for homogeneous fields given the highly non-local nature of such kernels in xx-space (Figure A1 in Appendix).

Figure A1. (left) Comparing the Gaussian (GℓG_\ell), Boxcar (HℓH_\ell), and Dirichlet (Sinc⁡ℓ\operatorname{Sinc}_\ell) kernels in physical space, with the filter width ℓ=Lx/8=0.4\ell= L_x/8 = 0.4. The Boxcar is exactly zero beyond a distance ℓ/2\ell/2 from the center, the Gaussian decays faster than exponentially, while the Dirichlet is highly non-local and decays as 1/∣x−x0∣1/\lvert x-x_0\rvert away from the center x0=0x_0 = 0. (right) The same kernels shown in Fourier kk-space.

In the appropriate domain, Fourier basis functions satisfy orthogonality, which is a highly prized property. However, the sum of any finite number of these basis functions2 suffers from significant spatial non-locality3. Such spatial non-locality is not a concern when working with homogeneous fields, but presents significant challenges when working with inhomogeneous fields such as in the presence of boundaries (see Figure A1 and associated discussion in Appendix). What has been shown in previous work [1, 5, 63, 64, 84] is the possibility of performing a meaningful scale decomposition of inhomogeneous fields and determine their spectra, satisfying both positive semi-definiteness and energy conservation, without the need for orthogonality.

The filtering spectrum as a method is especially valuable in permitting us to visualize (in physical space) the flow at different scales in a self-consistent manner [85], along with any associated anisotropy as we shall see below. A disadvantage of the filtering spectrum compared to the Fourier spectrum is that it involves smoothing as a function of scale [1]. This is the price paid for gaining spatially local information at different scales and generalizing the notion of a spectrum to non-homogeneous fields. Concurrently exact spatial and scale localization is forbidden by the uncertainty principle [52, 83]. The following section II is a brief review of the 1D filtering spectrum proposed by [1]. In section III, we generalize the filtering spectrum to multiple dimensions and define simple metrics to quantify scale anisotropy. In section IV, we demonstrate our approach using illustrative examples and then apply it to anisotropic inhomogeneous turbulence generated from the Rayleigh-Taylor (RT) instability in 2D and in 3D. The paper closes with a brief summary and practical comments about the approach’s usage.

One-dimensional filtering spectrum

For any field a(x)\mathbf{a}(\mathbf{x}), a coarse-grained or (low-pass) filtered version of this field, which contains spatial variations at scales >ℓ> \ell, is defined in nn-dimensional Euclidean space4 as [70, 87, 88, 89]

a‾ℓ(x)=∫dnr Gℓ(x−r)a(r).(1)\overline{\mathbf{a}}_\ell(\mathbf{x}) = \int\mathrm{d}^{n}\mathbf{r}\,G_\ell(\mathbf{x}-\mathbf{r})\mathbf{a}(\mathbf{r}). \tag*{(1)}

Kernel Gℓ(r)=ℓ−nG(r/ℓ)G_\ell(\mathbf{r}) = \ell^{-n}G(\mathbf{r}/\ell) is the dilated version of the “parent kernel” G(r)G(\mathbf{r}), which is normalized. Gℓ(r)G_\ell(\mathbf{r}) has its main support over a region of diameter ℓ\ell. Operation (1) may be interpreted as a local space average in a region of size ℓ\ell centered at point x\mathbf{x}. It is, therefore, a scale decomposition performed in x-space that partitions length scales in the system into large (≳ℓ\gtrsim\ell), captured by a‾ℓ\overline{\mathbf{a}}_\ell, and small (≲ℓ\lesssim\ell), captured by the residual

aℓ′=a−a‾ℓ.(2)\mathbf{a}'_\ell= \mathbf{a} - \overline{\mathbf{a}}_\ell. \tag*{(2)}

We assume GℓG_\ell is an even function such that, ∫r Gℓ(r) dnr=0\int\mathbf{r}\,G_\ell(\mathbf{r})\,\mathrm{d}^{n}\mathbf{r}=0, which ensures that local averaging is symmetric and operation (1) can be rewritten as

a‾ℓ(x)=∫dnr Gℓ(r)a(x+r).(3)\overline{\mathbf{a}}_\ell(\mathbf{x}) = \int\mathrm{d}^{n}\mathbf{r}\,G_\ell(\mathbf{r})\mathbf{a}(\mathbf{x}+\mathbf{r}). \tag*{(3)}

Coarse-graining is a very general scale-analysis framework and includes Fourier analysis (e.g. [90, 91]) and wavelet analysis (e.g. [92, 93]) as special cases with the appropriate choice of kernel GℓG_\ell. See [1] for further discussion. Coarse-graining also lays the foundational framework of large-eddy simulation (LES) [87, 88, 89, 94, 95]. For inhomogeneous flows such as in the presence of boundaries, it is traditional in LES to filter only along the homogeneous directions (e.g. [96, 97]). This is to prevent commutation errors that arise when deforming the kernel to avoid the boundary [98]. However, at least for the purpose of diagnosing scales, it was shown recently that these considerations are moot if regions beyond boundaries are treated in a manner that is consistent with the boundary conditions satisfied by the dynamics [64, 80, 99, 100]. We will discuss this point in greater detail when applying the method to data from Rayleigh-Taylor simulations.

Since spatial coarse-graining allows for extracting spatial information at varying length scales ℓ\ell, it can be used for calculating the one-dimensional filtering spectrum [1]. As an example, the filtering spectrum for the velocity field u\mathbf{u} is

E‾(kℓ)=ddkℓ{∣u‾ℓ(x)∣22}(4)\overline{E}(k_\ell) = \frac{d}{d k_\ell}\left\{\frac{\left|\overline{\mathbf{u}}_\ell(\mathbf{x})\right|^2}{2}\right\} \tag*{(4)}

where kℓ=L/ℓk_\ell= L/\ell is a ‘filtering wavenumber’, LL is a characteristic length-scale (e.g. domain size), ℓ\ell is the scale being probed, and {…}≡(Volume)−1∫(…) dnr\{\ldots\} \equiv(\mathrm{Volume})^{-1}\int(\ldots)\,\mathrm{d}^n r is spatial averaging in nn-dimensions. Eq. (4) measures the energy density (per wavenumber) at scale ℓ\ell by varying it and probing the associated variations in coarse KE, {∣u‾ℓ(x)∣2}/2\{|\overline{\mathbf{u}}_\ell(\mathbf{x})|^2\}/2, which is the cumulative spectrum at all scales larger than ℓ\ell. The main advantage of this method is that it does not rely on Fourier transforms and, therefore, can be easily applied to non-periodic or homogeneous data. In a periodic domain, Fourier and filtering spectra agree if GℓG_\ell has sufficient vanishing moments. In fact, the two spectra have an explicit relationship expressed by eq. (16) in [1]. Another advantage to the filtering spectrum is its ease of generalization to VD flows [1, 5] where KE, ρ∣u∣2/2\rho|\mathbf{u}|^2/2, is non-quadratic due to variations in the mass density field, ρ\rho. For VD flows, such as Rayleigh-Taylor flows considered in this paper, the KE filtering spectrum is [1]

E‾(kℓ)≡ddkℓ{12∣ρu‾ℓ(x)∣2ρ‾ℓ}(5)\overline{E}(k_\ell) \equiv\frac{d}{d k_\ell}\left\{\frac{1}{2}\frac{\left|\overline{\rho\mathbf{u}}_\ell(\mathbf{x})\right|^2}{\overline{\rho}_\ell}\right\} \tag*{(5)}

which reduces to eq. (4) when ρ\rho is constant. Energy spectra based on Fourier and wavelet transforms, on the other hand, are constrained to quadratic quantities to ensure that the spectrum is positive while also satisfying Parseval’s relation. Therefore, a Fourier or wavelet decomposition usually treats kinetic energy as quadratic by transforming the auxiliary field ρu\sqrt{\rho}\mathbf{u}, even when this treatment obscures the inertial-range dynamics [80]. In contrast, the filtering spectrum conserves energy due to the fundamental theorem of calculus and is not constrained by Parseval’s relation. Figure 5 in [5] shows the density, velocity, and kinetic energy filtering spectra of 2D and 3D RT turbulence obtained with this new approach.

Multi-dimensional filtering spectra and scale anisotropy

While the filtering spectrum in eqs. (4),(5) can be applied to anisotropic fields as was done in [5], it is a one-dimensional spectrum since it does not distinguish between scales along different directions. The contribution from the most energetic scale direction dominates the value of E‾(kℓ)\overline{E}(k_\ell). In order to distinguish the spectral content in different directions, we generalize the above definition to multiple dimensions by using anisotropic kernels.

This is most easily done when considering multi-dimensional filtering kernels that are separable. For example in 3D with Cartesian coordinates,

Gℓ(r)=Gℓ(rx)Gℓ(ry)Gℓ(rz),(6)G_\ell(\mathbf{r}) = G_\ell(r_x)G_\ell(r_y)G_\ell(r_z), \tag*{(6)}

where r=(rx,ry,rz)\mathbf{r} = (r_x,r_y,r_z), which is the case for a Gaussian or Boxcar kernel. For separable kernels, isotropic filtering in eq. (3) reduces to filtering in each direction separately:

u‾ℓ(x)=∫dr Gℓ(r) u(x+r)=∫drx Gℓ(rx)∫dry Gℓ(ry)∫drz Gℓ(rz) u(x+r).(7)\overline{\mathbf{u}}_\ell(\mathbf{x}) = \int\mathrm{d}r\,G_\ell(\mathbf{r})\,\mathbf{u}(\mathbf{x}+\mathbf{r}) = \int\mathrm{d}r_x\,G_\ell(r_x)\int\mathrm{d}r_y\,G_\ell(r_y)\int\mathrm{d}r_z\,G_\ell(r_z)\,\mathbf{u}(\mathbf{x}+\mathbf{r}) . \tag*{(7)}

Therefore, a natural generalization to anisotropic filtering is

u‾ℓ⃗(x)=∫drx Gℓx(rx)∫dry Gℓy(ry)∫drz Gℓz(rz) u(x+r).(8)\overline{\mathbf{u}}_{\vec{\ell}}(\mathbf{x}) = \int\mathrm{d}r_x\,G_{\ell_x}(r_x)\int\mathrm{d}r_y\,G_{\ell_y}(r_y)\int\mathrm{d}r_z\,G_{\ell_z}(r_z)\,\mathbf{u}(\mathbf{x}+\mathbf{r}) . \tag*{(8)}

This definition makes it clear how we can probe length scales along different directions by utilizing a vector of filtering scales, ℓ⃗=(ℓx,ℓy,ℓz)\vec{\ell} = (\ell_x,\ell_y,\ell_z). The “cumulative spectrum”

E(k)≡12{∣u‾ℓ⃗∣2ρ‾ℓ⃗}(9)\mathcal{E}(\mathbf{k}) \equiv\frac{1}{2}\left\{\frac{\lvert\overline{\mathbf{u}}_{\vec{\ell}}\rvert^2}{\overline{\rho}_{\vec{\ell}}}\right\} \tag*{(9)}

at filtering wavevector

k=(kx,ky,kz)≡(L/ℓx,L/ℓy,L/ℓz)(10)\mathbf{k} = (k_x,k_y,k_z) \equiv(L/\ell_x,L/\ell_y,L/\ell_z) \tag*{(10)}

yields the energy content at all scales larger than ℓi\ell_i in the iith-direction. Here, LL is a reference scale common to all directions. Figure 1(a) illustrates the cumulative spectrum in 2-dimensions, E(k1,k2)\mathcal{E}(k_1,k_2), which is shaded in blue and accounts for all energy at filtering wavevectors ∣kx∣<k1\lvert k_x\rvert< k_1 and ∣ky∣<k2\lvert k_y\rvert< k_2. Note that k\mathbf{k} in eq. (10) occupies only the first quadrant, kx,ky,kz≥0k_x,k_y,k_z \ge0. This is because E(k)\mathcal{E}(\mathbf{k}) in eq. (9) is invariant under rigid (Euclidean isometry) transformations of the fields ρ(x)\rho(\mathbf{x}) and u(x)\mathbf{u}(\mathbf{x}), which involve a sequence of rotations, translations, and reflections.

Figure 1. Panel (a): Schematic of extracting the multi-dimensional “filtering spectrum”. Filtering is performed in physical space to disentangle scales, without resorting to Fourier transforms. Filtering wavenumbers (kx,kz)=(L/ℓx,L/ℓz)(k_x,k_z)=(L/\ell_x,L/\ell_z) are only a proxy for length-scales, but may be thought of as Fourier wavenumbers conceptually. The cumulative spectrum E(k1,k2)\mathcal{E}(k_1,k_2) (shaded in blue) yields the energy content at all scales larger than ℓ1\ell_1 in the x-direction and larger than ℓ2\ell_2 in the z-direction, such that ∂kz∂kxE\partial_{k_z}\partial_{k_x}\mathcal{E} evaluated at (k1,k2)(k_1,k_2) yields the spectral energy (density) content at exactly those scales (ℓ1,ℓ2)(\ell_1,\ell_2). It is a generalization of the traditional multi-dimensional Fourier spectrum, 12∣u‾∣2(k1,k2)\frac{1}{2}|\overline{\mathbf{u}}|^2(k_1,k_2), and reduces to it when using a sharp-spectral filtering kernel. Panel (b): Shell (blue) in filtering wavenumber space, ∣k∣∈[k/2,2k]|\mathbf{k}|\in[k/\sqrt{2},\sqrt{2}k]. It has logarithmic thickness ×2\times2 [91]. Its centroid (yellow star) based on the distribution of E‾nD(k)\overline{E}^{\mathrm{nD}}(\mathbf{k}) within the shell (Kshell(k)\mathbf{K}^{\mathrm{shell}}(k) in eq. (13)) gives a measure of anisotropy at scale ℓ=L/k\ell=L/k from its proximity to the diagonal line (dashed orange).

Phase information, which may seem lost in the definition of E(k)\mathcal{E}(\mathbf{k}), is easily retrieved from the spatial field 12∣ρuℓ‾∣2(x)/ρ‾ℓ(x)\frac{1}{2}|\overline{\rho\mathbf{u}_{\ell}}|^{2}(\mathbf{x})/\overline{\rho}_{\ell}(\mathbf{x}) in eq. (9) without spatial averaging, {…}\{ \ldots\}.

