WorksPhysics and Astronomy
Orientation and slip statistics of spheroids settling in turbulence
Abstract
Small spheroids settling through turbulence are turned broadside-on by fluid inertia and tilted away by velocity gradients. We describe their orientation and slip statistics through the crossover from random orientation to alignment, in the overdamped limit. With quasi-static Stokes drag the leading inertial torque is an exact gradient flow on the orientation sphere, its potential quadratic in the cosine of the tilt and only rescaled by mobility anisotropy. The orientation performs an overdamped random walk in a quadratic well, and isotropic Gaussian white noise gives exactly a Boltzmann distribution. For forcing of finite correlation time, a closure keeps this form with its width set by linear response; it recovers the uniform density as settling vanishes and, for fixed-orientation forcing statistics, the aligned coefficient of rapid-settling theory, and errs most in the crossover. The root-mean-square lateral slip peaks where the alignment rate approaches the Kolmogorov rate, as faster settling strengthens the sideways drift at a given tilt but suppresses the tilt; the nearly shape-independent ratio, which tensor-forced simulations reproduce, lets a measured peak constrain the torque coefficient or dissipation rate for given forcing statistics. Dissipation intermittency enters as a lognormal superposition of Boltzmann distributions, which fattens the tails but leaves the leading aligned mean-square tilt unchanged. The superposition reproduces the shape of plate orientation distributions from direct numerical simulations, where a single Boltzmann distribution fails; published moments confirm the predicted spread beyond the weakest torques, and at strong alignment the measured fourth-order tilt moment lies between the two predictions.
Introduction
The orientation of small nonspherical particles settling through a turbulent fluid controls their optical signatures, their settling speeds and their collision rates. Ice crystals falling broadside-on produce the specular glints, polarimetric signals and enhanced backscatter used in lidar and radar retrievals [39, 5, 51], which return orientation averages, and observations place real cirrus between the two limits usually assumed, random orientation and perfect alignment [5, 44, 34]. The width of the orientation density matters as much as its mean: for plates above 200 µm, tilt widths across the plausible range – change computed lidar backscatter by more than two orders of magnitude at fixed size [36]. The same statistics enter collision models for ice crystals [32, 19, 45, 46] and set the orientation of frazil discs in supercooled water [9].
The torque due to fluid inertia is well understood in the near-viscous regime. At small but finite particle Reynolds number a fluid-inertial torque turns a settling spheroid broadside-on to its slip. It was obtained for slender bodies by Khayat & Cox [22] and for arbitrary aspect ratio by Dabade et al. [8], and an inertial slender-body theory extends it to diameter-based Reynolds numbers of Joshi et al. [18]. Experiments confirm the alignment in quiescent fluid [27, 42] and show that in air, where the particle is far denser than the fluid, the approach to it oscillates and the orientation fluctuates strongly [3].
The fluctuations about the broadside orientation are less well understood. Klett [24], following Cho et al. [6], balanced a restoring torque against an assumed white-noise turbulent torque; his density has the form derived below, but with coefficients assumed rather than obtained from the equations of motion. Rapid-settling theories obtain those coefficients for arbitrary aspect ratio [29, 11, 12, 1, 43], Gustavsson et al. [13] add particle inertia, Sheikh et al. [47] estimate when the fluid-inertial torque outweighs the turbulent one, and Roy et al. [43] map a shear-rate distribution onto an orientation density in that limit; direct numerical simulations span the transition from disorder to alignment [48, 19, 1, 40]. The derived results are, however, small-tilt expansions or rapid-settling asymptotics, whereas for crystals of tens of microns in cloud turbulence the alignment parameter is of order unity [13, 40]: the crossover, where those expansions fail and the available results come from simulation and from experiments on finite-size particles [43].
We treat that crossover in the overdamped limit, in which the response times of the particle are short and its inertia does not enter (§ 9 sets out its range of validity). The inertial torque is an exact gradient flow on the sphere of orientations, with a potential quadratic in the cosine of the tilt, and the Jeffery forcing at a fixed orientation has no preferred direction in the tangent plane. For white-noise forcing the stationary density is then a Boltzmann distribution, as for orientable Brownian particles [4], here exponential in the squared cosine of the tilt. Turbulent forcing has a finite correlation time, which leaves the orientation alone non-Markovian. We keep the Boltzmann form, fix its concentration from the exact linear response, and test that closure against stochastic simulations; it returns the uniform density as settling vanishes and, with fixed-orientation forcing statistics, the rapid-settling coefficient when the torque is strong (§ 4.3).
The even moments of that density give in closed form the orientation averages a scattering calculation needs and the slip statistics. The lateral slip first grows and then decays as the particle settles faster, peaking where the alignment rate is close to the Kolmogorov rate, at a ratio of the two that depends only weakly on shape, so the peak constrains the torque coefficient or, with the torque known, the dissipation rate, given the forcing statistics. The structure carries over to orientation-dependent torques and to dissipation intermittency, which enters as a lognormal superposition of Boltzmann densities and fits orientation densities of settling plates from direct numerical simulation much better than a single one. The published moments of those simulations follow the predicted spread except at the weakest torques, while at strong torque their fourth-order tilt moment lies between the predictions of the single density and the superposition. Figure 1 sets out the problem and the notation.
Figure 1. Schematic of the problem. (a) A spheroid settling with slip through a locally linear flow (dashed lines, the vertical and the horizontal): the anisotropic mobility turns away from , Eq. (10), and the fluid-inertial torque (red arcs) reduces the tilt . (b) Attracting sets: the girdle for columns () and the poles for plates (). (c) The orientational potential (4). (d) The density of the column at four settling numbers at .
Sections 2–4 set out the orientation dynamics, the turbulent forcing and the stationary statistics, Secs. 5 and 6 the slip and its maximum, Sec. 7 the extensions, Sec. 8 the tests against simulated turbulence, and Sec. 9 the application to atmospheric ice and frazil. Appendix B lists the checks behind the closed-form results.
Orientation dynamics
Equations of motion
Consider a rigid spheroid with symmetry semi-axis , transverse semi-axis , aspect ratio and unit symmetry axis , settling through homogeneous isotropic turbulence. A single covers elongated and flattened shapes, which enter the orientation dynamics through the Bretherton constant
and, through the mobility, the settling factors of Sec. 2.4. The sign of separates prolate columns () from oblate plates () and, as Sec. 2.3 shows, decides whether the particle is attracted to the girdle or to the poles. The particle is smaller than the Kolmogorov length , so it samples a locally linear flow; its translational and rotational response times are short, so the slip velocity relative to the local fluid is set quasistatically by gravity and Stokes drag and its rotation is overdamped; and its slip Reynolds number is small. We keep Jeffery’s rotation and the leading inertial torque due to the slip, and neglect the inertial corrections due to the ambient velocity gradient. The orientation then evolves as
the first two terms being Jeffery’s rotation [16] in the spin and strain of the local flow, and the last the fluid-inertial contribution. At the inertial torque on a spheroid in a uniform slip has the exact vector structure [22, 8, 17], so that, after division by the transverse rotational resistance,
where for is the bare dimensionless torque coefficient (Appendix A) and carries the prolate–oblate distinction: the torque turns any spheroid broadside-on to its slip, with the symmetry axis across the slip for a column and along it for a plate. Below, without a subscript denotes the effective coefficient defined by Eq. (4); § 2.4 relates it to .
Gradient-flow structure
For slip locked to gravity, , Eq. (3) reads with , which drives to 0 for a column and to for a plate. Since for any , the drift is exactly the negative surface gradient of the potential
where is the Kolmogorov time and the settling number based on the Kolmogorov velocity ; we call the alignment parameter. For slip locked to gravity ; § 2.4 shows that an orientation-dependent slip changes only this coefficient. More generally, any axisymmetric drift is a gradient flow on the sphere with , a fact used in § 7.
The strain term has the same structure. For any symmetric ,
where is the gradient on the unit sphere, the rotary-potential construction of Brenner & Condiff [4]. With and , Eq. (2) becomes
a rigid rotation at the ambient angular velocity together with gradient ascent on a single quadratic form. The rotation, with the vorticity, is not a gradient. A rotation history prescribed independently of turns the sphere rigidly and leaves a uniform density uniform; once the slip, and with it the path, depends on , so does the vorticity history, and the random-orientation limit rests instead on the joint statistical isotropy of and the gradients (§ 3). For a steady flow, by contrast, the rotary flux vanishes only if the flow is irrotational (Brenner & Condiff [4]); here the vorticity belongs to the stochastic forcing, and the potential comes from gravity.
The axisymmetric potential has two further consequences. Its drift has no azimuthal component, so the azimuth about gravity is exactly neutral, and its relaxation is solvable at any tilt. In the vertical plane that contains , which the drift leaves fixed, let be the signed angle of from the attracting direction, the horizontal for a column and the vertical for a plate, positive on a chosen side; its magnitude is the tilt, arcsin for a column and arccos for a plate. Then , and decays as . This is the overdamped limit of a damped oscillator. With rotational inertia, and the slip still quasi-static, the same torque gives , with the rotational response time, so that small tilts overshoot the broadside orientation when . Heavy spheroids settling in still air oscillate in this way (Bhowmick et al. [3]), though their full dynamics also couples the tilt to a lagging slip; for , the condition of § 9, the inertial term drops out and the gradient flow remains. Since the restoring drift is , a forcing that carries the same factor keeps the problem linear in , as for a thin disc in a vertical simple shear of rate , for which (Roy et al. [43]). In ,
so that for constant
which approaches the equilibrium
with no small-angle assumption. For a time-dependent the tangent is an exact linear filter of the shear rate, Gaussian when is Gaussian, a property special to this planar problem. Roy et al. [43] map a shear-rate distribution through Eq. (9) to obtain an orientation density (§ 4.3); the full competition between the shear and settling torques is mapped in Mishra & Roy [31].
Attracting sets
Columns and plates both settle broadside-on, but their attracting sets differ (Figure 1b). A column lies with its long axis horizontal: a tilt out of the horizontal is restored, while a swing within the horizontal plane meets no resistance, every horizontal direction being equally broadside-on. A plate settles face-down, and a tilt in any direction meets a restoring torque. In terms of the potential, the minimum of for is the equatorial girdle , with one restored and one neutral tangent direction; for it is the pole pair ( the lower pole, with pointing downwards), a single physical orientation by fore–aft symmetry, with both tangent directions restored at the same rate . With the two components equally forced (§ 3), the leading mean-square tilt of a plate is therefore twice that of a column at equal concentration. The same factor of two appears in the persistent limit of Anand et al. [1] and, as a shape coefficient of one for columns and two for plates, through the inertial regimes of Gustavsson et al. [13]. The random-orientation limits are not in the ratio two, because the spherical measure differs near a pole and near a girdle: they are for a column and for a plate [Eq. (26)], in the ratio 2.442.