From E(k)\mathcal{E}(\mathbf{k}), we can quantify the spectral energy density at any scale ℓ\boldsymbol{\ell} in nn-dimensions by defining the nn-dimensional filtering spectrum as

E‾nD(k)≡∂n∂k1…∂knE(k).(11)\overline{E}^{nD}(\mathbf{k}) \equiv\frac{\partial^{n}}{\partial k_{1}\ldots\partial k_{n}}\mathcal{E}(\mathbf{k}) . \tag*{(11)}

It satisfies energy conservation in a straightforward manner by the fundamental theorem of calculus,

12{∣ρu‾ℓ0∣2ρ‾ℓ0}+∫k0∞dnk E‾nD(k)=12⟨ρ∣u∣2⟩,(12)\frac{1}{2}\left\{\frac{|\overline{\rho\mathbf{u}}_{\ell_{0}}|^{2}}{\overline{\rho}_{\ell_{0}}}\right\}+\int_{k_{0}}^{\infty}\mathrm{d}^{n}\mathbf{k}\,\overline{E}^{nD}(\mathbf{k})=\frac{1}{2}\langle\rho|\mathbf{u}|^{2}\rangle, \tag*{(12)}

where ℓ0=(L/k0,L/k0,L/k0)\boldsymbol{\ell}_{0}=(L/k_{0},L/k_{0},L/k_{0}) is the largest length scale at which the flow is filtered.

Figure 1(a) illustrates the motivation for our definition (11) in 2D. The filtering spectrum E‾2D(k1,k2)\overline{E}^{2D}(k_{1},k_{2}) at filtering wavevector k=(k1,k2)\mathbf{k}=(k_{1},k_{2}) quantifies the spectral energy density by measuring the cumulative spectrum’s response, ∂k2∂k1E(k1,k2)\partial_{k_{2}}\partial_{k_{1}}\mathcal{E}(k_{1},k_{2}), to concurrent scale variations in all directions, Δkx\Delta k_{x} and Δkz\Delta k_{z}, at (k1,k2)(k_{1},k_{2}). Here, the filtering scale vector is (ℓ1,ℓ2)=(L/k1,L/k2)(\ell_{1},\ell_{2})=(L/k_{1},L/k_{2}). The one-dimensional spectrum E‾(kℓ)\overline{E}(k_{\ell}) in eq. (5) is essentially an integral of E‾2D(k)\overline{E}^{2D}(\mathbf{k}) at all filtering wavevectors k=(kx,ky)\mathbf{k}=(k_{x},k_{y}) within a thin shell of width Δk\Delta k,

∣kx2+ky2−kℓ∣<Δk.\left|\sqrt{k_{x}^{2}+k_{y}^{2}}-k_{\ell}\right|<\Delta k .

When mass density ρ\rho is constant, E‾nD(k)\overline{E}^{nD}(\mathbf{k}) is a generalization of the traditional Fourier spectrum, 12∣u^∣2(k)\frac{1}{2}|\widehat{\mathbf{u}}|^{2}(\mathbf{k}), where u^(k)\widehat{\mathbf{u}}(\mathbf{k}) is the Fourier transform of u(x)\mathbf{u}(\mathbf{x}). The filtering spectrum reduces to the Fourier spectrum when using a sharp-spectral filtering (Dirichlet or Bessel) kernel in eq. (8). However, a main advantage of our multi-dimensional filtering spectrum in equation (11) is that it can diagnose bounded or inhomogeneous flows in a straightforward manner. Moreover, as discussed in section II, it allows for measuring the spectrum of KE in VD flows while respecting its nonquadratic nature.

Scale-dependent Metrics for Shape Anisotropy

The multi-D filtering spectrum allows us to quantify anisotropy as a function of length-scale as follows. For any filtering wavenumber kk associated with length scale ℓ\ell, consider the logarithmic shell ∣k∣∈[k/2,2k]|\mathbf{k}|\in[k/\sqrt{2},\sqrt{2}k] in filtering wavevector space shown in Figure 1b. Within the shell, the normalized “scale-dependent centroid” (or first moment) of E‾nD\overline{E}^{\mathrm{nD}} is

Kshell(k)≡{1k∫∣k∣∈[k/2,2k]dnk k E‾nD(k)∫∣k∣∈[k/2,2k]dnk E‾nD(k),0if ∫∣k∣∈[k/2,2k]dnk E‾nD(k)<ϵnumeric.(13)\mathbf{K}^{\mathrm{shell}}(k) \equiv \begin{cases} \frac{1}{k}\frac{\int_{|\mathbf{k}|\in[k/\sqrt{2},\sqrt{2}k]} \mathrm{d}^{n} \mathbf{k}\,\mathbf{k}\,\overline{E}^{\mathrm{nD}}(\mathbf{k})}{\int_{|\mathbf{k}|\in[k/\sqrt{2},\sqrt{2}k]} \mathrm{d}^{n} \mathbf{k}\,\overline{E}^{\mathrm{nD}}(\mathbf{k})}, & \\[2ex] \mathbf{0} & \text{if }\int_{|\mathbf{k}|\in[k/\sqrt{2},\sqrt{2}k]} \mathrm{d}^{n}\mathbf{k}\,\overline{E}^{\mathrm{nD}}(\mathbf{k})<\epsilon^{\mathrm{numeric}}. \end{cases} \tag*{(13)}

Integration is performed in the first quadrant of k\mathbf{k}-space. Definition (13) is dimensionless and is zero if the shell is devoid of energy. We set ϵnumeric=10−14\epsilon^{\mathrm{numeric}} = 10^{-14} in this work to avoid numerical overflow errors. When multiplied by wavenumber, kKshell(k)k\mathbf{K}^{\mathrm{shell}}(k) is the centroid for the shell’s energy as shown in Figure 1b. Note the shells [k/2,2k]\left[k/\sqrt{2},\sqrt{2}k\right] in definition (13) have a width that is constant on a logarithmic (not linear) scale. This is important to capture spatially localized structures at different scales, sometimes called a Littlewood-Payley decomposition [91].

The centroid’s distance from the diagonal in Figure 1b is one measure of anisotropy at each kk, which is equivalent to the contrast between different components, KishellK_i^{\mathrm{shell}}, of Kshell\mathbf{K}^{\mathrm{shell}}. A convenient measure for such contrast is what we call the scale-dependent anisotropy metric (AR) tensor,

ARijshell(k)≡(Kishell)2−(Kjshell)2(Kishell)2+(Kjshell)2,for i,j=1,…,n.(14)\mathrm{AR}_{ij}^{\mathrm{shell}}(k) \equiv\frac{\left(K_i^{\mathrm{shell}}\right)^2-\left(K_j^{\mathrm{shell}}\right)^2}{\left(K_i^{\mathrm{shell}}\right)^2+\left(K_j^{\mathrm{shell}}\right)^2}, \qquad\text{for } i,j=1,\ldots,n. \tag*{(14)}

If Kishell=Kjshell=0K_i^{\mathrm{shell}}=K_j^{\mathrm{shell}}=0, we define ARijshell≡0\mathrm{AR}_{ij}^{\mathrm{shell}}\equiv0. For example, ARzxshell(k)>0\mathrm{AR}_{zx}^{\mathrm{shell}}(k)>0, indicates a dominant concentration of energy within shell kk at small kxk_x and large kzk_z. Therefore, ARzxshell(k)>0\mathrm{AR}_{zx}^{\mathrm{shell}}(k)>0 indicates structures that are horizontally elongated (along the xx-direction) at scale ℓ∼k−1\ell\sim k^{-1} as sketched in Figure 2. The converse is also true, where ARzxshell(k)<0\mathrm{AR}_{zx}^{\mathrm{shell}}(k)<0 indicates vertically elongated structures at scale ℓ∼k−1\ell\sim k^{-1}. The anisotropy metric tensor is antisymmeteric, ARijshell=−ARjishell\mathrm{AR}_{ij}^{\mathrm{shell}}=-\mathrm{AR}_{ji}^{\mathrm{shell}}, and each of its components lies within −1≤ARijshell≤1-1\leq\mathrm{AR}_{ij}^{\mathrm{shell}}\leq1. We shall demonstrate these definitions numerically using simple examples in the following section. In Appendix section B, we present a reduced (and less expensive) scale analysis based on what we call marginal spectra and discuss their limitations.

Figure 2. Anisotropic shapes and their corresponding anisotropy metric AR from eqs. (14),(A-3).

Numerical Implementation

In this section, we measure scale anisotropy using illustrative examples as well as data from Rayleigh-Taylor (RT) turbulence simulations in 3D and 2D. RT turbulence is markedly anisotropic and inhomogeneous, which makes it a good application for our method.

Illustrative Examples

For illustrative purposes, we consider two examples in a 2D periodic domain: a Gaussian scalar field and a Taylor-Green velocity field. Defining these in a periodic domain allows us to use a sharp spectral filtering kernel in Fourier space for comparison. The Gaussian scalar field ϕ(x,z)\phi(x,z) is of the form

ϕ(x,z)=12πσx12πσze−(x−π)22σx2−(z−π)22σz2,(15)\phi(x,z)=\frac{1}{\sqrt{2\pi}\sigma_x}\frac{1}{\sqrt{2\pi}\sigma_z}e^{-\frac{(x-\pi)^2}{2\sigma_x^2}-\frac{(z-\pi)^2}{2\sigma_z^2}}, \tag*{(15)}

where (σx,σz)=(π/32,π/8)(\sigma_x,\sigma_z)=(\pi/32,\pi/8) and x,z∈[0,2π]x,z\in[0,2\pi].

This Gaussian field is anisotropic, elongated in the zz-direction as shown in figure 3(a). It populates a broad band of Fourier wavenumbers (see Appendix Fig. A12).

The Taylor-Green (TG) velocity field

(ux,uz)=(pzsin⁡(pxx)cos⁡(pzz),−pxcos⁡(pxx)sin⁡(pzz)),(16)(u_x,u_z)=(p_z\sin(p_xx)\cos(p_zz),-p_x\cos(p_xx)\sin(p_zz)), \tag*{(16)}

where (px,pz)=(8,32)(p_x,p_z)=(8,32) and x,z∈[0,2π]x,z\in[0,2\pi].

is anisotropic and elongated in the xx-direction as shown in Figure 3(b). It consists of only a single Fourier wavevector (px,pz)(p_x, p_z). Both examples are discretized on a 512×512512 \times512 grid.

Illustrative examples of anisotropic fields

Figure 3. Illustrative examples of anisotropic fields. Left panel shows the Gaussian scalar field in eq. (15). Right panel shows the magnitude of the Taylor-Green velocity in eq. (16), where vortices are elongated along the xx-direction.

Gaussian scalar field

Figure 4 demonstrates the filtering spectrum, E‾2D(kx,kz)\overline{E}^{2D}(k_x,k_z), and how to quantify anisotropy for the Gaussian scalar field, ϕ\phi. The cumulative spectrum, E(kx,kz)\mathcal{E}(k_x,k_z), is calculated from eq. (9) by setting ρ=1\rho= 1 and replacing u\mathbf{u} with ϕ\phi. In defining filtering wavenumber, k=L/ℓk = L/\ell, we use the domain size L=2πL = 2\pi. The filtering kernel in eq. (8) is itself a normalized Gaussian of the form

Figure 4. Filtering spectra of the Gaussian scalar field (eq. (15)) using Gaussian filters (eq. (17)). Panel (a) is the cumulative spectrum in 2D filtering wavenumber space. Panel (b) is the 2D filtering spectrum obtained via eq. (11) (note the logarithmic color bar). Panel (c) shows the scale-dependent centroid obtained from integrating E‾2D\overline{E}^{2D} within a shell (eq. (13)). Panel (d) shows the anisotropy metric from eqs. (14).

Gℓ(rx,rz)=(6πℓx2)1/2e−6∣rx/ℓx∣2(6πℓz2)1/2e−6∣rz/ℓz∣2.(17)G_{\ell}(r_x,r_z) = \left(\frac{6}{\pi\ell_x^2}\right)^{1/2} e^{-6\lvert r_x/\ell_x\rvert^2}\left(\frac{6}{\pi\ell_z^2}\right)^{1/2} e^{-6\lvert r_z/\ell_z\rvert^2}. \tag*{(17)}

Figure 4(a) shows the cumulative spectrum, E(kx,kz)\mathcal{E}(k_x,k_z), as a function of filtering wavevector (kx,kz)=(L/ℓx,L/ℓz)(k_x,k_z) = (L/\ell_x,L/\ell_z). It is clear from Figure 4(a) that E(kx,kz)\mathcal{E}(k_x,k_z) is asymmetric with respect to the diagonal line, kx=kzk_x = k_z. E(kx,kz)\mathcal{E}(k_x,k_z) saturates more quickly with increasing kzk_z than

with increasing kxk_x. Equivalent, albeit more intuitive information is conveyed by the filtering spectrum, E‾2D(kx,kz)\overline{E}^{2D}(k_x,k_z), in Figure 4(b). It shows that E‾2D(kx,kz)\overline{E}^{2D}(k_x,k_z) is asymmetric with respect to the diagonal, kx=kzk_x = k_z, with more weight present at smaller kzk_z (i.e. larger vertical length scales ℓz∼kz−1\ell_z \sim k_z^{-1}). This is consistent with the elongated profile of ϕ\phi along the zz-direction relative to the xx-direction. Note that the asymmetry in Figure 4(b) is only over an intermediate range of scales. Symmetry seems to be recovered in the lower-left and upper-right corners of panel 4(b), which we shall now discuss.

Symmetry in the lower-left corner of Figure 4(b) is consistent with isotropy at the largest scales, comparable to the domain size, where filtering in eq. (8) approaches a domain average, ϕ‾ℓ⃗→{ϕ}\overline{\phi}_{\vec{\ell}} \to\{\phi\} as ℓ⃗→(∞,∞)\vec{\ell} \to(\infty,\infty). The spatial mean, {ϕ}\{\phi\}, is uniform and, thus, trivially isotropic. Figure 5 demonstrates this in physical space, which also highlights the advantages of coarse-graining. It may be helpful to think of a length scale decomposition as an operation analogous to removing one’s eyeglasses, blurring an image isotropically. If the blurring is only slight (small ℓ\ell in filtering kernel GℓG_\ell) as in Figure 5b, one can still detect any anisotropy present at the large scales. However, such anisotropy is undetectable if the blurring is severe (ℓ→∞\ell\to\infty) as in Figure 5a. Figure 5c,Figure 5d show how anisotropy of the field ϕ\phi in eq. (5) persists at smaller scales, which is explained by simple analysis in Appendix section I. How can this be reconciled with the seeming symmetry in the upper-right corner of Figure 4(b)?

Figure 5. Scale-dependent anisotropy demonstrated through the low-pass (ϕ‾ℓ\overline{\phi}_{\ell}) and high-pass (ϕℓ′\phi'_{\ell}) filtered field in eq. (15) (see Figure 3a). We use an isotropic Gaussian filtering kernel defined in eq. (17), with ℓx=ℓz=ℓ\ell_x = \ell_z = \ell. Note that ϕ\phi is isotropic at very large scales and becomes increasingly anisotropic at smaller scales, with the smallest scales in panel (d) exhibiting the most anisotropy among those shown here. These panels highlight the advantages of coarse-graining over Fourier space analysis or structure functions.