Effect of mobility anisotropy
At small translational Stokes number the quasi-static balance gives , where is the Stokes settling speed of the volume-equivalent sphere, , and is the dimensionless mobility, with and the axial and transverse settling factors (Appendix A). Thus
so both the direction and the magnitude of the slip depend on orientation, with running between and . Since the torque (3) is quadratic in , one might expect quartic and higher powers of in the potential, and hence in the exponent of the density. Substituting Eq. (10) and using rules this out:
exactly linear in . For the leading slip-induced torque with quasi-static Stokes drag, the potential therefore stays exactly quadratic, and the mobility only rescales the relaxation rate,
where the torque (3) is evaluated with the instantaneous slip (10) and Sv is the settling number based on the broadside terminal speed ( prolate, oblate). Writing for the mobility ratio that multiplies in the broadside form,
The effective of Eq. (4) is or according to which speed defines Sv. Orientation statistics depend on alone and are the same in either convention; the slip statistics of § 5 are written with , and the convention matters whenever is compared with a quiescent-flow coefficient or read off an experiment. Departures from the quadratic potential need physics beyond the quasi-static Stokes mobility, such as nonlinear drag at finite or unsteady slip; preferential sampling at finite Stokes number changes the forcing statistics rather than the potential.
Turbulent forcing
The Jeffery terms in Eq. (2) are linear in the velocity gradient and contribute, at orientation , the tangential angular velocity
tangential because . Treating as a prescribed stochastic forcing means evaluating its statistics at an externally fixed orientation; applying them along a trajectory assumes in addition that is independent of the instantaneous velocity gradient. We derive the covariance first and then examine that assumption.
The one-time statistics follow from the isotropic covariance of the velocity gradient in incompressible turbulence,
whose symmetric and antisymmetric parts give
each with trace ; the cross-covariance vanishes, no isotropic rank-four tensor being antisymmetric in one index pair and symmetric in the other. At fixed the gradient statistics are invariant under rotations about , and the only tangent-plane tensor with that invariance is , so no moment of can single out a direction in the tangent plane and the contraction fixes only the coefficient:
with total variance , five parts from the vorticity and from the strain. The independence assumption fails in general, because Jeffery rotation correlates with the flow: rods align with the vorticity more than with the principal strain axis [41], both aligning with the Lagrangian stretching direction (Ni et al. 2014). Either alignment weakens the forcing, since vanishes for along and the projected strain for along a strain eigenvector, so Eq. (18) overstates the forcing felt along a path, by more for slender columns. At the skew-torque end, with Sv small, gravity drops out and the joint statistics of and the gradients are isotropic, so the stationary density is uniform whatever the alignment of with the local gradient. At the strong-torque end the particle falls a distance per relaxation time, less than the gradient correlation length – (Appendix C), so a realignment sees a single frozen gradient, at an orientation imposed by gravity rather than by the flow. That suffices for a plate, restored in both tangent directions, but not for a column, whose azimuth in the girdle is set by the Jeffery dynamics at any settling number. The assumption is therefore controlled at both ends for plates, at the random end only for columns, and nowhere in the crossover, where its error adds to that of the closure (§ 8.1).
We model , whose time dependence along a path is not universal, as an Ornstein–Uhlenbeck process with correlation time and take , a single time although vorticity decorrelates more slowly than strain [52]. The model concerns at a fixed orientation. When and are themselves evolved as Ornstein–Uhlenbeck tensors sharing that time, as in the tensor-forced simulations (Appendix F), at fixed is a linear combination of their components and so exactly Ornstein–Uhlenbeck; along a path Eq. (14) is evaluated at the moving orientation and is no longer strictly so, a difference the closure neglects and the simulations retain (§ 4.2 and § 8.1). Settling faster than the eddies decorrelate shortens the correlation to the crossing time [12, 13], which falls below once exceeds –, according to shape and Reynolds number (Appendix C), well above the slip maximum studied here. The value of cancels from the aligned limit (§ 4.3) and moves the peak constant of § 6 for the column from at to at and at .
Stationary orientation statistics
Coloured-noise linear response
Write for the small tangential displacement of from the attracting set. Linearising the drift, each restored component obeys , with Ornstein–Uhlenbeck forcing of per-component variance from Eq. (18) and correlation time . Since , the tilt equals the displacement at leading order: for a girdle, with one restored direction, and for a bipolar state, with two. The stationary variance of this linear problem is exact,
which with and Eq. (4) gives, per restored component,
Equation (20) recovers the strong-alignment scalings of the rapid-settling theories [29, 43, 12] and identifies the coloured-noise correction : the effective rotational diffusivity is
reduced below the white-noise value by the factor , substantially once the relaxation is faster than the decorrelation of the forcing.
The Boltzmann density
For Gaussian white-noise forcing with isotropic statistics the full nonlinear problem is isotropic rotational diffusion on the sphere, with diffusivity , in the gradient drift of § 2.2. That drift gives , so a density uniform in azimuth obeys
and its stationary solution, which carries no probability current, is of Boltzmann form, , exactly and at every ; no small-tilt assumption enters, only the gradient structure. With the quadratic potential it is the Watson distribution of directional statistics [50, 28], which we call the Boltzmann density,
with concentration . Here and below is the density of on , the density per unit solid angle being . Since depends on alone we write it in , and statistics involving the folded variable, such as or the tilt of a plate, are written in . The normalisations are
with the Dawson function. The orientation performs an overdamped random walk in a quadratic well, with and in the roles of the spring constant and the temperature, each divided by the friction; confinement to the sphere turns the Gaussian of a harmonic well into the density (22) and makes it saturate at the uniform measure as . Unlike thermal noise, the turbulent forcing obeys no fluctuation–dissipation relation: is set by the velocity gradients and, within the prescribed forcing model, is independent of the particle’s size and rotational friction, while grows with the settling speed, so the particle size alone changes by orders of magnitude.
For forcing of finite correlation time we replace by , so that
or for general . This replacement is a finite-memory closure, tested below, and the only approximation beyond the forcing assumptions of § 3: the Boltzmann form, exact for isotropic Gaussian white noise, is kept and its one parameter is matched to the exact linear response. A systematic coloured-noise elimination of the Jung–Hänggi type [20, 14], exact for a linear drift, would not give a density of this form here, since the restoring drift is cubic in the tangent coordinate; it would produce an orientation-dependent prefactor and higher harmonics in the exponent. Equation (24) nevertheless has the correct limits. As settling vanishes at fixed shape, statistical isotropy makes the density uniform whatever the forcing (Sec. 3, and Eq. (24) returns it; switching off the torque while the particle keeps settling is a different limit, and the inertial runs of Anand et al. [1] with the torque suppressed are not uniform (insets of their Fig. 1). A dominant torque keeps the tilt small, where the linear response (19) that supplies is exact for forcing with the fixed-orientation statistics, so the closure recovers the aligned second-moment coefficient (§ 4.3); that assumption is controlled there for plates but not for columns (§ 3).
In between, we measure the accuracy of the closure by integrating the equation it describes,
with the upper sign for plates and the lower for columns, for two design shapes, a prolate spheroid of aspect ratio and an oblate one of , at the reference value with . Their physical coefficients are and , but orientation moments depend on alone, so the curves apply to the physical particles after the rescaling , which the root-mean-square (rms) slip, carrying an explicit , needs in its ordinate as well. An ambient vector , independent of the particle, is advanced as a three-component Ornstein–Uhlenbeck process with correlation time and per-component variance , and is taken at the current orientation. At a fixed orientation this reproduces Eq. (18), though not along a trajectory (§ 8.1). Unlike the Jeffery forcing ((14)), which is odd under as a director requires, is even in each realisation; its ensemble statistics are even, and the tensor-forced runs below give the same moments to within 1–2.5%. The velocity gradient itself is not simulated, so the test probes the closure within the forcing model rather than the model itself.
Figure 2(a) shows for the column and for the plate. The closure tracks the simulations across more than two decades, with its largest errors in the crossover (plotted in § 6). It overpredicts the prolate by up to 15%, near . The oblate error changes sign: the closure overpredicts the spread by up to 8% near and underpredicts it by up to 6.5% near , and the error falls to 1% by . The prolate error decays more slowly, to 5% at and 3% at 6. Under the projected-vector forcing a column’s azimuth is free, and the particle turns towards the horizontal part of the ambient vector, which then pulls it towards the horizontal as well and adds to the restoring rate at any settling number; freezing the azimuth in these simulations removes the deficit, leaving an error of 1–3% of the opposite sign (Appendix F). That mechanism belongs to the vector model, whereas the weakening of the forcing felt along a path appears under both forcings (§ 8.1). At the smallest settling numbers tested, and , the error in these second moments is only 1–6%, but the departure from the uniform density, , is about 1.5 times that of the closure for the column and 1.5–1.6 times for the plate, and from to 0.03 the ratio shows no resolved decrease towards unity. We attribute this to the ballistic motion below : the forcing turns the orientation through an angle of order before it decorrelates, and a weak drift then shifts the density more than diffusion of the same strength would. The closure thus has the uniform limit right but, at the torques tested, not the approach to it, and on this argument the discrepancy should grow with ; the plate simulations of § 8.2 show a factor of 1.6–1.9 at their weakest torques. Without the coloured-noise factor the concentration would be too low by a factor of two at and nine at (dashed curves).
Figure 2. Tests of the closure for the column () and the plate () at . (a) for the column and for the plate: solid, the closure (24); dashed, the same without the factor ; symbols, simulations with projected-vector forcing. (b), (c) Stationary densities at three settling numbers from tensor-forced simulations (symbols), against Eq. (22) at the concentration (24).
Panels (b) and (c) compare the full stationary densities. The forcing there is built from Ornstein–Uhlenbeck tensors and with the covariances ((16)–(17)), and is evaluated from Eq. (14), so neither the projection nor the tangential isotropy is imposed by hand; the two constructions agree on the variance to within 1% for the column and 2.5% for the plate, so the closure errors above carry over, the prolate maximum becoming 16% (Appendix F). The densities follow Eq. (22) across the crossover, the largest departure being in the far tail of the oblate case at , where the measured density lies above the closure. The forcing in these tests is Gaussian; resolved turbulence is intermittent, which reshapes the density (§ 7.2). § 8.2 tests the result against direct numerical simulation.
Limiting forms and the rapid-settling coefficient
Every even moment of under Eq. (22) is available in closed form (Appendix D), whereas angle moments such as need a quadrature (Figure 3a). As the mean-square tilt saturates at the uniform-measure values
Figure 3. Moments of the Boltzmann density at (solid, the column ; dashed, the plate ). (a) Mean-square tilt, with the random-orientation limits (26). (b) The slip shape factor of Eq. (30).
and once aligned it falls as
twice that for a bipolar state, the law of the rapid-settling theories [29, 43, 12]. Equation (27) drops the 1 in as well as taking the small-tilt limit, and the first causes most of the error: for the column at its error is 85% at and still 9% at , whereas the small-tilt result with the full concentration (24) is only 7% low at .
In the aligned limit , Eq. (24) gives , with cancelling between the diffusivity and the coloured-noise factor, so the limit carries no modelling parameter and can be compared directly with the rapid-settling theories. Following Gustavsson et al. [12, 13], write the relaxation rate as , with their dimensionless torque coefficient, a settling number built on the particle response time and the translational resistance along gravity, so that and . Equation (27) then becomes , with
where counts the restored directions (§ 2.3). This is the shape coefficient of the rapid-settling theory, obtained in the slender-body limit by Menon et al. [29] and Roy et al. [43] and for arbitrary aspect ratio by Gustavsson et al. [12, 13], recovered here under the same fixed-orientation forcing. A white-noise branch arises only if crossing sets with . Equation (19) then gives , the exponent of the many-eddy limit of a random field traversed at large mean speed [21], applied to settling pairs by Patra et al. [37], and the one with which Gustavsson et al. [13] corrected the of Klett [24]; it needs , hence , which only nearly spherical particles satisfy.
The density of Roy et al. [43] describes the persistent end of the same dynamics. They push a shear-rate distribution, from data or from a stochastic gradient model (Girimaji & Pope [10]), through the equilibrium (9), the limit in which the tilt follows the forcing instantaneously; the Boltzmann form is exact at the white-noise end, and Eq. (24) carries its concentration between the two. The densities coincide near alignment, where Eq. (9) linearises to , and part at large tilt, the map giving for Gaussian shear of variance against the Boltzmann form . The map also requires an equilibrium at each value of the forcing. A disc in vertical shear always has one, and so does a particle under the frozen projected-vector forcing of § 4.2, which adds the gradient of to the drift and leaves a gradient flow on the sphere. A column turned by a rigid rotation of rate about the horizontal axis normal to it obeys and loses its equilibrium once , as under strong vortical forcing; the Fokker–Planck description does not require one.
A worked example
Take the two design shapes at , in the middle of the crossover: at the reference value , or for the physical column and 4.83 for the physical plate (§ 4.2). Equation (24) gives for the column and 7.52 for the plate, and the moments follow at once: and 0.852. A Rayleigh scattering calculation needs the orientation average of the squared projection of on a polarisation direction, which for a horizontal direction is : 0.467 for the column and 0.074 for the plate. Random orientation gives , and perfect broadside alignment or 0. Neither idealisation is close for the plate, and the column lies between the two, nearer the aligned value. A calculation weighted towards large tilts also depends on the tails, which in turbulence differ from those of a single Boltzmann density (§ 7.2).
Slip-velocity statistics
With the quasi-static slip (10) in Kolmogorov units, with , the vertical and lateral components are
with the unit vector along the horizontal projection of . For an axisymmetric density, the azimuth being uniform,
and . Here is the mean vertical slip; the mean settling velocity adds the fluid velocity sampled along the path, biased downwards by preferential sweeping and absent from the overdamped slip. Equations (29) and (30) show that all orientation dependence enters through the mobility contrast , which vanishes for a sphere, and that the lateral slip is set by a fourth orientation moment, which makes it more sensitive to the tails of the density than the settling speed.
The lateral slip is odd in the tilt, so its rms is first order in the tilt spread, whereas the vertical fluctuation is even and second order: near alignment a particle wanders sideways before its fall speed changes measurably. In the random limit and ; once aligned collapses onto the mean-square tilt, so that while . The same coupling through , driven by a deterministic wave field rather than by turbulence, sets the drift of spheroids settling through surface gravity waves [30].
For the Boltzmann density is elementary (Appendix D). It runs from at random orientation to for a girdle and for a bipolar state once aligned, so that
is available at any settling number without quadrature. Figure 3(b) shows ; for plates it overshoots at intermediate , because, as grows, weight moves from the equator towards the poles through , where is largest.
Maximum of the instantaneous rms lateral slip
The lateral slip is a fluctuation, , and Eq. (31) is its root-mean-square. It vanishes at both ends, through the amplitude at small and through the loss of tilt at large , so it peaks in between, at . The maximum is over the settling speed in a fixed flow, for instance over particle size at fixed aspect ratio, and not over the dissipation rate for a fixed particle, whose dimensional slip simply follows (Figure 3b). Writing , so that , stationarity of requires
The right-hand side lies between and for any shape, and depends on alone, so at the peak falls in a narrow window: – for prolate shapes from to and – for oblate ones from to . Shape then enters only through the forcing variance,
The peak therefore depends only weakly on shape: runs from 5 to 8 between a sphere and a slender rod or thin disc, and varies roughly as its cube root, so that
from 0.860 at to 0.998 at , while the torque coefficient itself rises from 0.12 to 0.32 (Appendix A). The forcing statistics move by as much as the shape does (see below). Nearer the sphere tends to 0.814 from the prolate side and 0.870 from the oblate side, since the attracting set remains a girdle or a pole pair however small becomes, but there and vanish and the peak is unobservable. Elsewhere, locating the peak gives the torque coefficient,
Figure 4 illustrates the maximum. Integrating the orientation equation together with the particle position, Eq. (29), without the displacement by the carrier fluid, gives paths that wander sideways by a few Kolmogorov lengths over a hundred lengths of fall (panel a), and the rms slip in the simulations of Figure 2(a) follows Eq. (31) and peaks close to the settling number of Eq. (34) (panel b).
Figure 4. Lateral slip of the two design shapes at . (a) Slip-only paths in the vertical plane at the peak, eleven per shape from one release point. (b) Root-mean-square lateral slip in Kolmogorov units: curves, Eq. (31); circles, the simulations of Figure 2(a).
Figure 5(a) collects the second-moment test of § 4.2: the closure error in the spread about the attracting set under both forcings, and for the column with its azimuth frozen. The inversion below rests on the location of the maximum, which that test does not establish, since is set by the logarithmic slope of , a fourth-moment quantity, through the peak. We therefore located the maximum of in tensor-forced simulations of the two design shapes, on seven values of spaced by 0.1 in and driven by common random numbers to resolve that slope (Appendix F). They give for the column and for the plate, against 0.925 and 0.998 from the closure (Figure 5b); the uncertainty covers sampling and the choice of fit, and doubling the time step moves either value by at most 0.002. Across the grid the closure overestimates by 10% falling to 7% for the column, and by 4% for the plate, turning into a 1% underestimate: the error decreases through the peak, towards the aligned limit, so the simulated decays more slowly and peaks 3% and 4% higher in . Taken at its closure value, therefore underestimates in Eq. (35) by 5% and 8% and overestimates in Eq. (36) by 11% and 19%. Since at the peak varies little within each family, similar corrections should apply to the other aspect ratios of Eq. (34), though the simulations cover only these two shapes.
Figure 5. Closure accuracy and the slip maximum for the column (circles) and the plate (squares). (a) Closure error in the spread, , with for the column and for the plate: open symbols, projected-vector forcing; filled, tensor forcing; triangles, the column with its azimuth frozen. (b) The slope , whose value (dashed) marks the maximum, at : lines, the closure; symbols, tensor-forced simulations at (filled) and (crosses); vertical lines, from the closure (dotted) and the simulations (solid).
The location of the peak also depends on the forcing actually felt. If that is a fraction of the fixed-orientation variance (3.5), lowering from 1 to 0.6 lowers from 0.925 to 0.80 for the column and from 0.998 to 0.86 for the plate, so the fixed-orientation value overstates by 33% and 35%. The peak therefore constrains and the forcing statistics together: with calibrated it returns , and with computed from the shape it returns , a scalar measure of the sampling bias of § 8.1. Dissipation intermittency raises by up to 11% for the column and 14% for the plate (§ 7.2).
Inverting the peak
If the shape and size are known, so is , and the peak constrains the flow instead. From the Kolmogorov velocity is , with the settling speed at the peak in the convention in which is quoted, so that
The inversion amplifies the uncertainty in and in the forcing statistics: since , a 7% error in becomes 31% in . For columns in cirrus with and , the peak falls at a broadside speed (, ), or an equivalent-sphere speed of (), reached by ice columns about long; at it moves to and columns about long. Whether such crystals are overdamped is examined in § 9; in air the condition is restrictive.
Both readings use the slip speed, not a displacement, and the lateral dispersion need not peak where the slip does, since it depends also on the correlation time of . For an aligned plate the lateral slip follows the two restored tilt components, whose integral time is by the linear response (19); for a column , and the horizontal direction keeps diffusing in azimuth at any settling number, so its decorrelation enters as well. A displacement also carries the carrier-fluid velocity and its correlation with the slip, which differencing against a sphere does not remove, the two particles following different paths. We therefore state the result as a maximum of the instantaneous slip.
Because the maximum is flat, its position is hard to measure precisely, but the closed form can be used at any settling number. The ratio of the lateral rms to the mean vertical slip,
is dimensionless, needs no knowledge of , and depends only on Sv once the shape and the forcing model are fixed. For columns with , that is (Appendix D), which includes the range treated here, it falls monotonically, from at random orientation to a slope near once aligned, so one measurement of fixes Sv, and follows from the settling speed; for more slender columns rises by less than 1% before falling. For plates first rises, by about 18% to a maximum near at (, ), and only then falls, so a second observable must select the branch.
With computed from the shape (Appendix A), Eq. (34) places the peak at for the column and 4.8 for the plate, or and 1.8 on the broadside speed ( and 0.32), and at and 2.5 with the finite- factor of Appendix A.
Extensions
General torques
Since any axisymmetric drift is a gradient flow (Sec. 2.2), white forcing with isotropic statistics gives exactly for any such torque. Higher harmonics from finite- corrections, nonlinear drag entering through an orientation-dependent slip, and superposed body-force torques are all of this form, and the density of Sec. 4.2 is the quadratic member of the class. For coloured forcing the closure of Sec. 4.2 carries over only locally, through the relaxation rate about an isolated attractor.
The Boltzmann form also needs forcing that is isotropic in the tangent plane. Forcing that distinguishes a direction, as stratified or sheared turbulence may, makes the noise depend on orientation, and the stationary density then departs from that form. Settling alone produces such anisotropy, through the crossing integral of Appendix C, only in the white-noise branch, which only nearly spherical particles reach (Sec. 4.3); in the persistent regime the tilt follows the instantaneous forcing, which is isotropic in the tangent plane.