Symmetry in the upper-right corner of Figure 4(b) is merely due to the finite grid resolution, which precludes fully resolving shells beyond a certain wavenumber. Indeed, the closer we approach the upper-right corner along the diagonal kx=kzk_x = k_z, the less of the shell is resolved on either side of the diagonal. This is related to having isotropic numerical grid cells at the smallest scales. The analytical (non-discretized) ϕ\phi exhibits persistent anisotropy to arbitrarily small scales as explained in Appendix section I.

Figure 4c evaluates centroid components Kxshell(k)K_x^{\mathrm{shell}}(k) and Kzshell(k)K_z^{\mathrm{shell}}(k) from eq. (13). It shows that Kzshell≤KxshellK_z^{\mathrm{shell}} \leq K_x^{\mathrm{shell}} over the entire range of kk, which implies that the field is elongated along the zz-direction at all scales. This is also expressed by the anisotropy metric in Figure 4d, where ARzxshell≤0AR_{zx}^{\mathrm{shell}} \leq0 at all scales. As we discussed above, at the largest length scales (k→0k \to0), we see that Kzshell≈KxshellK_z^{\mathrm{shell}} \approx K_x^{\mathrm{shell}} and ARzxshell≈0AR_{zx}^{\mathrm{shell}} \approx0, indicating isotropy at very large scales. A similar behavior occurs as k→∞k \to\infty, but this is due to numerical resolution limits.

Figure 6 displays similar results as in Figure 4 but using a sharp spectral filter. This is defined in Fourier space as

Figure 6. Same as in Fig. 4 but filtering is done with a sharp-spectral kernel in Fourier space.

Sinc⁡^k⃗c(kx,kz)={1,if ∣kx∣<kc,x and ∣kz∣<kc,z,0,otherwise,(18)\widehat{\operatorname{Sinc}}_{\vec{k}_c}(k_x,k_z)= \begin{cases} 1, & \text{if } |k_x|<k_{c,x}\text{ and }|k_z|<k_{c,z},\\ 0, & \text{otherwise,} \end{cases} \tag*{(18)}

which is straightforward to use due the domain’s periodicity in our example. In eq. (18), k⃗c\vec{k}_c is the wavevector filtering cutoff. Sinc⁡^k⃗c\widehat{\operatorname{Sinc}}_{\vec{k}_c} can be written in xx-space as a product of Dirichlet kernels [101] (see Appendix Fig. A1).

Comparing Figure 4,Figure 6 shows that the results are broadly consistent. Some notable differences are that plots in Figure 6 are (i) sharper as a function of kk and (ii) decay more rapidly at large kk.

Differences in (i) sharpness arise from the filtering spectrum E‾(kℓ)\overline{E}(k_\ell) being essentially a weighted average of the Fourier spectrum E(k)E(k) over a range of Fourier wavenumbers kk

centered around kℓk_\ell [1],

E‾(kℓ)=∫0∞dp[ddkℓ∣G^(pkℓ)∣2]E(p).(19)\overline{E}(k_\ell)=\int_{0}^{\infty}\mathrm{d}p\left[\frac{\mathrm{d}}{\mathrm{d}k_\ell}\left|\widehat{G}\left(\frac{p}{k_\ell}\right)\right|^2\right]E(p). \tag*{(19)}

Here, G^\widehat{G} is the Fourier transform of filtering kernel GG. For a sharp spectral filtering kernel, the factor in square brackets in eq. (19) is a delta function, yielding E‾=E\overline{E}=E. Therefore, a comparison between Figure 6 and Figure 4 highlights that spatially localized filtering kernels such as the Gaussian are not strictly local in kk-space compared to a sharp-spectral filter, which can lead to additional smoothing as a function of scale. This is a necessary price for gaining spatially local information (e.g. Figure 5, showing the field in xx-space at different scales), since concurrently exact spatial and scale localization is forbidden by the uncertainty principle.

The difference in (ii) the decay rate at large kk between Figure 6 and Figure 4 was discussed in detail in [1] and is a consequence of eq. (19). It is due to using the first-order kernel in eq. (17) (using the terminology of [1]) to calculate the filtering spectrum, which precludes E‾(kℓ)\overline{E}(k_\ell) decaying faster than kℓ−3k_\ell^{-3} as kℓ→∞k_\ell\to\infty. In contrast, the Fourier spectrum decays rapidly, E(k)∼e−k2σ2E(k)\sim e^{-k^2\sigma^2}, for the Gaussian field ϕ\phi in eq. (15) (see Appendix section I). The slower kℓ−3k_\ell^{-3} decay is not a limitation of the filtering spectrum method, but of the filtering kernel itself. Faster decay rates can be captured by using higher-order kernels [1] (Appendix Figure A7). Further discussion of this issue along with a comparison between Fourier and filtering spectra are provided in Appendix section I. Despite differences in the decay rate, we see that anisotropy inferred from spatial filtering in Figure 4 is consistent with that in Figure 6 from Fourier analysis. Therefore, even when using a first-order kernel, the underlying anisotropy is still detected from the filtering spectrum E‾nD(kx,kz)\overline{E}^{\mathrm{nD}}(k_x,k_z), although it may be underestimated.

Figure A7. The Boxcar, first- and third-order Gaussian filtering kernels of width ℓ=Lx/8=0.4\ell=L_x/8=0.4.

Taylor-Green

Figure 7 shows the filtering spectrum, E‾2D(kx,kz)\overline{E}^{2D}(k_x,k_z), and anisotropy for the Taylor-Green flow (eq. (16)). For clarity, we use the sharp spectral kernel (eq. (18)) for filtering since the flow is localized at a single Fourier wavevector. We provide a corresponding Figure A15 in the Appendix using the Gaussian filter for completeness.

Cumulative spectrum $\mathcal{E}(k_x,k_z)$

(a) cumulative spectrum E(kx,kz)\mathcal{E}(k_x,k_z)

2D filtering spectrum $\overline{E}^{2D}(k_x,k_z)$

(b) 2D filtering spectrum E‾2D(kx,kz)\overline{E}^{2D}(k_x,k_z)

Scale-dependent centroid

(c) scale-dependent centroid

Scale-dependent anisotropy metric

(d) scale-dependent anisotropy metric

Figure 7. Similar to Figure 4 but analyzing the Taylor-Green velocity (eq. (16)) using a sharp-spectral filter in Fourier space. The only Fourier mode present is (kx,kz)=(8,32)(k_x, k_z) = (8, 32), seen in panel (b). Panels (c) and (d) show the centroid (eqs. (13),(A-2)), but with the integrals in eqs. (13),(A-2) done over thin bands k<∣k∣≤k+1k < |\mathbf{k}| \le k + 1 to highlight the precise wavenumber locations. A complementary Figure A14 in the appendix shows integrals over dyadic bands (k/2,2k](k/\sqrt{2}, \sqrt{2}k] as in eqs. (13),(A-2).

Figure A15. Similar to Figure 7 for the Taylor-Green velocity field but using the Gaussian filtering kernel. Panels (c) and (d) use bands k<∣k∣≤k+1k < |\mathbf{k}| \leq k + 1 as in Figure 7.

In Figure 7a, the cumulative spectrum, E(kx,kz)\mathcal{E}(k_x,k_z) from eq. (9), is calculated by setting ρ=1\rho=1. In defining filtering wavenumber, k=L/ℓk=L/\ell, we use the domain size L=2πL=2\pi. It is clear from Figure 7a,(Figure 7b) that E(kx,kz)\mathcal{E}(k_x,k_z) and E‾2D(kx,kz)\overline{\mathcal{E}}^{2\mathrm{D}}(k_x,k_z) are asymmetric with respect to the diagonal line, kx=kzk_x=k_z, with more weight toward smaller kxk_x and larger kzk_z. In fact, E‾2D(kx,kz)\overline{\mathcal{E}}^{2\mathrm{D}}(k_x,k_z) is zero everywhere except at (kx,kz)=(8,32)(k_x,k_z)=(8,32) as defined in eq. (16).

Figure 7c evaluates the normalized centroid Kshell\mathbf{K}^{\mathrm{shell}} from eq. (13). Note that unlike in eq. (13), here we use thin bands ∣k∣∈(k,k+1]|\mathbf{k}|\in(k,k+1] for clarity, since our field comprises of nonlocalized plane waves. For completeness, Figure A14 in the Appendix uses the dyadic bands as in eq. (13). Figure 7c shows a spike in Kshell\mathbf{K}^{\mathrm{shell}} at k=31k=31 with Kxshell=0.23KzshellK_x^{\mathrm{shell}}=0.23K_z^{\mathrm{shell}} as expected, where the field is elongated along the xx-direction. This is also expressed by the anisotropy metric in Figure 7d, where ARzxshell(31)=0.90\mathrm{AR}_{zx}^{\mathrm{shell}}(31)=0.90.

Scale-dependent centroid and anisotropy metric plots for the Taylor-Green velocity with a sharp-spectral filter

Figure A14. Similar to Figure 7 for the Taylor-Green velocity (eq. (16)) with a sharp-spectral filter in Fourier space but using the dyadic bands k∈[k/2,2k]\mathbf{k} \in[k/\sqrt{2},\sqrt{2}k]. Unlike Figure 7, no shift of the lines is performed.

Rayleigh-Taylor Turbulence

Our method is now applied to a fluid dynamics problem, that of Rayleigh-Taylor (RT) turbulence simulated in both 2D and 3D. RT turbulence occurs when a heavy fluid is supported from below by a light fluid in the presence of gravity or an equivalent acceleration field. An initial small perturbation at the fluid interface grows due to the buoyancy-driven RT instability. It is a fluid dynamics problem that is of both fundamental interest and practical importance in many fields. For example, RT instability plays a leading role in the propagation of the thermonuclear flame front in supernova explosions [102], and is a main obstacle degrading energy yields in inertial confinement fusion (ICF) [103].

It may be apparent to many fluid dynamics experts that modeling RT unstable 3D systems using 2D simulations is inferior to using 3D simulations. Yet, such propositions remain vague and speculative without clearly identifying the precise aspects of the dynamics that are misrepresented in a 2D simulation. This is necessary to convince and guide practitioners who are focused on global system modeling while facing finite computational resources. For example, ICF modelers aim to predict or postdict results from several shots (i.e. experiments) per day in the laboratory. Conducting a single 3D simulation can take several weeks on today’s high-end supercomputers, even without incorporating important system components such as the hohlraum and laser physics [104, 105]. The trade-off ICF modelers face is between including more physical processes in simulations at the expense of conducting them in 2D, despite the recognition that 3D modeling is preferable. To many, trade-offs between 2D and 3D flow physics are not as clear as trade-offs between whether or not to include components that are critical to the system’s evolution such laser deposition, hohlraum physics, or resolving small-scale effects such as from the tent or stalk-mount [106]. This is why in ICF, 2D simulations are still the main “work horse” for experimental design [107, 108, 109] as routine 3D simulations are prohibitively expensive. In the hope of improving 2D modeling, we shall present clear evidence of the stark differences in the fundamental RT scale physics between 2D and 3D models.

At high Reynolds numbers and at sufficiently late times, the system will develop into turbulence [4, 77, 110]. RT turbulence is inhomogeneous, anisotropic, and unsteady. Since RT turbulence is inhomogeneous in the vertical zz-direction (along gravity), it is a good representative of the challenges faced when analyzing scales in inhomogeneous flows. Previous work had been restricted to analyzing scales in the horizontal directions, which are homogeneous [4, 110]. An important exception is the work of Soulard and Griffond [55] who developed a successful theory for the anisotropic spectral scaling in the inertial range of RT flows. They employed Fourier analysis on the fluctuations in all directions, including the inhomogeneous direction, after removing Reynolds averages. The Soulard-Griffond RT theory applies to scales that are sufficiently small such that the nonlinear term dominates over buoyancy and anisotropy can be regarded as a perturbation around a Kolmogorov-Obukhov equilibrium spectrum [2, 43]. Soulard and Griffond [55] tested their theory using direct numerical simulations of RT and calculated Fourier transforms of the fluctuating fields along all directions, including the inhomogeneous vertical direction [66].

While we agree with [55] that Fourier analysis can be well-posed for a component of the RT flow (at times when mixing zone is sufficiently small such that the fluctuating fields are at sufficiently small scales relative to the vertical domain size), our goal here is to demonstrate an approach that generalizes Fourier analysis and can be used to diagnose scales of the entire RT flow, including the Reynolds mean fields. As mentioned in the introduction, there is often a lack of scale separation between the mean and fluctuating fields [5, 64], necessitating a characterization of spectral information for both. Moreover, Fourier analysis, due to its associated Dirichlet kernel that is highly nonlocal in xx-space (Figure A1 in Appendix) presents additional difficulties to scale analysis of inhomogeneous flows. Scales have to be limited to sufficiently large Fourier wavenumbers kk in the vertical and to sufficiently early times in the RT mixing layer growth to avoid domain boundary artifacts [55].

Recently, reference [5] measured the 1-dimensional filtering spectra (see Figure A9 below) for RT turbulence, including scales in the vertical inhomogeneous direction, but did not distinguish the spectral contribution from different directions. Here, we examine shape anisotropy as a function of scale. In addition, reference [5] showed the limitation of 2D RT simulations as surrogates for 3D configurations from the perspective of energy transfer across scales. Here we shall extend the comparison between 2D and 3D RT from the perspective of scale-dependent anisotropy.

Figure A9. Left panel: The shell-integrated (or 1D spectrum in eq. (5)) kinetic energy spectra associated with the 2D and 3D RT data used in the paper. Right panel: from 3D-RT data, we show the shell-integrated kinetic energy spectra using three different filtering kernels: first- and third-order Gaussian and a boxcar filter. A k−5/3k^{-5/3} and a k−3k^{-3} scaling is included in each figure.

Visualizations of the simulated flow in 2D and 3D are shown in Figure 8. They convey qualitatively the turbulent nature of these flows (see [5]). The simulations are of the fully compressible Navier-Stokes equation with an ideal gas equation of state. The governing equations are solved in 2D and 3D rectangular domains in Cartesian coordinates. The boundary conditions are periodic in the horizontal directions and no-slip rigid walls in the vertical zz-direction. The equations are solved using a hybrid solver, which uses the pseudospectral scheme in the horizontal and a 6th order compact finite difference scheme in the vertical. Here, we use data from simulations 3D1024 and 2D4096 described in Table 1 in [5], with the Atwood number 0.5 and the domain-size based Reynolds number 13854 and 44562 for 3D1024 and 2D4096, respectively. More details can be found in [5].

Figure 8. Visualizations of the density field from 2D (left) and 3D (right) Rayleigh-Taylor turbulence. Adapted from Figure 2 of reference [5], with permission from Cambridge University Press.

As discussed in Appendix section B, extracting the 3-dimensional cumulative spectrum, E(k)\mathcal{E}(\mathbf{k}) in eq. (9), can be computationally expensive. For expediency when analyzing the 3D RT data, we probe the following 2-dimensional scale sub-spaces, (ℓx,0,ℓz)(\ell_x,0,\ell_z) and (ℓx,ℓy,0)(\ell_x,\ell_y,0), separately. For example, analyzing (ℓx,0,ℓz)(\ell_x,0,\ell_z) is accomplished by convolving with kernel G⃗ℓ(x)=Gℓx(x) δ(y) Gℓz(z)\vec{G}_{\boldsymbol{\ell}}(\mathbf{x}) = G_{\ell_x}(x)\,\delta(y)\,G_{\ell_z}(z). In this case, no scale decomposition is performed in the yy-direction, but the convolution is done at every location (x,y,z)(x,y,z) in 3D physical space.

In this section, we use the Gaussian kernel in eq. (17). For the definition of filtering wavenumbers, k=L/ℓk = L/\ell, we use L=3.2L = 3.2, which is the domain size along the horizontal direction. In simulations with non-periodic boundary conditions, such as our RT flows with rigid walls at the top and bottom, filtering near the boundary requires a choice for the fields beyond the boundary. Following past work [5, 80, 99], we extend the domain beyond the physical boundaries in a manner consistent with the boundary conditions. For our RT problem, velocity is kept zero beyond the walls, density field is kept constant (zero normal gradient), and the extended pressure field satisfies the hydrostatic condition dP/dz=−ρgdP/dz = -\rho g.

The coarse-graining operation (3) can then be performed at every point in the flow domain.

3D RT

To quantify scale anisotropy in 3D RT, we first calculate E(kx,∞,kz)\mathcal{E}(k_x,\infty,k_z) and E‾2D(kx,kz)\overline{E}^{2\mathrm{D}}(k_x,k_z) in filtering wavenumber subspace (kx,kz)(k_x,k_z). These are shown in Figure 9. The asymmetry relative to the diagonal kz=kxk_z = k_x is especially clear from E‾2D\overline{E}^{2\mathrm{D}} in Figure 9b at wavenumbers smaller than ≈20\approx20, where E‾2D(kx,kz)\overline{E}^{2\mathrm{D}}(k_x,k_z) is skewed toward smaller kzk_z (i.e. larger vertical length scales ℓz∼kz−1\ell_z \sim k_z^{-1}). This indicates that the flow is elongated in the vertical, which is consistent with the notion of large scale vertically inter-penetrating bubbles and spikes. Further evidence for large scale anisotropy is derived from Figure 10, where vertically elongated shapes are visually obvious in panels (a)-(c) showing the flow at scales larger than ℓ=L/10\ell= L/10. Figure 10 underscores the utility of the coarse-graining approach to analyzing anisotropy in a natural and intuitive manner.

Figure 9. 3D-RT filtering spectra of KE along the xx-zz directions using Gaussian kernels (eq. (17)). Panel (a) is the cumulative spectrum (eq. (9)) in filtering wavenumber subspace (kx,kzk_x,k_z). Panel (b) is the associated 2D filtering spectrum obtained via eq. (11) (note the logarithmic color bar). Panel (c) shows the scale-dependent centroid obtained from integrating E‾2D\overline{E}^{2D} within a shell (eq. (13)). Panel (d) shows the anisotropy metric from eq. (14). Note the asymmetry relative to the diagonal in panels (a),(b), associated with vertically elongated flow structures at large scales (small kk). The asymmetry seems to become less at smaller scales (larger kk), which is made clear in panel (c) where the centroid moves closer to the diagonal at high kk (Kx≈KzK_x \approx K_z). This is also reflected in the anisotropy metric in (d), where AR→0AR \to0 as k→∞k \to\infty.

Figure 10. 3D-RT at different scales. Visualizations show a 2D slice at the mid-plane y=L/2y = L/2, where LL is the domain extent in the horizontal direction. Panels (a),(b) show the coarse-grained density, ρˉℓ\bar{\rho}_{\ell} in (a), and kinetic energy, E(kx,kz)\mathcal{E}(k_x,k_z) in (b), at scale kx=kz=L/ℓ=10k_x = k_z = L/\ell= 10. The filtering kernel is a Gaussian (eq. (17)) with ℓx=ℓz=ℓ\ell_x = \ell_z = \ell. Panel (c) shows ρˉℓ\bar{\rho}_{\ell} from panel (a) in the background and the streamlines based on the coarse-grained velocity field ((ρux)ℓ/ρˉℓ,(ρuz)ℓ/ρˉℓ)\left((\rho u_x)_{\ell}/\bar{\rho}_{\ell},(\rho u_z)_{\ell}/\bar{\rho}_{\ell}\right) with filtering as in panels (a),(b) at k=10k = 10. Panel (d) shows the small-scale (or fine) KE, E<ℓ\mathcal{E}_{<\ell} defined in eq. (20), with kx=kz=100k_x = k_z = 100. Anisotropy at large scales is evident in panels (a)-(c) while small scales seem to regain a semblance of isotropy.

The asymmetry relative to the diagonal is less clear at higher wavenumbers in Figure 9b. Indeed, visualizing the corresponding small-scales in physical space in Figure 10d suggests that they are in approximate isotropy. The quantity shown in Figure 10d is energy at scales smaller than ℓ\ell and defined as

E<ℓ⃗=12{(ρ∣u∣2)ℓ⃗‾−∣ρu‾ℓ⃗∣2ρ‾ℓ⃗},(20)\mathcal{E}_{<\vec{\ell}} = \frac{1}{2}\left\{\overline{(\rho|\mathbf{u}|^2)_{\vec{\ell}}}-\frac{|\overline{\rho\mathbf{u}}_{\vec{\ell}}|^2}{\overline{\rho}_{\vec{\ell}}}\right\}, \tag*{(20)}

which is positive semi-definite as proved in [1].

Panels (c) and (d) in Figure 9 provide a more quantitative measure of anisotropy. Discrepancy between KxshellK_x^{\mathrm{shell}} and KzshellK_z^{\mathrm{shell}} is clear around k=5k = 5, where their ratio reaches ≈4/3\approx4/3. This discrepancy is better captured by the anisotropy metric ARzxshell(k)AR_{zx}^{\mathrm{shell}}(k) in Figure 9(d), which has an extremum of ARzxshell(5)=−0.27AR_{zx}^{\mathrm{shell}}(5) = -0.27. ARzxshell(k)AR_{zx}^{\mathrm{shell}}(k) trends toward zero at higher kk, which indicates approximate isotropy field at smaller length scales, although it never reaches zero over the dynamic range of scales we resolve in the simulation. This is consistent with the presence of residual small-scale vector anisotropy observed in Livescu et al. [34] (Figure 14 in [34]). It is also consistent with anisotropic (or “directional”) spectra of the velocity fluctuations in Soulard and Griffond [55] (Figure 6-Figure 7 in [55]). We are unsure if this residual anisotropy at high kk in Figure 9(c),(Figure 9d) would persist in the limit of higher Reynolds numbers. Appendix Figure A7-Figure A8 repeat the anisotropy analysis using a Boxcar and a 3rd-order filtering kernel, which in some sense is more similar to a truncated Dirichlet kernel [111], showing that our results are insensitive to the Gaussian kernel we are using. Panels 9(c),(d) show that at scales ℓ\ell approaching the horizontal domain size LL (k=1k=1), we have Kxshell(1)=Kzshell(1)K_x^{\mathrm{shell}}(1)=K_z^{\mathrm{shell}}(1) and ARzxshell(1)≈0AR_{zx}^{\mathrm{shell}}(1)\approx0 indicating isotropy. At first glance, this may seem surprising since we expect RT flows to be highly anisotropic at large scales [34, 55, 110]. However, note that scale ℓ=L\ell=L (k=1k=1) is much larger than the largest bubble or spike in Figure 8. As the coarse-graining scale ℓ⃗\vec{\ell} approaches that of the domain size, spatially local KE, 12∣ρu⃗‾ℓ∣2/ρ‾ℓ\frac{1}{2}\left|\overline{\rho\vec{u}}_{\ell}\right|^{2}/\overline{\rho}_{\ell}, approaches that of a domain average, 12∣{ρu⃗}∣2/{ρ}\frac{1}{2}\left|\{\rho\vec{u}\}\right|^{2}/\{\rho\}. A domain average has trivial shape isotropy because it is spatially constant. We saw this in our illustrative example of an anisotropic Gaussian in Figure 3(a), which is isotropic at scales comparable to the domain size as shown in Figure 5. For 3D-RT, we have a similar situation in Figure 9(d) where ARshell(k=1)≈0AR^{\mathrm{shell}}(k=1)\approx0. This can be seen from the visualization of coarse KE at k=1k=1 in Figure A6 of the Appendix.

Three panels: (a) coarse density $\bar{\rho}_{\ell}$; (b) coarse KE $\mathcal{E}$; (c) coarse streamlines

Figure 14. 2D-RT has a quasi-isotropic large-scale overturning circulation due to an upscale cascade [5], which is absent in 3D-RT. Panels (a)-(b) are similar to those in Figure 13 but filtered at scale kx=kz=L/ℓ=3k_x = k_z = L/\ell= 3. Panel (c) demonstrates clearly the large-scale overturning circulation.

Visualization from 3D-RT of coarse kinetic energy

Figure A6. Visualization from 3D-RT of coarse KE 12ρˉ∣uˉ∣2\frac{1}{2}\bar{\rho}|\bar{\mathbf{u}}|^2 filtered at ℓ=Lx\ell= L_x, corresponding to k=1k = 1.

Figure A8. Centroid (eq. (A-2)) and anisotropic metric (eq. (A-3)) of the 3D Rayleigh-Taylor kinetic energy within a x−zx-z slice. Panels (a),(b) are for first-order kernels adopted in the main paper, panels (c),(d) are the results of top-hat or boxcar filter, while panels (c),(d) are obtained with third-order kernel of eq. A-4.

In contrast to Figure 9, comparing scales in the two horizontal (xx-yy) directions in Figure 11 shows almost perfect isotropy, consistent with physical expectations. We see that E(kx,ky,∞)\mathcal{E}(k_x,k_y,\infty) and E‾2D(kx,ky)\overline{E}^{2\mathrm{D}}(k_x,k_y) are symmetric relative to the diagonal ky=kxk_y=k_x, and ARyxshell=0AR_{yx}^{\mathrm{shell}}=0 for all kk.

Figure 11. (a) cumulative spectrum E(kx,ky,∞)\mathcal{E}(k_x,k_y,\infty); (b) 2D filtering spectrum E‾2D(kx,kz)\overline{E}^{2D}(k_x,k_z); (c) scale-dependent centroid; (d) scale-dependent anisotropy metric. Same as in Figure 9 but in the horizontal xx-yy directions for 3D-RT. The flow is isotropic at all scales as expected.

2D RT

To quantify scale anisotropy in 2D RT, we calculate E(kx,∞,kz)\mathcal{E}(k_x,\infty,k_z) and E‾2D(kx,kz)\overline{E}^{\mathrm{2D}}(k_x,k_z), shown in Figure 12. From panels (a) and (b), the asymmetry relative to the diagonal kz=kxk_z = k_x seems absent at small filtering wavenumbers but becomes discernible at k≳50k \gtrsim50, where E‾2D(kx,kz)\overline{E}^{\mathrm{2D}}(k_x,k_z) is skewed slightly toward smaller kxk_x (i.e. larger horizontal length scales ℓx∼kx−1\ell_x \sim k_x^{-1}). This indicates that the flow is elongated in the horizontal, which may be surprising for Rayleigh-Taylor flows where it is expected that the flow would be elongated in the vertical due to buoyancy forcing [3, 4, 112], similar to what we observed in section IV B 1 in 3D-RT.

Figure 12. (a) cumulative spectrum E(kx,kz)\mathcal{E}(k_x,k_z); (b) 2D filtering spectrum E2D(kx,kz)\mathcal{E}^{2\mathrm{D}}(k_x,k_z); (c) scale-dependent centroid; (d) scale-dependent anisotropy metric. 2D-RT filtering spectra of KE as in Figure 9. Note in panels (a),(b) the symmetry relative to the diagonal at large-scales (small kk), associated with a quasi-isotropic overturning circulation. However, asymmetry emerges at smaller scales (larger kk), which is made clear in panel (c) where the centroid moves away from the diagonal at high kk (Kx<KzK_x < K_z), associated with horizontal layering due to stable stratification. This is also reflected in the anisotropy metric in (d), where AR>0AR > 0 as k→∞k \to\infty.

Panels (c) and (d) in Figure 12 provide a more quantitative measure of anisotropy. We see that Kxshell(1)=Kzshell(1)K_x^{\mathrm{shell}}(1) = K_z^{\mathrm{shell}}(1) at the smallest wavenumbers. The discrepancy between KxshellK_x^{\mathrm{shell}} and KzshellK_z^{\mathrm{shell}} becomes discernible at k>10k > 10. This is also captured by ARzxshell(k)AR_{zx}^{\mathrm{shell}}(k) in Figure 12(d), which is positive and increases at larger kk, indicating that the flow is elongated in the horizontal and becomes increasingly anisotropic at those smaller length scales.

Isotropy at large scales and increasing anisotropy at small scales in 2D RT is consistent with recent results in [5]. In that work, it was shown that 2D RT is characterized by the emergence of a large scale overturning circulation, which isotropizes the 2D flow. This is in contrast to the picture of vertically rising bubbles and sinking spikes often invoked for RT flows, which nevertheless seems consistent with 3D-RT. Direct evidence for the existence of a large scale overturning circulation in 2D RT is provided in Figure 13,Figure 14. The overturning circulation in Figure 13c is quasi-isotropic and is absent at the same large scales (>ℓ=L/10> \ell= L/10) in 3D RT shown in Figure 10c, which are clearly anisotropic. Figure 14 shows even larger scales (>ℓ=L/3> \ell= L/3) in 2D RT, where the large-scale isotropic overturning circulation is clearer. In the Appendix Fig. A5, we calculate the anisotropy metrics at different locations along the vertical direction.

Figure 13. 2D-RT at different scales, as in Figure 10. Panels (a)-(c) are filtered at scale kx=kz=L/ℓ=10k_x = k_z = L/\ell= 10 using a Gaussian kernel. Panel (d) shows small-scale KE at kx=kz=100k_x = k_z = 100. Large-scale flow in panels (a)-(c) is approximately isotropic, which is especially clear in the large-scale overturning circulation in panel (c). In contrast, the small-scale flow in (d) shows a tendency for horizontal layering due to the stable stratification that results from overturning.

It was shown in [5] that 2D RT does not cascade energy to smaller scales, resulting in weaker small inertial scales relative to 3D-RT. Weaker inertial scales and an associated suppressed molecular mixing was also discussed in several previous studies [110, 113, 114, 115].