Intermittent dissipation
The forcing so far is Gaussian, with a variance set by the mean dissipation, but turbulent velocity gradients are intermittent: the local pseudo-dissipation fluctuates over orders of magnitude, its logarithm is close to normal ([52, 38]), and its variance grows with the Taylor-scale Reynolds number ([26, 49]). We use the parametrisation of Koch & Pope [25],
and represent intermittency by a scalar amplitude: , with normal of variance and mean , so that . This models the amplitude of the forcing, not the full gradient tensor conditioned on the true dissipation. The restoring rate is fixed by the settling speed and does not fluctuate, whereas the forcing variance does, . In the aligned limit, where and the correlation time has dropped out, the local concentration is therefore
The same form holds at any settling number, with replaced by the closure concentration (24), if the correlation time does not follow the local dissipation. If the orientation equilibrates faster than evolves, the stationary density is the lognormal average of Boltzmann densities over an effective diffusivity that follows the local dissipation, a superstatistics in the sense of Beck & Cohen [2],
with the signs of Eq. (22). It has no closed form, the Laplace transform of a lognormal not being elementary, and is evaluated by quadrature over .
The superposition does not change the leading aligned mean-square tilt. For a girdle
and for a bipolar state, with no dependence on : the aligned conditional moment is linear in and . At finite the conditional moments are no longer linear in and the tilt statistics shift; at and , falls by 18% for a column and by 19% for a plate, and the angle moments shift by different amounts.
The mixture is a superposition of exponentials in with positive weights, so is convex in , where a single Boltzmann density gives a straight line; upward curvature is therefore consistent with a mixture of concentrations, though it does not establish one, still less a lognormal one. The aligned-limit moments of the mixture have closed forms. With an average over a single density of concentration , from in the aligned limit and ,
and the first new information is at , where the flatness takes the closed form
Gaussian as and, with the refined-similarity form for a constant , growing as a power of , like velocity-derivative flatness [49] (Figure 6b). Higher moments need stronger alignment for these asymptotic forms to hold, since they weight rare intense events more heavily. The most visible effect is in the tails (Figure 6): at the mixture exceeds the Boltzmann density of the same by two orders of magnitude at , so large tilts, produced by rare intense dissipation events, are far more frequent than under Gaussian forcing. The peak of the girdle rises as well, by , at a fixed leading mean-square tilt, while the mean of falls by the factor , by 7% and 14% at and 1.2.
Figure 6. Dissipation intermittency in the aligned limit. (a) The superposition (40) for a girdle at with two log-dissipation variances (solid), against the Boltzmann density of the same (dashed). (b) The flatness (43) and its bipolar counterpart (46), with the Taylor-scale Reynolds number from Eq. (38) on the upper axis.
Pumir et al. [40] found a qualitatively similar sharpening when they made their stochastic model intermittent, their distributions becoming more peaked as grew while the mean absolute orientation cosine changed little. These leading-order moments also apply to other systems with locally quadratic restoring potentials, provided the conditional fluctuations are Gaussian and their variance follows the same slowly varying lognormal modulation.
The same factorisation gives in closed form the constants , with the density of the shear rate folded onto , that govern the rapid-settling orientation moments in the map construction of Roy et al. [43], previously obtained by quadrature. With the shear rate, in units of its root-mean-square value, written as , where is a unit-variance Gaussian independent of ([25]),
so that
a ratio of eighth to squared fourth moments of , independent of the mean of . The number of restored directions (§ 2.3) also sets the Gaussian baseline of the flatness. Equation (43) carries the value three because a girdle restores one tangent direction; for a bipolar state is a sum of two squared Gaussians, and
so a measured flatness must be normalised by two or by three according to the attracting set.
At the slip maximum, where the closure concentration is 6.0 for the column and 7.5 for the plate, the mixture with reduces by 25% and 30% if the correlation time does not follow the dissipation, so the rms slip falls by 14% and 16%. If instead follows the local Kolmogorov time, the local concentration follows from Eq. (19) with that time rather than from , and falls by 23% and 26%. The maximum moves too: rises from 0.925 to 1.031 for the column and from 0.998 to 1.141 for the plate, or to 0.941 and 1.045 when follows the dissipation. Read with the constant of a single Boltzmann density, a mixture peak would underestimate by up to 20% and 23% and, with known, overestimate by up to 55% and 71%.
The mixture also assumes that the dissipation stays fixed while the orientation equilibrates. Multiplying the forcing by a lognormal amplitude of correlation time (Appendix F) tests this assumption: for the column at and , the measured is 0.83, 0.87, 0.90 and 0.90 times the static superposition of closure densities for , 1, 4 and 20, and 0.91 when each particle keeps a frozen amplitude. The modulated runs level off by , about 1% below the frozen-amplitude value, and the remaining 9–10% is the closure’s own error, about twice its value of 5% at uniform dissipation (§ 4.2), because the mixture weights the weakly aligned states in which the closure is least accurate. Through the crossover the mixture is thus a sensitivity estimate, and it indicates that the single density overestimates the lateral slip; anything weighted towards large tilts, such as the fourth moment, the flatness or the probability of finding a crystal far from broadside-on, needs an account of the non-Gaussian forcing, of which the mixture is one tractable model.
Tests against simulated turbulence
This section examines the forcing felt along a trajectory, which the closure takes from fixed-orientation statistics, and compares the results with direct numerical simulation.
Correlation between orientation and the sampled gradients
The closure cannot absorb a correlation between and the instantaneous gradient, which § 3 assumes absent. Gravity alone does not spoil the one-time statistics: at fixed the velocity sampled by an inertialess particle is divergence free, which removes the usual mechanism of preferential concentration (though it does not prove uniform sampling, and evolving together), and settling changes the time correlation, through the crossing-trajectory effect [7] represented by the time . The orientation is, however, correlated with the gradient, because the orientation a particle has reached depends on the forcing that brought it there. For forcing continuous in time, as here, the conditional mean is an ordinary conditional expectation and does not vanish: in a stationary state without probability current it cancels the deterministic drift (Appendix E). Ideal white noise has no instantaneous value, and the corresponding statement is the zero-current balance of drift and diffusive flux from which the Boltzmann density follows. Neither is a drift to be added to a Fokker–Planck equation, and neither bears on the accuracy of the closure, which depends on the strength and distribution of the sampled forcing.
We measured the strength and distribution of the sampled forcing in both synthetic models. Under the projected-vector forcing, along the trajectories of Appendix F for the column, at and 0.218 at , against the fixed-orientation value : the forcing felt is 12% and 17% weaker. At the attained orientation and 0.176, against the isotropic , so the ambient vector points along more often than chance, having driven the particle there, and the tangential part that survives the projection is depleted. That mechanism is specific to the vector model, whose forcing is even under in each realisation (§ 4.2). The weakening is not: in the tensor-forced runs of table 3, at , the conditional second moments weighted by the population of their bins give for the column and 0.247 for the plate, 11% and 7% below the fixed-orientation values 0.263 and 0.266. The conditional second moments also depend on orientation and differ between the meridional and azimuthal directions (Appendix E).
We do not turn these measurements into a corrected equation for the orientation alone. For Gaussian gradients such an equation could start from the identity of Novikov [35], which writes averages of through the gradient covariance and the response of to the gradient history; that response decays as only in the aligned linear limit and only in the restored directions, and not at all for the azimuth of a column. A Gaussian calculation would also miss the source of the rod–vorticity alignment, which arises because and obey similar equations of motion (Pumir & Wilkinson [41]), whereas the vorticity of a prescribed Gaussian gradient is not stretched by its strain. The net effect of the correlation on the concentration can be judged only against direct simulation, with which the next section compares the results for small plates and columnar crystals.
Comparison with direct numerical simulation
The orientation density
Anand et al. [1] report orientation distributions of for oblate spheroids settling in homogeneous isotropic turbulence at , for three aspect ratios at fixed settling Reynolds number, . Their model shares our starting point: the gradient covariances they quote are Eqs. (3.3)–(3.4), their rapid-settling variance carries the combination of Eq. (3.5), and their gravity-induced torque is evaluated with the quasi-static slip, the mobility acting on gravity, as here. Their particles have translational and rotational inertia, at Stokes numbers –, but the slip fluctuations that inertia produces do not enter their torque, so the comparison tests the overdamped torque balance in resolved turbulence rather than the finite-inertia mechanism of Gustavsson et al. [13]. Rotational inertia is not negligible, however: for a thin disc the rotational response time is , with the particle response time of their model, so –. Under Ornstein–Uhlenbeck forcing the linearised oscillator of § 2.2 has a tilt variance times the overdamped value (19), so with rotational inertia would raise the linear tilt variance of these runs by 4–10%.
We digitise their Fig. 1(a) over , three to six decades in density, and plot it against , on which a single Boltzmann density is a straight line (Figure 7a). A single Boltzmann density does not describe these data: its residuals are large and systematic, the measured density curving upward in and rising towards the broadside orientation faster than a single exponential. § 7.2 predicts curvature of this sign. Mixing Boltzmann densities over a lognormal dissipation with the variance estimated from the Reynolds number, , and fitting only cuts the residuals five- to sevenfold with the same number of parameters; letting the variance float lowers them further and returns , 0.87 and 0.90, three values that agree with one another and lie 25–30% below Eq. (38). The fitted variance is an effective parameter of the model rather than a measurement of the dissipation statistics: it absorbs the torque approximation, the temporal response, the sampling along trajectories and the digitisation, and for it rises from 0.90 to 0.97 when the points above are dropped, those of the two thinner discs moving by 0.02 or less. The bin convention of the source matters near the pole. Its markers are spaced by 0.02 in ; read as means over bins of that width, the last marker, 52.3 at , would carry probability 1.05 by itself and that curve 1.12 in all, against 1.01 and 0.96 for the other two. Averaging the mean over bins of width 0.02 lowers the one-parameter residual of the mixture for from 0.26 to 0.16 and raises the variance fitted with to 0.97, with residual 0.084, and that of the thinner discs to 0.85 and 0.89; the preference for the mixture is unchanged. The fits constrain only the unconditional density, so they cannot establish that the density at a given local dissipation is a Boltzmann density.
Figure 7. Comparison with the plate DNS of Anand et al. [1] at . (a) Densities of at : symbols, the data; dashed, a single Boltzmann density with fitted; solid, the superposition (40) with and fitted; dotted, the superposition at the predicted concentration (24). (b) Their gradient-based ratio against , Eq. (46). (c), (d) Measured moments (triangles) against with the single density (dashed) and the superposition (solid) at the predicted concentration; dotted, the uniform density.