Therefore, our results in Figure 12 indicate that the combination of a large-scale overturning and suppressed small-scale mixing in 2D-RT yields strong stable stratification at small-scales. Such stratification leads to a flow that is elongated in the horizontal as quantified by the anisotropy metrics in Figure 12c,Figure 12d. Direct visual evidence for the anisotropic flow at small scales is provided in Figure 13d, which should be contrasted to the corresponding Figure 10d in 3D RT.

Conclusions

This work establishes how spatial coarse-graining can be used to quantify anisotropy at different length scales. It is based on generalization of ideas from [1], which we demonstrated via simple illustrative examples and an application to Rayleigh-Taylor turbulence.

In addition to the method, a main result of this work showed that anisotropy of Rayleigh-Taylor turbulence in 2D is fundamentally different from that of 3D-RT. We showed that 2D-RT is characterized by enhanced anisotropy at small scales in a manner that was perhaps unexpected, but in hindsight is consistent with recent work [5]. A clear identification of how 2D modeling of applications involving RT can misrepresent the actual 3D hydrodynamics is a necessary step to develop new modeling strategies without having to conduct fully 3D simulations. This is because 3D hydrodynamics modeling remains prohibitively expensive in many applications such as ICF as discussed above.

We envisage that our method for quantifying shape anisotropy would be especially valuable in flows that are inhomogeneous, where established methods such as Fourier analysis cannot be used in a straightforward manner. In this regard, it is important to be aware of limitations of filtering spectra. In particular, it was shown in [1] that when using first-order kernels such as a Gaussian, filtering spectra do not decay faster than k−3k^{-3} as k→∞k \to\infty, even if there is little or no energy at those small scales when probed with more complex (higher-order) kernels. Therefore, determining whether or not energy conveyed by filtering spectra is physical or a spurious artifact of the kernel being used requires using higher order kernels if a k−3k^{-3} scaling is observed. Regarding anisotropy, the potential for ‘locking’ in at a k−3k^{-3} scaling when using first-order filtering kernels would still allow us to detect anisotropy using the metrics, Kshell(k)\mathbf{K}^{\mathrm{shell}}(k) and ARijshell(k)\mathrm{AR}^{\mathrm{shell}}_{ij}(k), although the level of anisotropy may be underestimated.

Acknowledgments

We thank Olivier Soulard and two anonymous reviewers for their valuable comments that helped improve our paper. This research was funded by US DOE grant DE-SC0020229 and NSF grant PHY-2206380. Partial support from US NSF grants PHY-2020249 is acknowledged. HA was also supported by US DOE grants DE-SC0014318, DE-SC0019329, US NSF grant OCE-2123496, US NASA grant 80NSSC18K0772, and US NNSA grants DENA0003856, DE-NA0003914, DE-NA0004134. DZ also acknowledges financial support from the National Natural Science Foundation of China (No. 12202270). Computing time was provided by NERSC under Contract No. DE-AC02-05CH11231.

APPENDIX

Filtering kernels and Fourier analysis

Figure A1 shows three example kernels that can be used to calculate the filtering spectrum. When using the Dirichlet kernel, the filtering spectrum is exactly the same as the Fourier spectrum. The Dirichlet is highly non-local and decays as 1/∣x−x0∣1/\lvert x-x_0\rvert away from the center at x0x_0, where the filtered field u‾(x0)\overline{u}(x_0) is being evaluated. Therefore, the filtered field at location x0x_0 has contributions from spatial locations xx far away from x0x_0. This is in comparison to filtering with a more localized kernel, such as a Gaussian or Boxcar, which is practically (or exactly) zero beyond a distance ∼ℓ/2\sim\ell/2 away from x0x_0.

Therefore, in a bounded domain, regions outside the boundaries can make non-negligible (possibly spurious) contributions to the Fourier scale analysis (equivalent to filtering with a Dirichlet kernel) even in regions that are presumably far from the boundaries. When using a more compact kernel, such as the Gaussian or Boxcar, the “domain of dependence” is localized to be within ∼ℓ/2\sim\ell/2 from the location being analyzed. This makes it feasible to extend the fields beyond the domain boundaries (in a manner satisfying the boundary conditions) to analyze scales near those boundaries.

Marginal Filtering Spectra

The cumulative spectrum, E\mathcal{E}, or equivalently, the filtering spectrum, E‾nD\overline{\mathcal{E}}^{nD}, evaluated over the entire scale-grid (Figure 1(a)) contains all information about the energy content of spatial scales in different directions. However, E\mathcal{E} can be computationally expensive to calculate using data on NnN^n grid-points in nn-dimensions since it relies on spatial filtering via convolutions. Scanning the entire range of scales in each direction requires nominally O(Nn)O(N^n) different scale vectors (ℓ1,…,ℓn)(\ell_1,\ldots,\ell_n), each of which can cost of O(N2n)O(N^{2n}) operations for the convolution. Therefore, in 3-dimensions, the cost can be as high as O(N9)O(N^9). While there are several algorithmic optimizations to reduce the cost [63, 64, 116], such as subsampling the scale-grid (especially at high kk), the computational cost can still be considerable.

It is possible to quantify anisotropy by sampling a much smaller subset of the scale-grid but at the cost of accuracy and sensitivity. Figure A2(a) sketches what we call the “marginal spectra,” in analogy to the marginal probability distribution. A scale decomposition via filtering is performed along each direction separately rather than concurrently. The marginal spectra along different directions are defined as

Figure A2. Panel (a): similar to Figure 1(a) but for “marginal spectra”, ∂kxE\partial_{k_x}\mathcal{E} and ∂kzE\partial_{k_z}\mathcal{E}. We see that ∂kxE(k1,∞)\partial_{k_x}\mathcal{E}(k_1,\infty) (in red) accounts for field variations at scale ℓ1=L/k1\ell_1=L/k_1 in the xx-direction, but contains all scales in the zz-direction. Similarly, ∂kzE\partial_{k_z}\mathcal{E} (in green) accounts for variations in the zz-direction. Panel (b): (b) similar to Figure 1(b) but for the marginal centroid (Kmar(k)\mathbf{K}^{\mathrm{mar}}(k) in eq. (A-2)), depicted as a star. It is based on E‾mar(k)\overline{\mathcal{E}}^{\mathrm{mar}}(k), which is more easily computed than E‾shell(k)\overline{\mathcal{E}}^{\mathrm{shell}}(k) but at the expense of accuracy since it incorporates all scales in other directions. The restriction on Kmar(k)\mathbf{K}^{\mathrm{mar}}(k) to reside where the two (red and green) strips overlap also makes it a less sensitive gauge of anisotropy.

E‾xmar(k)≡∂∂kxE(kx,∞,∞)∣kx=k,(21)\overline{E}^{\mathrm{mar}}_x(k) \equiv\left.\frac{\partial}{\partial k_x}\mathcal{E}(k_x,\infty,\infty)\right|_{k_x=k}, \tag*{(21)}
E‾ymar(k)≡∂∂kyE(∞,ky,∞)∣kx=k,(22)\overline{E}^{\mathrm{mar}}_y(k) \equiv\left.\frac{\partial}{\partial k_y}\mathcal{E}(\infty,k_y,\infty)\right|_{k_x=k}, \tag*{(22)}
E‾zmar(k)≡∂∂kzE(∞,∞,kz)∣kz=k.(23)\overline{E}^{\mathrm{mar}}_z(k) \equiv\left.\frac{\partial}{\partial k_z}\mathcal{E}(\infty,\infty,k_z)\right|_{k_z=k}. \tag*{(23)}

Here, ∞\infty denotes the absence of filtering (ℓ=L/k=0\ell=L/k=0) along the corresponding direction, thereby retaining the contribution from all scales along that direction. The marginal spectrum E‾xmar(k)\overline{E}^{\mathrm{mar}}_x(k), for example, is obtained from u‾ℓx,0,0(x)\overline{\mathbf{u}}_{\ell_x,0,0}(\mathbf{x}) by convolving with kernel Gℓx(rx)δ(ry)δ(rz)G_{\ell_x}(r_x)\delta(r_y)\delta(r_z) in eq. (8), where the delta function δ(r)=Gℓ=0(r)\delta(r)=G_{\ell=0}(r).

Comparing E‾xmar(k)\overline{E}^{\mathrm{mar}}_x(k), E‾ymar(k)\overline{E}^{\mathrm{mar}}_y(k), and E‾zmar(k)\overline{E}^{\mathrm{mar}}_z(k) gives us a gauge of anisotropy at different scales. However, it is important to be mindful of what information is conveyed in eq. (A-1) and its limitations. When integrated over kk, all three quantities yield the same total energy. Therefore, if E‾zmar(k1)>E‾xmar(k1)\overline{E}^{\mathrm{mar}}_z(k_1)>\overline{E}^{\mathrm{mar}}_x(k_1) at k1k_1, it is guaranteed that E‾xmar(k2)>E‾zmar(k2)\overline{E}^{\mathrm{mar}}_x(k_2)>\overline{E}^{\mathrm{mar}}_z(k_2) for some other k2k_2. In comparison, E‾nD\overline{E}^{\mathrm{nD}} has no similar constraint.

Similar to eq. (13), the normalized scale-dependent centroid of marginal spectra within 1D shells are

Kimar(k)≡{1k∫k/22kdq q E‾imar(q)∫k/22kdq E‾imar(q),for i=1,…,n,0if ∫k/22kdq E‾imar(q)<ϵnumeric.(24)K_i^{\mathrm{mar}}(k)\equiv \begin{cases} \frac{1}{k}\frac{\int_{k/\sqrt{2}}^{\sqrt{2}k}dq\,q\,\overline{E}^{\mathrm{mar}}_i(q)}{\int_{k/\sqrt{2}}^{\sqrt{2}k}dq\,\overline{E}^{\mathrm{mar}}_i(q)}, & \text{for }i=1,\ldots,n,\\[1.2em] 0 & \text{if }\int_{k/\sqrt{2}}^{\sqrt{2}k}dq\,\overline{E}^{\mathrm{mar}}_i(q)<\epsilon^{\mathrm{numeric}}. \end{cases} \tag*{(24)}

We set ϵnumeric=10−14\epsilon^{\mathrm{numeric}}=10^{-14} below. Since E‾imar\overline{E}^{\mathrm{mar}}_i incorporates all scales in other directions as shown in Figure A2(a), the marginal centroid, KmarK^{\mathrm{mar}}, is a less accurate gauge of anisotropy at any one scale. Moreover, its restriction to reside where the two (red and green) strips overlap in Figure A2(b) also makes it a less sensitive gauge of anisotropy.

Similar to eq. (14), we define the anisotropy metric associated with marginal spectra as

ARijmar(k)≡(Kimar)2−(Kjmar)2(Kimar)2+(Kjmar)2,for i,j=1,…,n.(25)\mathrm{AR}^{\mathrm{mar}}_{ij}(k)\equiv\frac{\left(K_i^{\mathrm{mar}}\right)^2-\left(K_j^{\mathrm{mar}}\right)^2}{\left(K_i^{\mathrm{mar}}\right)^2+\left(K_j^{\mathrm{mar}}\right)^2},\qquad\text{for }i,j=1,\ldots,n. \tag*{(25)}

It has an interpretation similar to that in eq. (14), as sketched in Figure 2. However, the anisotropy metrics ARshell\mathrm{AR}^{\mathrm{shell}} and ARmar\mathrm{AR}^{\mathrm{mar}} can be differ quantitatively due to restrictions on Kmar\mathbf{K}^{\mathrm{mar}} explained in Figure A2.

Below, we provide numerical examples to demonstrate how these metrics quantify anisotropy and to highlight their limitations relative to shell-based metrics presented in the main text.

Centroids and anisotropic metric associated with marginal spectra

The anisotropy metric based on the marginal spectra is shown in Figure A3 for the two illustrative examples, and in Figure A4 for the 2D and 3D Rayleigh-Taylor fields.

Gauss centroid, Gaussian filter

(a) Gauss centroid, Gaussian filter

Gauss anisotropy metric, Gaussian filter

(b) Gauss anisotropy metric, Gaussian filter

Gauss centroid, sharp spectral filter

(c) Gauss centroid, sharp spectral filter

Gauss anisotropy metric, sharp spectral filter

(d) Gauss anisotropy metric, sharp spectral filter

TG vortex centroid, sharp spectral filter

(e) TG vortex centroid, sharp spectral filter

TG vortex anisotropy metric, sharp spectral filter

(f) TG vortex anisotropy metric, sharp spectral filter

Figure A3. Marginal spectra averaged centroid (eq. (A-2)) and anisotropic metric (eq. (A-3)) of the Gaussian scalar field using Gaussian filters in panels (a),(b), and of the Gaussian scalar using the sharp spectral filter in panels (c), (d), and of the Taylor-Green vortex using the sharp spectral filter in panels (e),(f).

Six-panel plots of marginal spectra based centroid and anisotropic metric for 3D and 2D Rayleigh-Taylor kinetic energy

Figure A4. Marginal spectra based centroid (eq. (A-2)) and anisotropic metric (eq. (A-3)) of the 3D Rayleigh-Taylor kinetic energy along the xx-zz directions in panels (a),(b), and of 3D RT KE along xx-yy directions in panels (c), (d), and of kinetic energy of 2D RT in panels (e),(f).

Figure A3a plots Kxmar(k)\mathbf{K}_{x}^{\mathrm{mar}}(k) and Kzmar(k)\mathbf{K}_{z}^{\mathrm{mar}}(k) from eq. (A-2) for the Gaussian data using Gaussian filters. Consistent with their Kshell\mathbf{K}^{\mathrm{shell}} analogue shown in Figure 4c, it shows that Kzmar≤Kxmar\mathbf{K}_{z}^{\mathrm{mar}} \leq\mathbf{K}_{x}^{\mathrm{mar}} over the entire range of kk, which implies that the field is elongated along the zz-direction at all scales. We see that Kzmar≈Kxmar\mathbf{K}_{z}^{\mathrm{mar}} \approx\mathbf{K}_{x}^{\mathrm{mar}} in the limits of k→0k \to0 and k→∞k \to\infty. The anisotropy metric ARzxmar(k)\mathrm{AR}_{zx}^{\mathrm{mar}}(k) in panel (b) from eq. (A-3) reflects this behavior and is qualitatively consistent with its analogue ARzxshell(k)\mathrm{AR}_{zx}^{\mathrm{shell}}(k) in Figure 4d, although with seemingly lower sensitivity. This is not surprising given that Kmar(k)\mathbf{K}^{\mathrm{mar}}(k) is derived from a highly reduced spectral representation as discussed in section B (see also Figure 1). The sharp spectral filtering results shown in panels (c) and (d) are similar to the Gaussian kernel results in panels (a) and (b).