The concentration can also be predicted from the run parameters. The supplementary material of Anand et al. [1] lists the Froude number of each run, 2.1, 2.7 and 3.7 for , 0.02 and 0.05. Their is the slip with the symmetry axis vertical, the broadside orientation of a plate, so ; and their rapid-settling variance (their equation 5) is the aligned limit of Eq. (24), with their resistance functions combining into , 0.325 and 0.311 (Appendix A). Equation (24) with then gives , 29.9 and 84.3, with nothing adjusted. The mixture fits return about a quarter above these at , and within 11% of them when the variance is fitted as well. With nothing fitted at all, the mixture at the predicted concentration follows the data to within a typical factor of 1.5 to 2.0 (Table 1) over three to six decades, whereas a single Boltzmann density at the same concentration falls short at large tilt by two to thirty decades. The densities are digitised, and their normalisation on is assumed.
| 2.1 | 2.7 | 3.7 | |
| 13.4 | 29.9 | 84.3 | |
| rms residual, nothing fitted | |||
| Boltzmann, | 3.7 | 13.4 | 49.5 |
| mixture, | 0.38 | 0.49 | 0.71 |
| power law, | 0.17 | 0.21 | 0.69 |
| rms residual, one parameter fitted | |||
| Boltzmann, | 0.71 | 1.14 | 1.92 |
| mixture, | 0.13 | 0.16 | 0.26 |
| power law, | 0.073 | 0.14 | 0.43 |
| rms residual, two parameters fitted | |||
| mixture, and | 0.052 | 0.058 | 0.18 |
| power law, and | 0.072 | 0.050 | 0.11 |
| fitted parameters | |||
| Boltzmann, | 7.7 | 10.3 | 14.2 |
| mixture, at | 16.6 | 37.2 | 107 |
| mixture, and | 0.84, 14.8 | 0.87, 30.6 | 0.90, 79.7 |
| power law, (implied ) | 3.51 (1.08) | 3.90 (0.75) | 4.42 (0.53) |
Table 1. Prediction and fits for the orientation distributions of Anand et al. [1] in Figure 7(a), digitised over and grouped by the number of fitted parameters. Every density is normalised on , with no free amplitude, and evaluated at the abscissae of the markers (§ 8.2 gives the effect of averaging over bins). Residuals are rms deviations of over the digitised points, so is the typical factor between model and data. is Eq. (24) with and from Appendix A. The mixture is Eq. (40) with from Eq. (38) unless is fitted; the power-law form is Eq. (47), bipolar, with from Eq. (48) unless is fitted, and the implied by a fitted follows by inverting Eq. (48).
Moments through the crossover
These densities come from three runs at one settling Reynolds number. The main panels of Figure 2 of Anand et al. [1] give, for every run, the measured moments , the spread about the broadside orientation, and , which is of fourth order in the tilt near alignment and weights large tilts more strongly; at the runs span the crossover, from to 4.2. We read seven of the nine runs (two overlap near ; Appendix F) and compare them in Figure 7(c,Figure 7d) with the single Boltzmann density and the mixture at , both at the concentration (4.6) predicted from the run parameters, nothing fitted; the curves are drawn for , those for and 0.05 differing by less than 1%. The spread is predicted: from upwards the single density is within 3–13% of the measured and the mixture within 0.5–11%. The tails are not, or not uniformly: from upwards the single density underestimates by factors of 2.4–2.9, and the mixture overestimates it by an amount that grows with settling, from 16% to 61%. At the three weakest torques, –0.26, the data lie between the two forms, their departure from the uniform value being 1.6–1.9 times that of the single density, a deficit of the sign and size found against the stochastic simulations (§ 4.2). Reading the markers at their pixel centroids, or the settling numbers from the rounded values of the supplementary table, changes these ranges (Appendix F) but not their ordering: the single density stays low by factors of 2.4–3.1 in , and the excess of the mixture grows with settling. The runs have Stokes numbers 0.06–0.40, so the comparison is not strictly overdamped. Agreement in supports the predicted width through the crossover, but it cannot identify the dissipation mechanism, and it does not locate the slip peak, which depends on the fourth moment through (§ 6). A corresponding test for columns, and a measurement of the slip peak in resolved turbulence, remain to be made.
The fourth moment
A third test, with no fitted parameter, uses the fourth moment; it tests the amplitude model of § 7.2 rather than the three densities of Figure 7(a). Anand et al. [1] evaluate (with the signed , would tend to 2, not 0, for a plate equally likely at either pole) for a thin disc in the rapid-settling limit, where the tilt follows the instantaneous gradient, from their theory with fourth moments of the velocity gradients measured in their direct numerical simulations (DNS; inset of their Figure 2(b)), and trace its growth with to the intermittency of the gradients. In that limit their expression reduces to , with , so is 2/9 of the flatness of a transverse velocity gradient: a property of the gradient statistics, not a fourth moment of the measured densities. For small tilt , and the superposition gives it the closed form through Eq. (46). The Gaussian value is because a disc is restored in two tangent directions, and for Gaussian gradients gives the same value; the unity that Anand et al. [1] quote as Gaussian holds for a single tilt component. With from Eq. (38) (Figure 7b) the prediction runs 12–22% above their four values, 1.27, 1.70, 1.99 and 2.04 at , 96, 150 and 200, but reproduces their level and trend: the log-dissipation variance implied by grows as , against 0.31 for the parametrisation, a consistency check of the amplitude model rather than a measurement of the exponent from four points. Every value lies far above , and above unity as well, so normalising by three, as though the tilt had one component, understates the departure from Gaussianity. The moments of § 8.3 give directly for the three runs with : 2.14, 2.32 and 2.17, close to their inset value 1.99 at and to the aligned limit 2.21 of the superposition, whereas the superposition at the concentrations of those runs gives 2.5–3.1 and the single density 0.68–0.83.
Finite Stokes number
The runs of Pumir et al. [40] are at finite Stokes number, outside the overdamped regime (§ 9), and their distributions show what particle inertia does to the shape. Their Figure 2 gives for a column of aspect ratio four in four flows, at Stokes numbers from 0.31 to 4.9, on linear axes. We use the vertices of the published curves, read from the vector graphics of the article, which resolve the density to about , three to four decades below the peak (Appendix F). Fitted on the linear axes of the source to the points above a fiftieth of the peak, curves describe the cores to 3–3.5% of the peak height in rms, and to 1.5% in the least aligned flow, with concentrations 555, 171, 50.7 and 4.37 (Figure 8a). The core concentration depends on the extraction and on the fitted range: the same fit to an independent digitisation of the plotted curves gave 667, 187, 52 and 4.4, and raising the threshold to half the peak gives 741, 222, 64 and 4.9. Divided by the fitted amplitude and plotted against (Figure 8b), the three aligned flows approximately collapse onto a common profile, which curves upward on these semilogarithmic axes where a single Boltzmann density is straight. At these Stokes numbers that curvature need not come from intermittency (§ 7.2) alone. Within the fitted range the data rise above the fit by factors up to 12–16, and by 150–190 at . The normalised Boltzmann density at the fitted concentration underestimates the tabulated mean-square tilt by 36–38% in the three aligned flows. The fits therefore describe the width of the core, not the shape of the distribution, and give no evidence that the Boltzmann form survives inertia. The fitted concentrations describe the simulated distributions; they are not predictions. The prediction is the Stokes-drag overdamped reference of § 9, computed from the run parameters with nothing fitted, and the three quantities differ: in flow F1, the least turbulent, the tabulated rms tilt is , the Boltzmann density at the fitted core concentration gives , and the overdamped reference gives . Over the four flows the core concentrations are 1.7–2.9% of Eq. (24), a core six to eight times broader in angle, and the tabulated rms tilt is 6.6–9.6 times the reference, which indicates what inertia adds beyond the closure; the finite- drag left out of the reference would raise it by about 42% for these particles.
Figure 8. Orientation distributions of a column of aspect ratio four from Fig. 2 of Pumir et al. [40], in four flows labelled by their Stokes number. (a) The data (symbols) and fitted curves (lines). (b) The data divided by , against , on which each fit is the dashed line ; filled symbols lie in the fitted range.
A form for comparison with data
The comparisons above use the mixture (7.3) directly: a one-dimensional quadrature with one free parameter, , once is estimated from Eq. (38). Where a closed-form density is wanted, as in a scattering calculation, a power-law density, normalised on , can
stand in for it:
where is a fitting scale multiplying for a girdle and for a bipolar state; the Boltzmann form is recovered as at fixed . Matching the aligned-limit variance (41) and flatness (43) of the mixture, or for a bipolar state and Eq. (46), fixes both parameters in terms of and :
The exponent falls towards and 3 as grows, so the tail is a power law of low index in the tilt. At , where , Eq. (48) gives for a bipolar state, just below the 3.51, 3.90 and 4.42 fitted, together with , to the three distributions of Figure 7(a).
Table 1 compares the two forms at equal numbers of fitted parameters: with none fitted the surrogate does better, and with one or two neither is consistently better, as expected of a form that carries the mixture’s aligned-limit variance and flatness. The forms differ in the meaning of their second parameter. In the mixture it is an effective log-dissipation variance, and fitting it returns nearly one value, 0.84 to 0.90, for all three shapes, as expected of a property of the flow; in the surrogate it is an exponent, tied to that variance only through the aligned-limit moments, and the variances implied by its fitted exponents drift with shape, from 1.08 to 0.53. We would use Eq. (47) where a closed-form density is wanted and Eq. (40) where an effective dissipation variance is to be fitted or carried from one flow to another.
Application to atmospheric ice and frazil
The crossover is the regime of practical interest: for crystals of tens of microns in cloud turbulence, and for frazil discs in supercooled water, the alignment parameter passes through order unity within a single cloud or plume. The formulas apply as they stand only if the translational and rotational response times of the particle are short against the forcing and alignment times, , so a small Stokes number based on alone is not enough when the alignment is fast. In air this is restrictive: ice is about a thousand times denser than air, and Gustavsson et al. [13] showed that translational inertia then feeds slip-velocity fluctuations into the angular equation and raises the mean-square tilt by orders of magnitude for slender columns and thin plates, even at small Stokes number, finding overdamped dynamics for platelets in cloud conditions only for . Outside that window the present results are the overdamped reference, the alignment a crystal would reach if its slip did not fluctuate, and not a prediction of the observed distribution: at finite inertia is itself a fluctuating vector coupled to the orientation and to the sampled flow, and eliminating it can produce multiplicative forcing and non-gradient drift, so whether the Boltzmann form survives there is an open question.
The overdamped condition is easier to meet in water. A frazil disc 1 mm across and 50 thick () rises broadside-on at about 1.5 mm s at 0 °C, with a diameter-based Reynolds number near 0.8. In turbulence with – its diameter is 0.36–0.65 Kolmogorov lengths, its Stokes number is of order , and –1.3, so with it sits in the crossover, at –3. For a nearly neutrally buoyant disc the added-mass and history forces need assessment before quantitative use; the particles of the settling experiments of Lopez & Guazzelli [27] and Roy et al. [42] are in a similar position.
A scattering measurement returns the orientation average of a cross-section. In the Rayleigh regime of centimetre- and millimetre-wavelength radar the polarisability of a spheroid is , so the co-polar return at horizontal polarisation is governed by and its square, which for a density axisymmetric about gravity average to
The orientation factors in reflectivity and differential reflectivity are thus set by and , which follow from the recursion (D1) for a Boltzmann density and with one further average over for the mixture. Lidar needs the full density and a non-Rayleigh calculation; the sensitivity to the tilt width found by Okamoto et al. [36] shows why the width matters, although their plates, larger than 200 , lie beyond the overdamped regime and do not test Eq. (24). For small crystals the mean-square tilt saturates at the random-orientation values (26) as falls, where the aligned-limit law would grow without bound as ; that saturation is a testable signature of the crossover, and intermittency predicts tails that broaden with the Reynolds number at a fixed leading aligned variance.
Swept-volume collision kernels for crystal–droplet and crystal–crystal encounters (Naso et al. [32]; Jucha et al. [19]; Sheikh et al. [45, 46]) need , and the slip statistics (30), all in closed form at any , where the strong-alignment expansions err by tens of per cent at the operating point (§ 4.3); crystal–crystal kernels also need pair statistics. In laboratory settling experiments (Lopez & Guazzelli [27]), where the overdamped condition is under the experimenter’s control, the lateral slip relative to the local fluid tests the peak relation of § 6, with the conventions of § 2.4 connecting it to quiescent-flow theory.
Comparing the overdamped reference with recent direct simulations first requires reconciling conventions. Pumir et al. [40] tabulate the rms tilt of columns with – and semi-major axes 100–200 m in four flows spanning –1120 . Their equations of motion carry a calibrated inertial-torque factor , so their overdamped limit has relaxation coefficient , and their settling number , built on the response time with the semi-minor axis, exceeds the physical broadside settling number by the factor –. Table 2 sets their measured tilt against the overdamped reference of their own model with Stokes drag, computed from their particle and flow parameters with no adjustable constant. For the smallest columns (, 100 m) the measured rms exceeds the reference by 3.2–3.7 in all four flows, at Stokes numbers up to three. For the largest (, 200 m) the gap opens to 10–16, and the measured tilt grows from the slow to the fast settlers of a given flow, where the overdamped scaling would have it fall, as expected when the slip fluctuations that force the orientation grow with particle size (Gustavsson et al. [13]). The reference omits two effects of opposite sign. The finite- drag of their model (their equations (C3) and (C4)) slows the broadside settling, from 21.6 to 19.1 for the smallest columns and from 37.3 to 28.3 for the largest, and strengthens the coupling of tilt to lateral slip; in the aligned limit the two together raise the reference by about 24% and 59%, which would bring the ratios to 2.6–3.0 and 6.5–10. The orientation–gradient correlation of § 8.1 would lower the reference. The ratios therefore indicate the inertial inflation without bounding it; both columns, with and 89 µm, are far larger than the platelets that Gustavsson et al. [13] find overdamped. The conditional settling speeds that Pumir et al. [40] report increase as the axis turns towards the vertical, as the anisotropic mobility of § 5 implies, with a contrast between the extreme orientations of the order of the quiescent mobility ratio in the two flows that sample near-vertical orientations; but these are laboratory-frame velocities conditioned on orientation, which do not test the instantaneous slip relation, and preferential sweeping, outside the overdamped theory, contributes substantially to the mean settling enhancement.
| run | St | DNS (°) | reference (°) | ratio | |||
| F1/P1 | 3 | 0.185 | 39.0 | 22.6 | 1.77 | 0.477 | 3.71 |
| F1/P3 | 5 | 0.439 | 92.8 | 39.1 | 2.31 | 0.145 | 16.0 |
| F2/P1 | 3 | 0.370 | 27.6 | 16.0 | 3.47 | 0.945 | 3.67 |
| F3/P1 | 3 | 0.739 | 19.5 | 11.3 | 6.40 | 1.84 | 3.47 |
| F3/P3 | 5 | 1.76 | 46.4 | 19.6 | 7.71 | 0.573 | 13.5 |
| F4/P1 | 3 | 2.94 | 9.79 | 5.66 | 21.2 | 6.54 | 3.24 |
| F4/P3 | 5 | 6.98 | 23.3 | 9.81 | 22.5 | 2.19 | 10.3 |
Table 2. Measured rms tilt of settling ice columns in the DNS of Pumir et al. [40] (their Table III; flows F1–F4, particles P1 and P3) against the Stokes-drag overdamped reference of their model dynamics, with the torque factor of their equations of motion and the physical broadside settling number , with their settling number. St and are recomputed from the particle and flow parameters of their Tables I and II, which reproduce the values printed in their Table III to within 3%, and every computed column follows from them. The reference is the rms angle , with at the concentration (24), evaluated by quadrature rather than from the aligned asymptote, since run F4/P1 sits at . The finite- drag of their model, not included, would raise the reference by about 24% for P1 and 59% for P3 in the aligned limit.
Conclusions
A settling spheroid is held broadside-on by the fluid-inertial torque and turned away by the turbulent velocity gradients. In the overdamped limit the torque is an exact gradient flow on the sphere of orientations, with a potential quadratic in the cosine of the tilt that an anisotropic mobility only rescales, and the gradients force the orientation with no preferred direction in the tangent plane, so isotropic Gaussian white-noise forcing gives exactly a Boltzmann density, exponential in the squared cosine of the tilt. Turbulent forcing is not white, and keeping that density with a width fixed by the linear response is a finite-memory closure, accurate to 16% or better in the second moment against stochastic simulations of a column and a plate with , although at the weakest torques tested the simulated departure from random orientation is about 1.5 times the closure’s. The width depends on the alignment parameter , the alignment rate in Kolmogorov units, and on the shape through the forcing factor .
The observables then follow from the density without further approximation. The even moments give the tilt statistics, the mean vertical slip and the orientation averages of Rayleigh scattering, which at crossover settling numbers need not be close to either limiting assumption. The rms lateral slip does not grow without limit as particles settle faster, because alignment quenches the tilt that produces it; its maximum sits at between 0.74 and 1.0 for shapes from a moderately elongated column to a thin plate, or 5% and 9% higher for the two design shapes in tensor-forced simulations, so a measured peak constrains the torque coefficient or, read in reverse, the Kolmogorov scales, given the forcing statistics.
Beyond the closure, the results rest on three approximations. The forcing is prescribed rather than taken from the Navier–Stokes equations, its correlation time is uncalibrated, and the orientation it produces is correlated with the gradients it samples, which weakens and redistributes the forcing felt in a way that a single coefficient does not capture. Resolved turbulence has so far tested the closure only for plates. Their measured spread follows the prediction with nothing fitted, to within 13% from upwards, and the weak-torque deficit of the closure appears there as well. At strong torque their fourth-order tilt moment lies between the predictions of the single density, 2.4–2.9 times too low, and of the superposition, whose excess grows with settling. A test for columns, and a measurement of the slip peak, remain to be made. Intermittency enters through a lognormal amplitude that varies slowly compared with the alignment. The superposition reproduces the shape of simulated densities for small plates: with nothing fitted it follows them within a typical factor of 1.5 to 2, and fitting its concentration returns values about a quarter above Eq. (24) with , although such fits cannot identify the mechanism (§ 8); through the crossover the superposition is a sensitivity estimate. Finally, the analysis is overdamped, whereas atmospheric crystals larger than a few tens of microns carry enough inertia to inflate the tilt several times over. Within these limits, the closed-form density and its moments can replace the usual choice between random orientation and perfect alignment in scattering and microphysics calculations.
Acknowledgements. The digitisation of the published data and the Langevin simulations were carried out with Claude Opus 5 to 5.5; all scientific content, analysis and conclusions are the author’s.
Funding. This work was supported by the Anusandhan National Research Foundation under project ANRF/ARG/2025/010666/ENS and by the ISRO–IIT Madras Space Technology Cell project SP26271464 AMISRO008746.
Declaration of interests. The author reports no conflict of interest.
Data availability statement. The Python scripts that generate the figures, the tables and the numbers quoted in the text, the Langevin simulation drivers with their job lists and seeds, the simulation outputs, and the data taken from Anand et al. (2020) and Pumir et al. (2026) (digitised points, vertices of published curves and transcribed parameters) are archived at Zenodo (https://doi.org/10.5281/zenodo.23922242); the scripts and data added in this version for § 8.1 and § 8.2 are available from the author.
Appendix A. Ellipsoidal mobility integrals
For a spheroid with semi-axes the Oberbeck resistance gives the settling speed along the symmetry axis as and along a transverse axis as under a force , with
where , so that, relative to the volume-equivalent sphere of radius ,
All three integrals reduce to inverse hyperbolic (prolate) or inverse trigonometric (oblate) functions of the eccentricity, tabulated by Happel & Brenner [15] and Kim & Karrila [23]. For the two design shapes, gives , , and , and gives , , and . The forcing factor varies only between 5 and 8 over the whole range of aspect ratio, which is why the peak constant of § 6 depends only weakly on shape.
The torque coefficient
The bare drift coefficient of Eq. (3) is , where is the torque amplitude in
with [8, 17], and is the couple resisting rotation about a transverse axis,
We take from the closed-form expressions of Dabade et al. [8] (their equations 4.1 and 4.2), with eccentricity for prolate and for oblate shapes; these return the near-sphere limit , the slender-body limit, and the flat-disc value given there, and the prolate amplitude peaks at . In the broadside convention of Eq. (13) they give , 0.158, 0.175, 0.183 and 0.189 for columns of , 3, 4, 5 and 10, and 0.329, 0.325, 0.311 and 0.320 for plates of , 0.02, 0.05 and 0.03; in the equivalent-sphere convention the column has . These are creeping-flow values. At the particle Reynolds numbers of real crystals, , the amplitude falls below them [17], and Pumir et al. [40] use a calibrated correction ; for slender fibres the inertial slender-body theory of Joshi et al. [18] should be used instead.