Figure A3e and (f) show for the Taylor-Green data the centroid Kmar\mathbf{K}^{\mathrm{mar}} based on marginal spectra, which exhibits a spike in Kxmar\mathbf{K}_{x}^{\mathrm{mar}} at k=7k = 7 and a spike in Kzmar\mathbf{K}_{z}^{\mathrm{mar}} at k=31k = 31. This is because E‾xmar\overline{E}_{x}^{\mathrm{mar}} and E‾zmar\overline{E}_{z}^{\mathrm{mar}} spike at those respective wavenumbers. However, they highlight a limitation of marginal spectra we mentioned; since E‾xmar\overline{E}_{x}^{\mathrm{mar}} and E‾zmar\overline{E}_{z}^{\mathrm{mar}} have to yield the same total energy when integrated, if E‾xmar(k1)>E‾zmar(k1)\overline{E}_{x}^{\mathrm{mar}}(k_{1}) > \overline{E}_{z}^{\mathrm{mar}}(k_{1}) at k1k_{1}, then we necessarily have E‾xmar(k2)<E‾zmar(k2)\overline{E}_{x}^{\mathrm{mar}}(k_{2}) < \overline{E}_{z}^{\mathrm{mar}}(k_{2}) at another wavenumber k2k_{2}. This is also expressed by the anisotropy metric ARijmar(k)\mathrm{AR}_{ij}^{\mathrm{mar}}(k) in panel (f), where ARzxmar(7)=−1\mathrm{AR}_{zx}^{\mathrm{mar}}(7) = -1 and ARzxmar(31)=1\mathrm{AR}_{zx}^{\mathrm{mar}}(31) = 1.

The marginal spectra based centroids and anisotropy metrics for the 2D and 3D RT data are shown in Figure A4. For the shape anisotropy along the xx-zz directions for 3D RT, metrics based on marginal spectra in Figure A4(a),(Figure A4b) show qualitatively consistent results as the shell-based results in Figure 9(c),(Figure 9d), albeit with a much weaker anisotropic signal. This is consistent with our earlier remarks about their lower sensitivity to anisotropy (Figure 1).

After all, E‾mar\overline{E}^{\mathrm{mar}} has significantly less scale information than E‾2D\overline{E}^{2\mathrm{D}}. The marginal spectra based shape anisotropy for 3D RT along the xx-yy direction in Figure A4(d) is zero, indicating the flow is isotropy among the horizontal directions as expected. For 2D RT data in panels (e), (f), metrics based on marginal spectra fail to capture the small-scale anisotropy compared to the shell-based results in Figure 12(c),(Figure 12d), due to their lower sensitivity discussed in the main text.

Local measurement of shape anisotropy

One advantage of our filtering approach is that we can measure the anisotropy associated with local sub-regions of the flow in addition to the whole domain. Figure A5 shows the centroids of the 2D RT corresponding to three different sub-regions along vertical direction. The results indicate that the three sub-regions are qualitatively similar but exhibit slight quantitative differences in their anisotropy. We defer further analysis to future work.

Centroids and anisotropy metrics of 2D Rayleigh-Taylor turbulence for three vertical regions

Figure A5. The centroids (a)-(c) and anisotropy metrics (d)-(f) of 2D RT associated with different vertical regions within z∈[0,Lz]z \in[0,L_z]. Panels (a,d),(b,e),(c,f) are associated with regions 0<z≤Lz/30 < z \leq L_z/3, Lz/3<z≤2Lz/3L_z/3 < z \leq2L_z/3, and 2Lz/3<z≤Lz2L_z/3 < z \leq L_z, respectively.

Isotropy of 3D-RT at the domain-size scales

As we discuss in the main text, at the largest possible scales, coarse-grained fields approach the spatial mean, which is trivially isotropic. Specifically, ϕ⃗ℓ→{ϕ}\vec{\phi}_{\ell} \to\{\phi\} as ℓ⃗→(∞,∞)\vec{\ell} \to(\infty,\infty). The spatial mean, {ϕ}\{\phi\}, is uniform and, thus, trivially isotropic. This can be easily seen from the illustrative example of an anisotropic Gaussian in Figure 3(a), which is isotropic at scales comparable to the domain size as shown in Figure 5. This is why the anisotropic metric ARzxshell≈0AR^{\mathrm{shell}}_{zx} \approx0 at k=1k = 1 in Figure 4(d).

For 3D-RT, we have a similar situation in Figure 9(d) where ARshell≈0AR^{\mathrm{shell}} \approx0 at the horizontal domain scale k=1k = 1 (i.e.i.e. ℓ=L\ell= L, where L=3.2L = 3.2 is the domain size along the xx-direction). This can be seen from the visualization of coarse KE at k=1k = 1 in Figure A6. Coarse KE is isotropic because the filter width is comparable to the horizontal domain size, such that the coarse-grained fields approach the spatial mean, which is trivially isotropic. Note that these large scales (k=1k = 1) where isotropy exists are larger than the size of biggest bubble or spike in 3D-RT seen in Figure 8 and Figure 10.

Insensitivity of our results to filtering kernels

To verify that our results are not sensitive to the particular filtering kernel adopted in the paper, which is a first-order Gaussian [1], we repeat our analysis with a Boxcar kernel, and with a third-order modified Gaussian filter. The expression of the third-order kernel GℓI(x)G_{\ell}^{I}(x) with filtering size ℓ\ell is

GℓI(x)=cGℓ(x,ℓ)−c′Gℓ′(x−x0)−c′Gℓ′(x+x0)(26)G_{\ell}^{I}(x) = cG_{\ell}(x,\ell) - c'G_{\ell'}(x-x_{0}) - c'G_{\ell'}(x+x_{0}) \tag*{(26)}

where xx is the spatial variable, Gℓ(x,ℓ)=6πℓ2e−6x2ℓ2G_{\ell}(x,\ell) = \sqrt{\frac{6}{\pi\ell^{2}}}e^{-\frac{6x^{2}}{\ell^{2}}} is the first-order Gaussian kernel as in eq. (17), c=1.1c=1.1, c′=0.05c'=0.05, ℓ′=0.5ℓ\ell'=0.5\ell, x0=4348ℓx_{0}=\sqrt{\frac{43}{48}}\ell are parameters to satisfy the third-order kernel constraints ∫GℓI dx=0\int G_{\ell}^{I}\,dx=0 and ∫x3GℓI dx=0\int x^{3}G_{\ell}^{I}\,dx=0. This higher-order kernel is, in some sense, more similar to a truncated Dirichlet kernel [111] than a Gaussian. The shapes of the first- and the third-order Gaussian, as well as the Boxcar, kernels are shown in Figure A7.

For convenience, we sample an xx-zz slice from the 3D Rayleigh-Taylor data and perform the anisotropy metrics using the above three kernels. Figure A8 and Figure A9(b) show the centroids, the anisotropic metrics, and the shell-averaged 1-D filtering spectra of KE, in which both the qualitative trends and the quantitative values are similar. Thus our shape anisotropy results for RT are not sensitive to the particular filtering kernel we have chosen.

Structure function scalings of 1D fields

Here, we present the scaling of structure functions associated with periodic 1D field data ϕ(x)\phi(x) whose Fourier spectra follows a power-law scaling Eϕ(k)∼k−αE_{\phi}(k) \sim k^{-\alpha}, for a few different α\alpha values. The field ϕ(x)\phi(x) is defined in a periodic domain x∈[0,2π)x \in[0, 2\pi), discretized on a uniform numerical grid of 32768 points. ϕ(x)\phi(x) and its associated Fourier spectra are shown in Figure A10. The corresponding second-order structure function S2(r)=⟨(u(x+r)−u(x))2⟩S_2(r) = \langle(u(x+r)-u(x))^2 \rangle, where the spatial average ⟨…⟩\langle\ldots\rangle is over x∈[0,2π)x \in[0, 2\pi), is shown in Figure A11 for 1D fields with Fourier spectra ranging from k−4k^{-4} to k0k^0. Figure A11(a) demonstrates that power-law scaling of a 2nd-order structure function, S2(r)∼rα−1S_2(r) \sim r^{\alpha-1}, is related to scaling of the Fourier spectrum E(k)∼k−αE(k) \sim k^{-\alpha} (legend in Figure A11), but only when 1<α<31 < \alpha< 3. We see from Figure A11(a) that the scaling of S2(r)S_2(r) is no longer related to that of the Fourier spectrum when E(k)E(k) is steeper than k−3k^{-3} (purple) or shallower than k−1k^{-1} (blue).

Figure A10. One-dimensional fields in a periodic domain [0,2π)[0,2\pi) whose Fourier spectra scale as E(k)∼k0E(k) \sim k^{0} (top row) and E(k)∼k−3E(k) \sim k^{-3} (bottom row).

Power-law scaling of a 2nd-order structure function and filtering spectra

Figure A11. Left panel: power-law scaling of a 2nd-order structure function, S2(r)∼rα−1S_2(r) \sim r^{\alpha-1}, is related to scaling of the Fourier spectrum E(k)∼k−αE(k) \sim k^{-\alpha} (legend), but only when 1<α<31 < \alpha< 3. We see that the scaling of S2(r)S_2(r) is no longer related to that of the Fourier spectrum when E(k)E(k) is steeper than k−3k^{-3} (purple) or shallower than k−1k^{-1} (blue). The fields analyzed here are 1D periodic data similar to those shown in Figure A10. Right panel: filtering spectra E‾(kℓ)\overline{E}(k_\ell) using a Gaussian kernel applied to the same data used in the left panel. We see that the filtering spectrum has the same scaling as the Fourier spectrum, E‾(k)∼E(k)∼k−α\overline{E}(k) \sim E(k) \sim k^{-\alpha}, for α<3\alpha< 3. Specifically, it can correctly capture power-law scaling that is shallower than k−1k^{-1} (blue) but fails for power-law scaling steeper than k−3k^{-3} (purple) since the Gaussian kernel we are using to calculate E‾(kℓ)\overline{E}(k_\ell) is a first-order kernel. It is possible for E‾(kℓ)\overline{E}(k_\ell) to correctly capture power-laws steeper than k−3k^{-3} by using a higher-order kernel [1].

Figure A11(b) shows filtering spectra E‾(kℓ)\overline{E}(k\ell) using a Gaussian (1st-order) filtering kernel applied to the same data used in panel A11(a). We see that the filtering spectrum has the same scaling as the Fourier spectrum, E‾(k)∼E(k)∼k−α\overline{E}(k) \sim E(k) \sim k^{-\alpha}, for α<3\alpha< 3. Specifically, it can correctly capture power-law scaling that is shallower than k−1k^{-1} (blue) but fails for power-law scaling steeper than k−3k^{-3} (purple) since the Gaussian kernel we are using to calculate E‾(kℓ)\overline{E}(k\ell) is a 1st-order kernel. It is possible for E‾(kℓ)\overline{E}(k\ell) to correctly capture power-laws steeper than k−3k^{-3} by using a higher-order kernel [1].

Fourier Spectrum of a Gaussian

For a 1-dimensional Gaussian function with zero mean and standard deviation σ\sigma,

ϕ(x)=12πσe−x22σ2,(27)\phi(x) = \frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{x^2}{2\sigma^2}}, \tag*{(27)}

defined over the periodic domain x∈[−π,π]x \in[-\pi,\pi], its Fourier transform when σ≪π\sigma\ll\pi is

ϕ^(k)=12π∫−ππe−ikx12πσe−x22σ2 dx=12πe−12k2σ2.(28)\hat{\phi}(k) = \frac{1}{2\pi}\int_{-\pi}^{\pi}e^{-ikx}\frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{x^2}{2\sigma^2}}\,dx = \frac{1}{2\pi}e^{-\frac{1}{2}k^2\sigma^2}. \tag*{(28)}

The factor 1/2π1/2\pi in the Fourier transform definition ensures that ϕ^(0)\hat{\phi}(0) yields the domain average. Eq. (A-6) is an elementary textbook result. Thus, the Fourier spectrum of a Gaussian is

E(k)=∣ϕ^∣2(k)+∣ϕ^∣2(−k)=12π2e−k2σ2.(29)E(k) = |\hat{\phi}|^2(k) + |\hat{\phi}|^2(-k) = \frac{1}{2\pi^2}e^{-k^2\sigma^2}. \tag*{(29)}

We drop the factor 1/21/2 in defining E(k)E(k) in the Appendix for convenience. It is straightforward to verify that Parseval’s theorem, ∫0∞E(k)dk=12π∫−ππϕ(x)2dx\int_0^\infty E(k)dk = \frac{1}{2\pi}\int_{-\pi}^{\pi}\phi(x)^2dx, is satisfied for σ≪π\sigma\ll\pi. Figure A12 compares the Fourier spectra of Gaussians with σ=π/8\sigma= \pi/8 and σ=π/32\sigma= \pi/32, showing that the narrower Gaussian (σ=π/32\sigma= \pi/32) extends to higher kk.

Figure A12. Fourier spectra of Gaussians with standard deviations σ=π/8\sigma= \pi/8 and σ=π/32\sigma= \pi/32

In our eq. (15) example of the 2D anisotropic Gaussian, we had

ϕ(x,z)=12πσx12πσze−x22σx2−z22σz2,where(σx,σz)=(π/32,π/8)andx,z∈[−π,π].(30)\begin{aligned} \phi(x,z) &= \frac{1}{\sqrt{2\pi}\sigma_x}\frac{1}{\sqrt{2\pi}\sigma_z}e^{-\frac{x^2}{2\sigma_x^2}-\frac{z^2}{2\sigma_z^2}},\\ \text{where}\qquad(\sigma_x,\sigma_z) &= (\pi/32,\pi/8)\qquad\text{and}\qquad x,z \in[-\pi,\pi]. \tag*{(30)} \end{aligned}

Its 2-dimensional Fourier transform follows from eq. (A-6),

ϕ^(kx,kz)=1(2π)2e−12kx2σx2e−12kz2σz2,(31)\widehat{\phi}(k_x,k_z)=\frac{1}{(2\pi)^2}e^{-\frac{1}{2}k_x^2\sigma_x^2}e^{-\frac{1}{2}k_z^2\sigma_z^2}, \tag*{(31)}

and its 2D Fourier spectrum is, therefore,

E2D(kx,kz)=4(2π)4e−kx2σx2e−kz2σz2.(32)E^{\mathrm{2D}}(k_x,k_z)=\frac{4}{(2\pi)^4}e^{-k_x^2\sigma_x^2}e^{-k_z^2\sigma_z^2}. \tag*{(32)}

Filtering Spectrum of a Gaussian

To obtain the analytical expression for the filtering spectrum of a 1-D Gaussian function ϕ(x)=12πσe−x22σ2\phi(x)=\frac{1}{\sqrt{2\pi\sigma}}e^{-\frac{x^2}{2\sigma^2}}, using the filtering kernel Gℓ(x)=1πℓe−x2ℓ2G_{\ell}(x)=\frac{1}{\sqrt{\pi\ell}}e^{-\frac{x^2}{\ell^2}}, we first evaluate the filtered quantity ϕ‾ℓ\overline{\phi}_{\ell}, then obtain {ϕ‾ℓ2}\{\overline{\phi}_{\ell}^{2}\}, before calculating the filtering spectrum E‾(k)=ddk{ϕ‾ℓ2}\overline{E}(k)=\frac{d}{dk}\{\overline{\phi}_{\ell}^{2}\}, with k=L/ℓk=L/\ell and L=2πL=2\pi. We now show the details.