Appendix B. Validation summary
The closed-form results have been checked against quadrature or simulation as follows: the forcing law (18) by Monte Carlo sampling ( realisations) of Gaussian and at four values of , including the tangent-plane isotropy of the strain contribution; the moments (D.2)–(D.3) of the Boltzmann density and the shape factor (D.4) against 40-digit quadrature for from to on both branches, to about , the quadrature of the text using a 400-point Gauss–Legendre rule in with the exponent shifted by its maximum for , and for a 200-point rule in the scaled variables (girdle) or (bipolar), which resolve the boundary layers of width and ; the aligned mixture moments (7.5)–(7.9) by quadrature at and , to 0.01% in the variance and 0.3% in the flatness for a girdle and to 0.5% and 2.4% for a bipolar state, whose corrections are of relative order ; the random limits (26), the slip identities and and the relaxation (8) to quadrature and integrator precision; the mobility values against the sphere limit, the thin-disc speed ratio , approached as , and the logarithmically slow approach of to 2 (1.526 at ); the identity (11) to a residual of ; and the linear response (19) by Ornstein–Uhlenbeck-forced simulation at three pairs, to 0.3%.
Appendix C
Correlation time along a settling path
In the rapid-crossing, frozen-field approximation the time-integrated gradient correlation along a straight settling path follows from the energy spectrum. With as in Eq. (15) and in Kolmogorov units, the time integral collapses onto the plane ,
with the wavenumber magnitude and a unit vector in that plane. Dividing the contracted integral by the zero-lag variance gives the correlation time of the tangential forcing, , with fixed by the spectrum rather than assumed. Evaluating Eq. (53) with the model spectrum of Pope [38], at its high-Reynolds-number constants, and contracting it with Eq. (14) and the tangent projector at the attracting orientation, which averages over the two tangent directions, gives , 16.4 and 17.7 for the design column at , and , and 13.9, 18.7 and 20.2 for the design plate; the integral converges as , so saturates near twenty Kolmogorov lengths rather than growing with the Taylor microscale. The crossing time therefore falls below only for –20 (§ 3), and where it does, as in the faster runs of table 2, the alignment is strong enough, , for to cancel from Eq. (24).
The same integral shows that sampling along a settling path is axisymmetric about gravity rather than isotropic. At the girdle the meridional part of the integrated forcing exceeds the azimuthal part by the factor , independent of but strongly shape dependent: 1.32 for the column, 1.08 at , and as , so the sense of the anisotropy reverses for nearly spherical columns. The splitting matters only in the white-noise limit branch (§ 7.1), where the tilt of a column is set by the restored, meridional component alone, whose correlation length is , 14% above the average for the column. At a pole the two tangent directions are equivalent by symmetry, with no splitting.
Appendix D
Moments of the Boltzmann density and the slip shape factor
Integrating over relates successive even moments , with :
the superscript denoting the sign in the exponent of Eq. (22): the girdle branch for columns, the bipolar branch for plates. Angle moments such as , with for a girdle and for a bipolar state, , are evaluated by quadrature.
The single moments used in the text are, with for the girdle branch,
the bipolar pair written with the Dawson function for numerical stability. The random limits (26) follow from and , the difference being the measure on the sphere near a girdle or a pole.
The slip shape factor of Eq. (30) is , so that
with and from Eq. (23), that is
Both reduce to as , and for strong alignment
Equation (D 4) loses precision below , where the numerator is a difference of nearly equal terms, and quadrature should be used there. The vertical fluctuation follows from the same moments,
whose bracket tends to in the random limit and falls as once aligned, so that : near a pole is exponentially distributed with variance , and near a girdle is a scaled one-degree-of-freedom chi-square variable.
The threshold quoted for the slip ratio (6.6) in § 6.1 follows from the same moments. With and primes denoting derivatives with respect to on the girdle branch, where ,
which increases with at fixed . Near random orientation and , so at the slope is , which vanishes at , that is . At that ratio Eq. (D 8) is negative for all (evaluated for ), so falls monotonically for every column with ; for larger ratios it rises first, by an amount that grows with the ratio and tends to about 0.72% in the slender limit .
Appendix E
Conditional forcing statistics
Resolving the forcing in the tangent frame, into the meridional direction , along which the tilt is restored, and the azimuthal direction , gives the second moments of Table 3. They are not equal, and their ratio varies across the sphere, from 0.64 to 2.33 for the column and from 2.1 to 4.3 for the plate.
| (attractor ) | (attractor ) | ||||||
| ratio | ratio | ||||||
| 0.08 | 0.073 | 0.114 | 0.64 | 0.222 | 0.105 | 2.11 | |
| 0.25 | 0.123 | 0.117 | 1.05 | 0.254 | 0.103 | 2.47 | |
| 0.42 | 0.208 | 0.123 | 1.70 | 0.310 | 0.096 | 3.23 | |
| 0.58 | 0.300 | 0.133 | 2.26 | 0.353 | 0.088 | 4.04 | |
| 0.75 | 0.353 | 0.151 | 2.33 | 0.330 | 0.077 | 4.29 | |
| 0.92 | 0.312 | 0.179 | 1.74 | 0.154 | 0.068 | 2.27 |
Table 3. Raw second moments of the forcing conditioned on orientation, from tensor-forced runs at (, , the first discarded; Appendix F), resolved into the meridional and azimuthal tangent directions and binned in . Standard errors over 16 groups of particles are below 0.002 except in the least populated bins, where they reach 0.006. The fixed-orientation value assumed by Eq. (18) is for the column and 0.133 for the plate.
Stationarity fixes the conditional mean. Since , a stationary marginal with no probability current requires
which the simulations reproduce to 0.2% when binned in signed and the right-hand side is averaged over the samples in each bin: for the column at the measured values in bins centred at to are 0.093, 0.270, 0.416, 0.493 and 0.406, and the bin averages of Eq. (E 1) agree to the precision shown, whereas its values at the bin centres, 0.100, 0.287, 0.434, 0.500 and 0.393, differ by up to 7% through the curvature within the bins. Binning in cancels it, since the mean is odd in . Equation (E 1) is neither a torque of turbulent origin nor a correction the closure omits (§ 8.1), and its agreement with the simulations checks their stationarity rather than the closure. Because the mean is non-zero, the entries of Table 3 are raw second moments and not centred variances; at the mean contributes about three quarters of for the column and two thirds for the plate.
These measurements show how the sampled forcing departs from the fixed-orientation forcing, but they do not separate the drift and memory corrections to the generator: , with the integral time of , is an ordinary correlation integral, whereas is response weighted, so the coloured-noise factor need not absorb them.
Appendix F
Numerical methods
The Langevin simulations integrate Eq. (25) ( prolate, oblate) with two forcings. The projected-vector forcing is , with an isotropic Ornstein–Uhlenbeck vector of per-component variance and correlation time ; it has the covariance (18) at a fixed orientation but not the dynamics of the tensor, which enters through the orientation-dependent map (14). The tensor forcing evolves and themselves. At each step a Gaussian matrix with independent standard normal entries is split into its traceless symmetric and antisymmetric parts,
with an independent draw, and and are advanced by the exact update
Since for these draws, and in Kolmogorov units reproduce and from Eqs. (16)–(17). The realised statistics reproduce , , and both traces to better than one per cent, with zero to machine precision. Halving the time step to changes by less than 1%, within about two standard errors, over –2.2, and five independent runs with each forcing at , with and the first discarded, over (projected vector) and (tensor), scatter by 0.4% and 0.5% about their means, so the discrepancies reported in § 4.2 are not numerical. The densities of Figure 2(b,Figure 2c) come from tensor runs with particles over , the first discarded, binned in 40 equal intervals of , each bin value being plotted where the closure density equals its own average over the bin. The vector is updated exactly over each step,
with independent standard normal vectors, and is advanced by a Heun predictor–corrector with renormalisation. At finite the forcing is continuous in time, though not differentiable, so the orientation equation is a random ordinary differential equation and no Itô correction arises.
Figure 2(a) shows the projected-vector runs of the next paragraph, at with , and figure 4(b) the slip from the same runs. For the amplitude-modulated runs of § 7.2: the column with and , forced by , with the unmodulated vector above and an independent Ornstein–Uhlenbeck process for each particle, of mean , variance and correlation time , updated exactly; , particles for 100, 100 and for , 1 and 4, and over for , discarding the first 20, 20 and and with two to four independent runs per case. The frozen-amplitude value is a 16-point Gauss–Hermite quadrature in over runs at fixed amplitude (, , the first discarded), and the reference is , with the closure concentration (4.6) at uniform dissipation. The fits to the data of Anand et al. [1] minimise the unweighted rms residual in over the digitised points, every density being normalised so and the mixture (7.3) evaluated by 80-point Gauss–Hermite quadrature in ; the four points of figure 7(b) are the centroids of the symbols in the embedded raster of their Fig. 2(b), calibrated against the frame of the inset. The moments of figure 7(c,d) come from the main panels of the same raster: each logarithmic axis is calibrated on its decade ticks, the filled triangles of the runs are segmented by colour, and each is placed at the centre of its bounding box, where one pixel is 2.1% in and 3.9% in ; two markers near overlap and are not used. Each marker is matched to a run of their supplementary table by the nearest tabulated , and is read from the plotted abscissa. Placing the markers at their pixel centroids instead, or taking from the rounded values of the table, keeps the single density within 10% and the mixture within 15% of the measured from upwards, gives factors of 2.5–3.1 and excesses of 9–54% in from upwards, and a weak-torque departure ratio of 1.3–2.2. For the data of Pumir et al. [40] we use the vertices of the curves of their Fig. 2, read from the vector paths of the published article and converted with its axis frame; the vertices sit at the centres of histogram bins of width 0.01 and are quantised in steps of in density, so points below are not used, and each fit minimises the unweighted squared deviation of from the vertices above a fiftieth of the peak. Table 2 uses the parameters of their Table 1 and II, , and , with buoyancy neglected as in their response time, and the drag estimate of § 9 uses their equations (C3) and (C4) with their drag correction .
The second moments quoted in § 4.2 come from repeat runs with both forcings, , and total time , the first discarded, with standard errors of 0.1–0.4%. For the slip maximum of § 6: tensor forcing at , from to 0.2 (column) or to 0.4 (plate) in steps of 0.1, all runs of a shape driven by the same random sequence; in 16 groups, the same timing, at and . The maximum of is located from a cubic spline through added to the closure curve (D 4), and checked against a cubic fit of , a local quadratic, and the crossing of by the finite-difference slope of ; the sampling error is one standard deviation over 400 bootstrap resamplings of the groups, the same at every . Independent runs on the coarser grid –2.2, at and , agree within their larger sampling error. For the azimuthal test of § 4.2: the column under the projected-vector forcing at and 6 (, , the first discarded, eight seeds for each case, the two cases sharing their random numbers seed by seed), with and without the azimuthal component of : is 0.956 and 0.972 of the linear response ((19)) with it and 1.027 and 1.008 without, each with a standard error of 0.0015, and the mean of along the horizontal direction of , 0.118 with it and zero without, accounts as an added restoring rate for over 90% of the difference. The forcing statistics of § 8.1 come from projected-vector runs of the column with the same timing at and 0.079, two seeds each.
Moments of the Boltzmann density are evaluated as in Appendix B, and the lateral-slip maxima of Eq. (33) were located by bounded scalar minimisation to .
The digitisation of the published data and the simulations of this appendix were carried out between August and November 2026 with Claude Opus 5 to 5.5 (Anthropic), which wrote and ran the extraction scripts and the simulation drivers; their availability is given in the data availability statement.
References
- [1]Anand, P., Ray, S. S. & Subramanian, G. 2020 Orientation dynamics of sedimenting anisotropic particles in turbulence. Phys. Rev. Lett. 125, 034501.DOI
- [2]Beck, C. & Cohen, E. G. D. 2003 Superstatistics. Physica A 322, 267.
- [3]Bhowmick, T., Seesing, J., Gustavsson, K., Guettler, J., Wang, Y., Pumir, A., Mehlig, B. & Bagheri, G. 2024 Inertia induces strong orientation fluctuations of nonspherical atmospheric particles. Phys. Rev. Lett. 132, 034101.DOI
- [4]Brenner, H. & Condiff, D. W. 1974 Transport mechanics in systems of orientable particles. IV. Convective transport. J. Colloid Interface Sci. 47, 199.DOI
- [5]Bréon, F.-M. & Dubrulle, B. 2004 Horizontally oriented plates in clouds. J. Atmos. Sci. 61, 2888.DOI
- [6]Cho, H.-R., Iribarne, J. V. & Richards, W. G. 1981 On the orientation of ice crystals in a cumulonimbus cloud. J. Atmos. Sci. 38, 1111.doi.org/10.1175/1520-0469(1981)038<1111:otooic>2.0.co;2
- [7]Csanady, G. T. 1963 Turbulent diffusion of heavy particles in the atmosphere. J. Atmos. Sci. 20, 201.doi.org/10.1175/1520-0469(1963)020<0201:tdohpi>2.0.co;2
- [8]Dabade, V., Marath, N. K. & Subramanian, G. 2015 Effects of inertia and viscoelasticity on sedimenting anisotropic particles. J. Fluid Mech. 778, 133.DOI
- [9]Daly, S. F. 1984 Frazil ice dynamics. CRREL Monograph 84-1. U.S. Army Cold Regions Research and Engineering Laboratory, Hanover, NH.
- [10]Girimaji, S. S. & Pope, S. B. 1990 A diffusion model for velocity gradients in turbulence. Phys. Fluids A 2, 242.DOI
- [11]Gustavsson, K., Jucha, J., Naso, A., Lévêque, E., Pumir, A. & Mehlig, B. 2017 Statistical model for the orientation of nonspherical particles settling in turbulence. Phys. Rev. Lett. 119, 254501.DOI
- [12]Gustavsson, K., Sheikh, M. Z., Lopez, D., Naso, A., Pumir, A. & Mehlig, B. 2019 Effect of fluid inertia on the orientation of a small prolate spheroid settling in turbulence. New J. Phys. 21, 083008.DOI
- [13]Gustavsson, K., Sheikh, M. Z., Naso, A., Pumir, A. & Mehlig, B. 2021 Effect of particle inertia on the alignment of small ice crystals in turbulent clouds. J. Atmos. Sci. 78, 2573.
- [14]Hänggi, P. & Jung, P. 1995 Colored noise in dynamical systems. Adv. Chem. Phys. 89, 239.
- [15]Happel, J. & Brenner, H. 1983 Low Reynolds Number Hydrodynamics. The Hague: Martinus Nijhoff.DOI
- [16]Jeffery, G. B. 1922 The motion of ellipsoidal particles immersed in a viscous fluid. Proc. R. Soc. Lond. A 102, 161.DOI
- [17]Jiang, F., Zhao, L., Andersson, H. I., Gustavsson, K., Pumir, A. & Mehlig, B. 2021 Inertial torque on a small spheroid in a stationary uniform flow. Phys. Rev. Fluids 6, 024302.DOI
- [18]Joshi, A., Roy, A., Sharma, A. & Koch, D. L. 2026 An inertial slender-body theory. arXiv:2607.02993.arxiv.org/abs/2607.02993
- [19]Jucha, J., Naso, A., Lévêque, E. & Pumir, A. 2018 Settling and collision between small ice crystals in turbulent flows. Phys. Rev. Fluids 3, 014604.DOI
- [20]Jung, P. & Hänggi, P. 1987 Dynamical systems: a unified colored-noise approximation. Phys. Rev. A 35, 4464.pubmed.ncbi.nlm.nih.gov/9898048
- [21]Kesten, H. & Papanicolaou, G. C. 1979 A limit theorem for turbulent diffusion. Commun. Math. Phys. 65, 97.
- [22]Khayat, R. E. & Cox, R. G. 1989 Inertia effects on the motion of long slender bodies. J. Fluid Mech. 209, 435.
- [23]Kim, S. & Karrila, S. J. 1991 Microhydrodynamics: Principles and Selected Applications. Boston: Butterworth-Heinemann.
- [24]Klett, J. D. 1995 Orientation model for particles in turbulence. J. Atmos. Sci. 52, 2276.doi.org/10.1175/1520-0469(1995)052<2276:omfpit>2.0.co;2
- [25]Koch, D. L. & Pope, S. B. 2002 Coagulation-induced particle-concentration fluctuations in homogeneous, isotropic turbulence. Phys. Fluids 14, 2447.
- [26]Kolmogorov, A. N. 1962 A reference of previous hypotheses concerning the local structure of turbulence in a viscous incompressible fluid at high Reynolds number. J. Fluid Mech. 13, 82.
- [27]Lopez, D. & Guazzelli, E. 2017 Inertial effects on fibers settling in a vortical flow. Phys. Rev. Fluids 2, 024306.
- [28]Mardia, K. V. & Jupp, P. E. 2000 Directional Statistics. Chichester: Wiley.DOI
- [29]Menon, U., Roy, A., Kramel, S., Voth, G. A. & Koch, D. L. 2017 Theoretical predictions of the orientation distribution of high-aspect-ratio, inertial particles settling in isotropic turbulence. Abstract Q36.00011, 70th Annual Meeting of the APS Division of Fluid Dynamics, Denver, CO.
- [30]Mishra, H. & Roy, A. 2025 Transport and orientation of anisotropic particles settling in surface gravity waves. arXiv:2512.24190.
- [31]Mishra, H. & Roy, A. 2026 Orientation dynamics of a settling spheroid in simple shear flow: bifurcations and stochastic alignment. arXiv:2604.13676.
- [32]Naso, A., Jucha, J., Lévêque, E. & Pumir, A. 2018 Collision rate of ice crystals with water droplets in turbulent flows. J. Fluid Mech. 845, 615.
- [34]Noel, V. & Chepfer, H. 2010 A global view of horizontally oriented crystals in ice clouds from Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO). J. Geophys. Res. Atmos. 115, D00H23.
- [35]Novikov, E. A. 1965 Functionals and the random-force method in turbulence theory. Sov. Phys. JETP 20, 1290.
- [36]Okamoto, H., Sato, K., Borovoi, A., Ishimoto, H., Masuda, K., Konoshonkin, A. & Kustova, N. 2020 Wavelength dependence of ice cloud backscatter properties for space-borne polarization lidar applications. Opt. Express 28, 29178.
- [37]Patra, P., Koch, D. L. & Roy, A. 2026 Collision efficiency of rapidly settling particle pairs in a turbulent flow. arXiv:2608.16528.
- [38]Pope, S. B. 2000 Turbulent Flows. Cambridge: Cambridge University Press.DOI
- [39]Pruppacher, H. R. & Klett, J. D. 1997 Microphysics of Clouds and Precipitation, 2nd edn. Dordrecht: Kluwer.
- [40]Pumir, A., Sheikh, M. Z., Gustavsson, K., Lévêque, E., Mehlig, B. & Naso, A. 2026 Statistical orientation and distribution of columnar ice crystals in turbulent flows. Phys. Rev. Fluids 11, 064608.
- [41]Pumir, A. & Wilkinson, M. 2011 Orientation statistics of small particles in turbulence. New J. Phys. 13, 093030.
- [42]Roy, A., Hamati, R. J., Tierney, L., Koch, D. L. & Voth, G. A. 2019 Inertial torques and a symmetry breaking orientational transition in the sedimentation of slender fibres. J. Fluid Mech. 875, 576.
- [43]Roy, A., Kramel, S., Menon, U., Voth, G. A. & Koch, D. L. 2023 Orientation of finite Reynolds number anisotropic particles settling in turbulence. J. Non-Newtonian Fluid Mech. 318, 105048.
- [44]Sassen, K. & Benson, S. 2001 A midlatitude cirrus cloud climatology from the Facility for Atmospheric Remote Sensing. Part II: Microphysical properties derived from lidar depolarization. J. Atmos. Sci. 58, 2103.
- [45]Sheikh, M. Z., Gustavsson, K., Lévêque, E., Mehlig, B., Pumir, A. & Naso, A. 2022 Colliding ice crystals in turbulent clouds. J. Atmos. Sci. 79, 2205.DOI
- [46]Sheikh, M. Z., Gustavsson, K., Lévêque, E., Mehlig, B., Pumir, A. & Naso, A. 2024 Effect of turbulence on the collision rate between settling ice crystals and droplets. J. Atmos. Sci. 81, 887.
- [47]Sheikh, M. Z., Gustavsson, K., Lopez, D., Lévêque, E., Mehlig, B., Pumir, A. & Naso, A. 2020 Importance of fluid inertia for the orientation of spheroids settling in turbulent flow. J. Fluid Mech. 886, A9.
- [48]Siewert, C., Kunnen, R. P. J., Meinke, M. & Schröder, W. 2014 Orientation statistics and settling velocity of ellipsoids in decaying turbulence. Atmos. Res. 142, 45.
- [49]Sreenivasan, K. R. & Antonia, R. A. 1997 The phenomenology of small-scale turbulence. Annu. Rev. Fluid Mech. 29, 435.DOI
- [50]Watson, G. S. 1965 Equatorial distributions on a sphere. Biometrika 52, 193.DOI
- [51]Yang, P., Baum, B. A., Heymsfield, A. J., Hu, Y. X., Huang, H.-L., Tsay, S.-C. & Ackerman, S. 2003 Single-scattering properties of droxtals. J. Quant. Spectrosc. Radiat. Transfer 79–80, 1159.DOI
- [52]Yeung, P. K. & Pope, S. B. 1989 Lagrangian statistics from direct numerical simulations of isotropic turbulence. J. Fluid Mech. 207, 531.DOI