The filtered Gaussian ϕ‾ℓ\overline{\phi}_{\ell} is

ϕ‾ℓ=1πℓ12πσ∫−∞∞dξ e−(x−ξ)2ℓ2e−12(ξσ)2.(33)\overline{\phi}_{\ell}=\frac{1}{\sqrt{\pi\ell}}\frac{1}{\sqrt{2\pi\sigma}}\int_{-\infty}^{\infty}\mathrm{d}\xi\ e^{-\frac{(x-\xi)^2}{\ell^2}}e^{-\frac{1}{2}\left(\frac{\xi}{\sigma}\right)^2}. \tag*{(33)}

Note that in a domain [−L/2,L/2)[-L/2,L/2) that is periodic, ϕ‾ℓ\overline{\phi}_{\ell} is calculated in Fourier space. This ensures that eq. (A-11) is evaluated correctly when ℓ→L\ell\to L, since the filtering kernel e−x2/ℓ2e^{-x^2/\ell^2} extends beyond [−L/2,L/2)[-L/2,L/2) for large ℓ\ell. Doing so is equivalent to extending the domain periodically in xx-space beyond its boundaries.

The cumulative spectrum is

{ϕ‾ℓ2}=1L∫−L/2L/2dx 1πℓ212πσ2∫−∞∞dξ1∫−∞∞dξ2 e−(x−ξ1)2ℓ2e−(x−ξ2)2ℓ2e−ξ12+ξ222σ2=1L2πℓ2+2σ2=1L2πL2k−2+2σ2.(34)\begin{aligned} \{\overline{\phi}_{\ell}^{2}\} &=\frac{1}{L}\int_{-L/2}^{L/2}\mathrm{d}x\ \frac{1}{\pi\ell^2}\frac{1}{2\pi\sigma^2}\int_{-\infty}^{\infty}\mathrm{d}\xi_1\int_{-\infty}^{\infty}\mathrm{d}\xi_2\ e^{-\frac{(x-\xi_1)^2}{\ell^2}}e^{-\frac{(x-\xi_2)^2}{\ell^2}}e^{-\frac{\xi_1^2+\xi_2^2}{2\sigma^2}}\\ &=\frac{1}{L\sqrt{2\pi}\sqrt{\ell^2+2\sigma^2}}=\frac{1}{L\sqrt{2\pi}\sqrt{L^2k^{-2}+2\sigma^2}}. \tag*{(34)} \end{aligned}

where the second line assumes σ≪L\sigma\ll L when evaluating integrals involving Gaussians. The assumption is only for the convenience of obtaining an analytical expression. Finally, the filtering spectrum is

E‾(k)=ddk{ϕ‾ℓ2}=12πk−3L2(k−2+2σ2/L2)3/2,(35)\overline{E}(k)=\frac{d}{dk}\{\overline{\phi}_{\ell}^{2}\}=\frac{1}{\sqrt{2\pi}}\frac{k^{-3}}{L^2\left(k^{-2}+2\sigma^2/L^2\right)^{3/2}}, \tag*{(35)}

which is plotted in Figure A13 for two different values of σ\sigma. In the limit k→∞k\to\infty, we see that E‾(k)∼k−3\overline{E}(k)\sim k^{-3}. As discussed in [1], this ‘locking’ at k−3k^{-3} scaling overestimates the energy at large kk and occurs when calculating the filtering spectrum using a 1st order kernel such as the Gaussian kernel used here. In comparison, the Fourier spectrum E(k)E(k) decays faster than k−3k^{-3} as shown in eq. (A-6) and Figure A12. For the filtering spectrum to be meaningful, one has to use a kernel with at least pp vanishing moments when analyzing a function whose Fourier spectrum decays as E(k)∼k−αE(k) \sim k^{-\alpha} with α<p+2\alpha< p+2. In practice, when Fourier transforms are difficult to perform, one calculates the filtering spectrum using a 1st-order kernel then decides if a higher order kernel is required based on whether or not E‾(k)\overline{E}(k) may be exhibiting a locked k−3k^{-3} scaling. See [1] for a detailed discussion.

Figure A13. Filtering spectra of Gaussians with standard deviations σ=π/8\sigma= \pi/8 and σ=π/32\sigma= \pi/32.

Note that even with a 1st-order kernel, anisotropy is still detected using filtering spectra. This can be seen from the filtering spectra of the two Gaussians in Figure A13, which shows differences in energy at different kk. However, the differences are less than those in Figure A12 for the Fourier spectra, i.e. the level of anisotropy can be underestimated when using a 1st-order kernel.

Taylor-Green using the dyadic bands

Figure A14 shows centroid and anisotropy metric for the Taylor-Green example using sharp-spectral filter in Fourier space as in Figure 7 but with dyadic bands k∈[k/2,2k]k \in[k/\sqrt{2},\sqrt{2}k].

Taylor-Green using Gaussian filter

Figure A15 shows filtering spectra for the Taylor-Green example using a Gaussian kernel. In defining filtering wavenumber, k=L/ℓk = L/\ell, we use the domain size L=2πL = 2\pi.

References

References

  1. [1]Mahmoud Sadek and Hussein Aluie. Extracting the spectrum of a flow by spatial filtering. Physical Review Fluids, 3(12):124610, 2018.arxiv.org/abs/1811.08259
  2. [2]Michael Chertkov. Phenomenology of rayleigh-taylor turbulence. Physical review letters, 91(11):115001, 2003.
  3. [3]Olivier Soulard. Implications of the monin-yaglom relation for rayleigh-taylor turbulence. Physical review letters, 109(25):254501, 2012.
  4. [4]Ye Zhou. Rayleigh–taylor and richtmyer–meshkov instability induced flow, turbulence, and mixing. ii. Physics Reports, 723:1–160, 2017.
  5. [5]Dongxiao Zhao, Riccardo Betti, and Hussein Aluie. Scale interactions and anisotropy in rayleigh–taylor turbulence. Journal of Fluid Mechanics, 930:A29, 2022.DOI
  6. [6]Michael D Graham. Fluid dynamics of dissolved polymer molecules in confined geometries. Annual Review of Fluid Mechanics, 43:273–298, 2011.DOI
  7. [7]Roberto Benzi and Emily SC Ching. Polymers in fluid flows. Annual Review of Condensed Matter Physics, 9:163–181, 2018.
  8. [8]L Biferale, Dmytro Khomenko, V L’vov, A Pomyalov, I Procaccia, and G Sahoo. Superfluid helium in three-dimensional counterflow differs strongly from classical flows: Anisotropy on small scales. Physical review letters, 122(14):144501, 2019.DOI
  9. [9]Satoshi Yui, Hiromichi Kobayashi, Makoto Tsubota, and Wei Guo. Fully coupled two-fluid dynamics in superfluid he 4: Anomalous anisotropic velocity fluctuations in counterflow. Physical review letters, 124(15):155301, 2020.DOI
  10. [10]R Manning. An anisotropic turbulence model for wave propagation near the surface of the earth. IEEE transactions on antennas and propagation, 34(2):258–261, 1986.
  11. [11]GM Grechko, AS Gurvich, V Kan, SV Kireev, and SA Savchenko. Anisotropy of spatial structures in the middle atmosphere. Advances in Space Research, 12(10):169–175, 1992.DOI
  12. [12]Italo Toselli, Brij Agrawal, and Sergio Restaino. Light propagation through anisotropic turbulence. JOSA A, 28(3):483–488, 2011.
  13. [13]Linyan Cui, Bindang Xue, and Fugen Zhou. Generalized anisotropic turbulence spectra and applications in the optical waves’ propagation through anisotropic turbulence. Optics express, 23(23):30088–30103, 2015.
  14. [14]Annick Pouquet, D Rosenberg, and Raffaele Marino. Linking dissipation, anisotropy, and intermittency in rotating stratified turbulence at the threshold of linear shear instabilities. Physics of Fluids, 31(10):105116, 2019.
  15. [15]John V Shebalin, William H Matthaeus, and David Montgomery. Anisotropy in mhd turbulence due to a mean magnetic field. Journal of Plasma Physics, 29(3):525–547, 1983.
  16. [16]David Montgomery and William H Matthaeus. Anisotropic modal energy transfer in interstellar turbulence. The Astrophysical Journal, 447:706, 1995.DOI
  17. [17]Jungyeon Cho and Ethan T Vishniac. The anisotropy of magnetohydrodynamic alfvénic turbulence. The Astrophysical Journal, 539(1):273, 2000.
  18. [18]Timothy S Horbury, Miriam Forman, and Sean Oughton. Anisotropic scaling of magnetohydrodynamic turbulence. Physical Review Letters, 101(17):175005, 2008.
  19. [19]EJ List. Turbulent jets and plumes. Annual review of fluid mechanics, 14(1):189–212, 1982.
  20. [20]RA Antonia, J Kim, and LWB Browne. Some characteristics of small-scale turbulence in a turbulent duct flow. Journal of Fluid Mechanics, 233:369–388, 1991.DOI
  21. [21]Amin Rasam, Geert Brethouwer, Philipp Schlatter, Qiang Li, and Arne V Johansson. Effects of modelling, resolution and anisotropy of subgrid-scales on large eddy simulations of channel flow. Journal of turbulence, (12):N10, 2011.
  22. [22]Claude Cambon, Nagi N Mansour, and Fabien S Godeferd. Energy transfer in rotating turbulence. Journal of Fluid Mechanics, 337:303–332, 1997.
  23. [23]Claude Cambon. Turbulence and vortex structures in rotating and stratified flows. European Journal of Mechanics-B/Fluids, 20(4):489–510, 2001.DOI
  24. [24]Fabien S Godeferd and Chantal Staquet. Statistical modelling and direct numerical simulations of decaying stably stratified turbulence. part 2. large-scale and small-scale anisotropy. Journal of Fluid Mechanics, 486:115–159, 2003.DOI
  25. [25]Claude Cambon and Robert Rubinstein. Anisotropic developments for homogeneous shear flows. Physics of Fluids, 18(8):085106, 2006.
  26. [26]Alan Burlot, B-J Gréa, Fabien S Godeferd, Claude Cambon, and Olivier Soulard. Large reynolds number self-similar states of unstably stratified homogeneous turbulence. Physics of Fluids, 27(6):065114, 2015.DOI
  27. [27]Andrey Nikolaevich Kolmogorov. The local structure of turbulence in incompressible viscous fluid for very large reynolds numbers. Cr Acad. Sci. URSS, 30:301–305, 1941.
  28. [28]Stephen B Pope. Turbulent flows. IOP Publishing, 2001.
  29. [29]P Goldreich and S Sridhar. Magnetohydrodynamic turbulence revisited. The Astrophysical Journal, 485(2):680, 1997.
  30. [30]Stanislav Boldyrev. On the spectrum of magnetohydrodynamic turbulence. The Astrophysical Journal Letters, 626(1):L37, 2005.
  31. [31]Xin Bian and Hussein Aluie. Decoupled cascades of kinetic and magnetic energy in magnetohydrodynamic turbulence. Physical review letters, 122(13):135101, 2019.DOI
  32. [32]Alexander A Schekochihin. Mhd turbulence: a biased review. Journal of Plasma Physics, 88(5):155880501, 2022.
  33. [33]John L Lumley and Gary R Newman. The return to isotropy of homogeneous turbulence. Journal of Fluid Mechanics, 82(1):161–178, 1977.
  34. [34]Daniel Livescu, JR Ristorcelli, RA Gore, SH Dean, WH Cabot, and AW Cook. High-reynolds number rayleigh–taylor turbulence. Journal of Turbulence, (10):N13, 2009.
  35. [35]F Thiesset, L Danaila, and RA Antonia. Dynamical effect of the total strain induced by the coherent motion on local isotropy in a wake. Journal of Fluid Mechanics, 720:393–423, 2013.DOI
  36. [36]Douglas W Carter and Filippo Coletti. Scale-to-scale anisotropy in homogeneous turbulence. Journal of Fluid Mechanics, 827:250–284, 2017.DOI
  37. [37]Amrapalli Garanaik and Subhas Karan Venayagamoorthy. Assessment of small-scale anisotropy in stably stratified turbulent flows using direct numerical simulations. Physics of Fluids, 30(12):126602, 2018.DOI
  38. [38]PK Yeung and James G Brasseur. The response of isotropic turbulence to isotropic and anisotropic forcing at the large scales. Physics of Fluids A: Fluid Dynamics, 3(5):884–897, 1991.DOI
  39. [39]Itai Arad, Victor S L’vov, and Itamar Procaccia. Correlation functions in isotropic and anisotropic turbulence: The role of the symmetry group. Physical Review E, 59(6):6753, 1999.
  40. [40]Luca Biferale and Itamar Procaccia. Anisotropy in turbulent flows and in turbulent transport. Physics Reports, 414(2-3):43–164, 2005.
  41. [41]Susan Kurien and Katepalli R Sreenivasan. Anisotropic scaling contributions to high-order structure functions in high-reynolds-number turbulence. Physical Review E, 62(2):2206, 2000.
  42. [42]Luca Biferale and Federico Toschi. Anisotropic homogeneous turbulence: hierarchy and intermittency of scaling exponents in the anisotropic sectors. Physical review letters, 86(21):4831, 2001.
  43. [43]Takashi Ishihara, Kyo Yoshida, and Yukio Kaneda. Anisotropic velocity correlation spectrum at small scales in a homogeneous turbulent shear flow. Physical review letters, 88(15):154501, 2002.
  44. [44]X Shen and Z Warhaft. Longitudinal and transverse structure functions in sheared and unsheared wind-tunnel turbulence. Physics of fluids, 14(1):370–381, 2002.DOI
  45. [45]CM Casciola, P Gualtieri, B Jacob, and R Piva. The residual anisotropy at small scales in high shear turbulence. Physics of Fluids, 19(10):101704, 2007.DOI
  46. [46]Alan Burlot, B-J Gréa, Fabien S Godeferd, Claude Cambon, and Olivier Soulard. Large reynolds number self-similar states of unstably stratified homogeneous turbulence. Physics of Fluids, 27(6):065114, 2015.DOI
  47. [47]GE Elsinga and Ivan Marusic. The anisotropic structure of turbulence and its energy spectrum. Physics of Fluids, 28(1):011701, 2016.
  48. [48]Olivier Soulard and Benoît-Joseph Gréa. Influence of zero-modes on the inertial-range anisotropy of rayleigh-taylor and unstably stratified homogeneous turbulence. Physical Review Fluids, 2(7):074603, 2017.DOI
  49. [49]Timothy S Horbury, Miriam Forman, and Sean Oughton. Anisotropic scaling of magnetohydrodynamic turbulence. Physical Review Letters, 101(17):175005, 2008.
  50. [50]RT Wicks, TS Horbury, CHK Chen, and AA Schekochihin. Anisotropy of imbalanced alfvénic turbulence in fast solar wind. Physical review letters, 106(4):045001, 2011.
  51. [51]David C Champeney and DC Champeney. A handbook of Fourier theorems. Cambridge University Press, 1987.
  52. [52]Steven G Krantz. A panorama of harmonic analysis, volume 27. American Mathematical Society, 2019.
  53. [53]Alan V. Oppenheim and Ronald W. Schafer. Discrete-Time Signal Processing. Prentice Hall, Upper Saddle River, New Jersey, 1999.
  54. [54]R E Thomson and W J Emery. Data Analysis Methods in Physical Oceanography. Elsevier Science, 3rd edition, 2001.
  55. [55]Olivier Soulard and Jérôme Griffond. Inertial-range anisotropy in rayleigh-taylor turbulence. Physics of Fluids, 24(2):025101, 2012.DOI
  56. [56]Anna C. Savage, Brian K. Arbic, James G. Richman, Jay F. Shriver, Matthew H. Alford, Maarten C Buijsman, J. Thomas Farrar, Hari Sharma, Gunnar Voet, Alan J Wallcraft, and Luis Zamudio. Frequency content of sea surface height variability from internal gravity waves to mesoscale eddies. Journal of Geophysical Research: Oceans, 122(3):2519–2538, 3 2017.DOI
  57. [57]Amanda K. O’Rourke, Brian K. Arbic, and Stephen M. Griffies. Frequency-domain analysis of atmospherically forced versus intrinsic ocean surface kinetic energy variability in GFDL’s CM2-O model hierarchy. Journal of Climate, 31(5):1789–1810, 2018.DOI
  58. [58]Robert B. Scott and Faming Wang. Direct evidence of an oceanic inverse kinetic energy cascade from satellite altimetry. Journal of Physical Oceanography, 35(9):1650–1666, 2005.DOI
  59. [59]Hemant Khatri, Jai Sukhatme, Abhishek Kumar, and Mahendra K Verma. Surface ocean enstrophy, kinetic energy fluxes, and spectra from satellite altimetry. Journal of Geophysical Research: Oceans, 123(5):3875–3892, 2018.DOI
  60. [60]Raffaele Ferrari and Carl Wunsch. Ocean Circulation Kinetic Energy: Reservoirs, Sources, and Sinks. Annual Review of Fluid Mechanics, 41(1):253–282, January 2009.DOI
  61. [61]Hector S Torres, Patrice Klein, Dimitris Menemenlis, Bo Qiu, Zhan Su, Jinbo Wang, Shuiming Chen, and Lee-Lueng Fu. Partitioning Ocean Motions Into Balanced Motions and Internal Gravity Waves: A Modeling Study in Anticipation of Future Space Missions. Journal of Geophysical Research-Oceans, 123(11):8084–8105, November 2018.DOI
  62. [62]Patrice Klein, Guillaume Lapeyre, Lia Siegelman, Bo Qiu, Lee-Lueng Fu, Hector Torres, Zhan Su, Dimitris Menemenlis, and Sylvie Le Gentil. Ocean-Scale Interactions From Space. Earth and Space Science, 6(5):795–817, May 2019.
  63. [63]Benjamin A Storer, Michele Buzzicotti, Hemant Khatri, Stephen M Griffies, and Hussein Aluie. Global energy spectrum of the general oceanic circulation. Nature communications, 13(1):1–9, 2022.DOI
  64. [64]M Buzzicotti, BA Storer, H Khatri, SM Griffies, and H Aluie. Spatio-temporal coarse-graining decomposition of the global ocean geostrophic kinetic energy. Journal of Advances in Modeling Earth Systems, 15(6):e2023MS003693, 2023.DOI
  65. [65]Robert H Kraichnan. Direct-interaction approximation for shear and thermally driven turbulence. The Physics of Fluids, 7(7):1048–1062, 1964.DOI
  66. [66]B-J Gréa. The rapid acceleration model and the growth rate of a turbulent mixing zone induced by rayleigh-taylor instability. Physics of fluids, 25(1):015118, 2013.DOI
  67. [67]Uriel Frisch. Turbulence: the legacy of AN Kolmogorov. Cambridge university press, 1995.
  68. [68]Armando Babiano, Claude Basdevant, and Robert Sadourny. Structure functions and dispersion laws in two-dimensional turbulence. Journal of Atmospheric Sciences, 42(9):941–949, 1985.
  69. [69]Dhruv Balwada, Jin-Han Xie, Raffaele Marino, and Fabio Feraco. Direct observational evidence of an oceanic dual kinetic energy cascade and its seasonality. Science Advances, 8(41):eabq2566, 2022.
  70. [70]Gregory L Eyink. Locality of turbulent cascades. Physica D: Nonlinear Phenomena, 207(1):91–116, 2005.
  71. [71]Gregory L Eyink and Hussein Aluie. Localness of energy cascade in hydrodynamic turbulence. i. smooth coarse graining. Physics of Fluids, 21(11):115107, 2009.DOI
  72. [72]Hussein Aluie. Coarse-grained incompressible magnetohydrodynamics: analyzing the turbulent cascades. New Journal of Physics, January 2017.DOI
  73. [73]Hussein Aluie and Gregory L Eyink. Scale locality of magnetohydrodynamic turbulence. Physical review letters, 104(8):081101, 2010.
  74. [74]L Biferale, M Cencini, A Lanotte, D Vergni, and A Vulpiani. Inverse statistics of smooth signals: The case of two dimensional turbulence. Physical Review Letters, 87(12):124501, 2001.
  75. [75]G. L. Eyink. Besov spaces and the multifractal hypothesis. J. Stat. Phys., 78:353–375, 1995.
  76. [76]Daniel Livescu and JR Ristorcelli. Variable-density mixing in buoyancy-driven turbulence. Journal of Fluid Mechanics, 605:145–180, 2008.
  77. [77]Daniel Livescu. Turbulence with large thermal and compositional density variations. Annual Review of Fluid Mechanics, 52:309–341, 2020.DOI
  78. [78]Shigeo Kida and Steven A Orszag. Energy and spectral dynamics in forced compressible turbulence. Journal of Scientific Computing, 5(2):85–125, 1990.DOI
  79. [79]Andrew W Cook and Ye Zhou. Energy transfer in Rayleigh-Taylor instability. Physical Review E, 66(2):192, August 2002.
  80. [80]Dongxiao Zhao and Hussein Aluie. Inviscid criterion for decomposing scales. Physical Review Fluids, 3(5):054603, 2018.arxiv.org/abs/1804.07715
  81. [81]Ingrid Daubechies. Ten lectures on wavelets. SIAM, 1992.
  82. [82]Valérie Perrier, Thierry Philipovitch, and Claude Basdevant. Wavelet spectra compared to fourier spectra. Journal of mathematical physics, 36(3):1506–1519, 1995.
  83. [83]Christopher D Sogge. Fourier integrals in classical analysis, volume 210. Cambridge University Press, 2017.
  84. [84]Shikhar Rai, Matthew Hecht, Matthew Maltrud, and Hussein Aluie. Scale of oceanic eddy killing by wind from global satellite observations. Science Advances, 7(28):eabf4920, 2021.DOI
  85. [85]Michele Buzzicotti, Benjamin A Storer, Stephen M Griffies, and Hussein Aluie. A coarse-grained decomposition of surface geostrophic kinetic energy in the global ocean. Earth and Space Science Open Archive, page 58, 2021.
  86. [86]Hussein Aluie. Convolutions on the sphere: commutation with differential operators. GEM-International Journal on Geomathematics, 10(1):9, 2019.
  87. [87]A Leonard. Energy cascade in large-eddy simulations of turbulent fluid flows. Advances in geophysics, 18:237–248, 1975.DOI
  88. [88]M Germano. Turbulence: the filtering approach. Journal of Fluid Mechanics, 238:325–336, 1992.
  89. [89]C. Meneveau and J. Katz. Scale-Invariance and Turbulence Models for Large-Eddy Simulation. Ann. Rev. Fluid Mech., 32:1–32, 2000.DOI
  90. [90]S. G. Krantz. A Panorama of Harmonic Analysis. The Mathematical Association of America, USA, 1999.
  91. [91]Hussein Aluie and Gregory L Eyink. Localness of energy cascade in hydrodynamic turbulence. ii. sharp spectral filter. Physics of Fluids, 21(11):115108, 2009.
  92. [92]Charles Meneveau. Analysis of turbulence in the orthonormal wavelet representation. Journal of Fluid Mechanics, 232:469–520, 1991.DOI
  93. [93]Charles Meneveau. Dual spectra and mixed energy cascade of turbulence in the wavelet representation. Physical review letters, 66(11):1450, 1991.
  94. [94]Aditya Suresh Ghate and SK Lele. Gabor mode enrichment in large eddy simulations of turbulent flow. Journal of Fluid Mechanics, 903:A13, 2020.
  95. [95]Perry L Johnson. A physics-inspired alternative to spatial filtering for large-eddy simulations of turbulent flows. Journal of Fluid Mechanics, 934:A30, 2022.DOI
  96. [96]Ugo Piomelli, William H Cabot, Parviz Moin, and Sangsan Lee. Subgrid-scale backscatter in turbulent and transitional flows. Physics of Fluids A: Fluid Dynamics, 3(7):1766–1771, 1991.
  97. [97]XIA Yang, J Sadique, R Mittal, and Charles Meneveau. Integral wall model for large eddy simulations of wall-bounded turbulent flows. Physics of Fluids, 27(2):025112, 2015.
  98. [98]Pierre Sagaut. Large eddy simulation for incompressible flows: an introduction. Springer Science & Business Media, 2006.
  99. [99]Hussein Aluie, Matthew Hecht, and Geoffrey K. Vallis. Mapping the Energy Cascade in the North Atlantic Ocean: The Coarse-graining Approach. Journal of Physical Oceanography, 48:225–244, 2018.DOI
  100. [100]Shikhar Rai, Matthew Hecht, Matthew Maltrud, and Hussein Aluie. Scale of oceanic eddy killing by wind from global satellite observations. Science Advances, 7(28):eabf4920, 2021.DOI
  101. [101]M K Rivera, H Aluie, and R E Ecke. The direct enstrophy cascade of two-dimensional soap film flows. Physics of Fluids, 26(5), May 2014.DOI
  102. [102]Wolfgang Hillebrandt and Jens C Niemeyer. Type ia supernova explosion models. Annual Review of Astronomy and Astrophysics, 38(1):191–230, 2000.
  103. [103]R Betti and OA Hurricane. Inertial-confinement fusion with lasers. Nature Physics, 12(5):435, 2016.
  104. [104]Brian M. Haines, R. C. Shah, J. M. Smidt, B. J. Albright, T. Cardenas, M. R. Douglas, C. Forrest, V. Yu Glebov, M. A. Gunderson, C. E. Hamilton, K. C. Henderson, Y. Kim, M. N. Lee, T. J. Murphy, J. A. Oertel, R. E. Olson, B. M. Patterson, R. B. Randolph, and D. W. Schmidt. Observation of persistent species temperature separation in inertial confinement fusion mixtures. Nature Communications, 11(1):544, Jan 2020.
  105. [105]M. Gatu Johnson, B.M. Haines, P.J. Adrian, C. Forrest, J.A. Frenje, V.Yu. Glebov, W. Gribble, R. Janezic, J.P. Knauer, B. Lahmann, F.J. Marshall, T. Michel, F.H. Séguin, C. Stoeckl, and R.D. Petrasso. 3d xrage simulation of inertial confinement fusion implosion with imposed mode 2 laser drive asymmetry. High Energy Density Physics, 36:100825, 2020.
  106. [106]D S Clark, C R Weber, J L Milovich, J D Salm̄onson, A L Kritcher, S W Haan, B A Hammel, D E Hinkel, O A Hurricane, O S Jones, M M Marinak, P K Patel, H F Robey, S M Sepke, and M J Edwards. Three-dimensional simulations of low foot and high foot implosion experiments on the national ignition facility. Physics of Plasmas, 23(5):056302, May 2016.
  107. [107]D. J. Schlossberg, G. P. Grim, D. T. Casey, A. S. Moore, R. Nora, B. Bachmann, L. R. Benedetti, R. M. Bionta, M. J. Eckart, J. E. Field, et al. Observation of hydrodynamic flows in imploding fusion plasmas on the national ignition facility. Phys. Rev. Lett., 127:125001, Sep 2021.
  108. [108]A. L. Kritcher, C. V. Young, H. F. Robey, C. R. Weber, A. B. Zylstra, O. A. Hurricane, D. A. Callahan, J. E. Ralph, J. S. Ross, K. L. Baker, et al. Design of inertial fusion implosions reaching the burning plasma regime. Nature Physics, 18(3):251–258, Mar 2022.
  109. [109]A. L. Kritcher, A. B. Zylstra, D. A. Callahan, O. A. Hurricane, C. R. Weber, D. S. Clark, C. V. Young, J. E. Ralph, D. T. Casey, A. Pak, et al. Design of an inertial fusion experiment exceeding the lawson criterion for ignition. Phys. Rev. E, 106:025201, Aug 2022.
  110. [110]Ye Zhou. Rayleigh–taylor and richtmyer–meshkov instability induced flow, turbulence, and mixing. i. Physics Reports, 720:1–136, 2017.
  111. [111]Evangelos T Spyropoulos and Gregory A Blaisdell. Evaluation of the dynamic model for simulations of compressible decaying isotropic turbulence. AIAA Journal, 34:51–60.
  112. [112]Daniel Livescu and JR Ristorcelli. Variable-density mixing in buoyancy-driven turbulence. Journal of Fluid Mechanics, 605:145–180, 2008.
  113. [113]W Cabot. Comparison of two-and three-dimensional simulations of miscible rayleigh-taylor instability. Physics of Fluids, 18(4):045101, 2006.
  114. [114]Huasen Zhang, Riccardo Betti, R Yan, D Zhao, D Shvarts, and H Aluie. Self-similar multimode bubble-front evolution of the ablative rayleigh-taylor instability in two and three dimensions. Physical review letters, 121(18):185002, 2018.
  115. [115]Xin Bian, Hussein Aluie, Dongxiao Zhao, Huasen Zhang, and Daniel Livescu. Revisiting the late-time growth of single-mode rayleigh–taylor instability and the role of vorticity. Physica D: Nonlinear Phenomena, page 132250, 2019.
  116. [116]Benjamin A. Storer and Hussein Aluie. FlowSieve: A Coarse-Graining Utility for Geophysical Flows on the Sphere. Journal of Open Source Software, 8(84):4277, 4 2023.

Paper details

vlnc:W_rdgzcn45
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents