1. Introduction
Particle transport and the resulting concentration in wall-bounded turbulent flows are of central importance to a wide range of industrial and environmental processes (Guha Reference Guha2008; Brandt & Coletti Reference Brandt and Coletti2022; Marchioli et al. Reference Marchioli2025). The majority of fundamental studies have focused on canonical planar configurations, such as plane channels and flat-plate boundary layers. In these geometries, inertial particles are known to drift towards the walls and preferentially concentrate within low-speed streaks (Marchioli & Soldati Reference Marchioli and Soldati2002; Sardina et al. Reference Sardina, Schlatter, Brandt, Picano and Casciola2012). The near-wall particle accumulation is primarily dominated by turbophoresis (Caporaloni et al. Reference Caporaloni, Tampieri, Trombetti and Vittori1975; Reeks Reference Reeks1983), which denotes the tendency of particles to migrate from regions of high to low turbulence intensity.
Many practical flows occur in conduits with wall curvature, which introduces inertial forces that can fundamentally alter particle transport. For example, particles in straight pipe flows show a considerably stronger near-wall concentration compared to plane channels (Picano, Sardina & Casciola Reference Picano, Sardina and Casciola2009). This is because for the concave wall in turbulent pipe flows, the transverse curvature works in accordance with turbophoresis, as the ‘centrifugal’ force induced by azimuthal velocity fluctuations acts radially towards the outer wall. Consequently, both planar and concave geometries exhibit a consistent trend of wallward transport.
A fundamental gap in this understanding, however, lies in the behaviour of particles near a convex wall, where the effects of turbophoresis and the centrifugal force are expected to compete. The canonical configuration to isolate this regime is the concentric annular duct. Unlike the pipe, the annulus possesses a convex inner wall where the physics presents a unique contradiction: while turbophoresis drives particles inwards, the inertial centrifugal force driven by azimuthal turbulent fluctuations acts away from the inner wall. Although the single-phase flow in annuli has been well documented (Chung, Rhee & Sung Reference Chung, Rhee and Sung2002; Bagheri, Wang & Yang Reference Bagheri, Wang and Yang2020; Orlandi, Soldati & Pirozzoli Reference Orlandi, Soldati and Pirozzoli2025), confirming the asymmetry of turbulence statistics, the resulting transport of dispersed phases remains scarcely explored. Recent work by Jiang et al. (Reference Jiang, Lu, Wang, Meng, Shen and Chong2025) investigated particles in Taylor–Couette flow, revealing particle depletion at the inner wall. However, that system is dominated by imposed mean rotation and Taylor vortices. To date, the pure effect of wall curvature on particle transport, isolated from mean rotation, remains unquantified. It is unknown whether the convex curvature of the inner wall is sufficient to overcome the robust mechanism of turbophoretic deposition.
Another key topic regarding particle transport is the transient evolution of concentration, which has received far less quantitative scrutiny. Particle concentrations are known to require extremely long times to reach a statistical steady state (Rouson & Eaton Reference Rouson and Eaton2001; Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008; Picano et al. Reference Picano, Sardina and Casciola2009; Sardina et al. Reference Sardina, Schlatter, Brandt, Picano and Casciola2012), often exceeding the duration of standard direct numerical simulations (DNS). While theoretical models for steady-state particle flux exist (Guha Reference Guha2008) and recent DNS-informed models have been proposed (Zahtila et al. Reference Zahtila, Chan, Ooi and Philip2023), the characteristic time scales governing the non-stationary evolution remain poorly quantified.
In this paper, we present the first DNS of particle-laden turbulent flows within concentric annular ducts. We report phenomena absent in channels and pipes, including particle depletion and a non-monotonic evolution of particle concentration at the inner wall, which is driven by the complex interplay of transport mechanisms such as turbophoresis and curvature-induced centrifugal effects, among others. We also move beyond phenomenological modelling by deriving a particle concentration transport equation. Using a Sturm–Liouville modal analysis, we identify the multi-time-scale asymmetric transport modes that govern this process, providing a reduced-order physical model that captures the transient evolution of the concentration field.
2. Numerical set-up
The numerical configuration is illustrated in figure 1. The geometry of the concentric annular duct is described in a cylindrical coordinate system
$(r,\theta ,z)$
, where
$r$
,
$\theta$
and
$z$
denote the radial, azimuthal and axial directions, respectively. As shown by the cross-section in figure 1, the inner and outer radii of the annulus are denoted by
$R_i$
and
$R_o$
, respectively. Here,
$\delta$
represents the half-height of the channel, and
$R=(R_i+R_o)/2$
denotes the radial distance from the centre
$O$
to the centreline. Accordingly, the computational domain extends over
$L_r = 2\delta$
in the
$r$
direction, and
$L_\theta = 2\pi R$
in the
$\theta$
direction. It is noted that the axial domain length
$L_z = 12\pi \delta$
is chosen to be sufficiently long to accurately capture near-wall streaks (Bagheri et al. Reference Bagheri, Wang and Yang2020), which have been shown to play an important role in particle distribution (Sardina et al. Reference Sardina, Schlatter, Brandt, Picano and Casciola2012). Based on the mapping function
$r = R + \delta \tanh (2\zeta )/\tanh (2)$
, with
$\zeta$
uniformly discretised over the interval
$[-1,1]$
, the grid in the
$r$
direction is refined towards the walls, where no-slip and isothermal boundary conditions are imposed. Uniform grid spacing is employed in the periodic
$\theta$
and
$z$
directions. To isolate curvature effects, two annular configurations with radius ratios
$\eta = R_i/R_o = 0.5$
and
$0.1$
are considered, representing mild- and strong-curvature cases, respectively (Ishida, Duguet & Tsukahara Reference Ishida, Duguet and Tsukahara2016). A plane channel flow is also included as a zero-curvature reference case. In this situation,
$L_\theta$
denotes the spanwise domain size, and the axial length is shortened to
$L_z = 4\pi \delta$
to reduce the computational cost (Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008). The parameters of all simulations are summarised in table 1, where case 1 is post-processed using symmetry with respect to the centreline.
Parameters of the different cases. Here,
$ {\textit{Re}}_\tau$
is the friction Reynolds number, and the
$N$
and
$\Delta ^+$
terms denote the number of grid nodes and the grid resolution in wall units in each direction, respectively. ‘Inner’ and ‘outer’ denote quantities normalised by the viscous scales at the inner and outer walls, respectively.

Table 1. Long description
The table presents parameters for different cases in a numerical configuration of concentric annular duct flow. It includes friction Reynolds number, grid nodes, grid resolution in wall units, and normalized quantities for inner and outer walls. The table has seven columns: Case, Reτ inner and outer, η, Lr, Lθ, Lz, Nr, Nθ, Nz, Δr+ inner and outer, Δθ+ inner and outer, Δz+ inner and outer. The rows provide specific values for each case, detailing the configuration and resolution. Notable trends include varying grid resolutions and Reynolds numbers across different cases, with specific focus on inner and outer wall normalization.
Numerical configuration of the annular duct: (a) the cross-section of the duct; (b) an overview of the computational domain. The contours of axial velocity are shown within the volume domain, and the density contours are displayed on the inner and outer walls.

Figure 1. Long description
The diagram consists of two parts: (a) a cross-section of an annular duct and (b) an overview of the computational domain. In the cross-section, the duct is depicted with inner and outer radii labeled as R_i and R_o, respectively. The radial distance r and angular coordinate theta are also indicated. The overview shows the computational domain with contours of axial velocity within the volume and density contours on the inner and outer walls. Particles are represented as small dots distributed within the domain, illustrating their transport and concentration patterns. The diagram highlights the interaction between particles and turbulent flow, emphasizing the accumulation of particles in specific regions.
The turbulent flows are considered to operate at Reynolds number
$ {\textit{Re}}_{{f}} = (\rho _b U_b \delta )/\mu _\infty = 2450$
and Mach number
$Ma_{{f}} = U_b/a_\infty = 0.25$
, where
$\rho _b$
and
$U_b$
denote the mean bulk density and velocity, respectively. Here,
$\mu _\infty$
and
$a_\infty$
denote the reference viscosity and acoustic velocity. The flow field is obtained by solving the Navier–Stokes equations using the finite-difference solver HiPSTAR (Sandberg et al. Reference Sandberg, Michelassi, Pichler, Chen and Johnstone2015), which ensures fourth-order accuracy in both time and space. Temperature and density fluctuations remain below 0.15 % throughout the domain, as detailed in Appendix A, confirming that compressibility effects are negligible at
$Ma_{{f}}=0.25$
.
To validate the present carrier flow, figure 2 compares the near-wall statistics obtained from the present simulations with those reported in previous incompressible studies conducted under comparable conditions (Kim, Moin & Moser Reference Kim, Moin and Moser1987; Chung et al. Reference Chung, Rhee and Sung2002; Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008). Specifically, the comparison includes the mean axial fluid velocity
$u^{+}_z$
and the root mean square (rms) fluid velocity fluctuations in the axial, radial and azimuthal directions, denoted by
$u^{\prime +}_{z,{\textit{rms}}}$
,
$u^{\prime +}_{r,{\textit{rms}}}$
and
$u^{\prime +}_{\theta ,{\textit{rms}}}$
, respectively. Here, the superscript
$+$
denotes normalisation using the viscous velocity scales (
$u_{\tau ,i}$
and
$u_{\tau ,o}$
) and viscous length scales (
$\delta _{\nu ,i}$
and
$\delta _{\nu ,o}$
) at the inner and outer walls, respectively. In particular,
$r^+$
denotes the wall-normal distance in wall units, defined as
$r^+=(r-R_i)/\delta _{\nu ,i}$
in the inner-wall region, and
$r^+=(R_o-r)/\delta _{\nu ,o}$
in the outer-wall region. The close agreement demonstrates that the present single-phase DNS provides a reliable turbulent carrier flow.
Near-wall fluid velocity statistics:
$(a)$
mean axial fluid velocity
$u_z^+$
,
$(b)$
axial rms fluid velocity fluctuation
$u^{\prime +}_{z,{\textit{rms}}}$
,
$(c)$
radial rms fluid velocity fluctuation
$u^{\prime +}_{r,{\textit{rms}}}$
, and
$(d)$
azimuthal rms fluid velocity fluctuation
$u^{\prime +}_{\theta ,{\textit{rms}}}$
. The superscript
$+$
denotes viscous scaling, and
$r^+$
denotes the wall-normal distance in wall units. The markers denote the data from Kim et al. (Reference Kim, Moin and Moser1987) (plane channel,
$ {\textit{Re}}_{{f}}=3300$
), the University of Udine (UUD) dataset of Marchioli et al. (Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008) (plane channel,
$ {\textit{Re}}_{{f}}=2100$
) and the data from Chung et al. (Reference Chung, Rhee and Sung2002) (concentric annuli,
$ {\textit{Re}}_{{f}}=2225$
).

Figure 2. Long description
The image contains four graphs showing near-wall fluid velocity statistics. The first graph (a) displays the mean axial fluid velocity with various datasets and cases plotted against the wall-normal distance in wall units. The second graph (b) shows the axial rms fluid velocity fluctuation, the third graph (c) illustrates the radial rms fluid velocity fluctuation, and the fourth graph (d) presents the azimuthal rms fluid velocity fluctuation. Each graph includes data from Kim et al. (1987), the UUD dataset of Marchioli et al. (2008), and Chung et al. (2002). The markers denote different datasets and cases, with lines representing theoretical laws and specific cases. The graphs collectively analyze the behavior of fluid velocity near walls in different configurations.
The dispersed phase is modelled as rigid spheres with uniform diameter
$d_{{p}}=0.004\delta$
and fixed density
$\rho _{{p}}$
. For cases 1–3, the particle density is set to
$\rho _{{p}}=1200\rho _b$
to quantify the effect of curvature. For the strong-curvature annular configuration (cases 3–6), four types of particles with different response times,
$\tau _{{p}}=\rho _{{p}}d_{{p}}^2/(18 \mu _\infty )$
, are considered to assess the influence of the Stokes number, where
$\tau _{{p}}$
is varied by changing
$\rho _{{p}}$
. The detailed parameters are summarised in table 2, and the resulting nominal viscous Stokes number
$St^+=\tau _{{p}}/\tau _{{f}}$
spans the range 6–81. Here,
$\tau _{{f}}$
denotes the viscous time scale, defined as
$\tau _{{f}}=\delta _{\nu ,i}/u_{\tau ,i}$
at the inner wall, and
$\tau _{{f}}=\delta _{\nu ,o}/u_{\tau ,o}$
at the outer wall. Note that the particle diameter is smaller than both the Kolmogorov and viscous length scales, thereby justifying the point-particle assumption. Owing to the large density ratio, particle dynamics are dominated by drag (Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008), with the drag coefficient modelled using the correlation proposed by Wen & Yu (Reference Wen and Yu1966). Third-order Lagrange polynomials (Ruan & Xiao Reference Ruan and Xiao2024) are employed for spatial interpolation, and the velocity Verlet algorithm (Tartakovsky & Panchenko Reference Tartakovsky and Panchenko2016) is adopted for time integration. Particle–wall interactions are assumed to be perfectly elastic (Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008).
Parameters of the different cases. Here,
$N_{{p}}$
denotes the total particle number,
$\varPhi _{{p}}$
denotes the particle volume fraction, and
$\eta _{{w}}$
denotes the smallest eddy scale near the wall.

Table 2. Long description
The table presents parameters for different cases modeling the dispersed phase as rigid spheres with uniform diameter and fixed density. It includes columns for case number, particle diameter over channel half-height, particle density over fluid density, particle response time over the ratio of channel half-height to friction velocity, particle volume fraction, total particle number, particle diameter over Kolmogorov length scale near the wall for inner and outer regions, particle diameter over viscous length scale for inner and outer regions, and Stokes number for inner and outer regions. The table has six rows and twelve columns, detailing specific values for each case. Notable trends include variations in particle density and response times to assess the influence of the Stokes number.
Near-wall particle velocity statistics:
$(a)$
mean axial particle velocity
$v_z^+$
,
$(b)$
axial rms particle velocity fluctuation
$v^{\prime +}_{z,{\textit{rms}}}$
,
$(c)$
radial rms particle velocity fluctuation
$v^{\prime +}_{r,{\textit{rms}}}$
, and
$(d)$
azimuthal rms particle velocity fluctuation
$v^{\prime +}_{\theta ,{\textit{rms}}}$
. The superscript
$+$
denotes viscous scaling, and
$r^+$
denotes the wall-normal distance in wall units. The markers denote the UUD dataset of Marchioli et al. (Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008) (plane channel,
$ {\textit{Re}}_{{f}}=2100$
).

Figure 3. Long description
The image contains four line graphs that illustrate near-wall particle velocity statistics. The graphs show mean axial particle velocity, axial rms particle velocity fluctuation, radial rms particle velocity fluctuation, and azimuthal rms particle velocity fluctuation. The x-axis represents the wall-normal distance in wall units, while the y-axis varies for each graph. The superscript denotes viscous scaling. The markers denote the UUD dataset of Marchioli et al. (2008) for a plane channel. Each graph uses different colors and line styles to distinguish between various cases and conditions. The data points and trends are clearly marked, showing how particle velocity changes with distance from the wall. The graphs are overlaid to compare different velocity components and their fluctuations. All values are approximated.
All simulations are performed with the one-way coupling strategy. Starting from an initially uniform random particle distribution as shown in figure 1, the particles are initialised to follow the fully turbulent fluid motion at
$t=0$
, where
$t$
denotes the time non-dimensionalised by
$\delta /U_b$
. The simulations are carried out for a total non-dimensional duration
$T/(\delta /U_b)=3000$
. Note that this duration is longer than typical channel flow studies (Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008) to account for the slow transients of concentration evolution, although, as discussed in § 3, case 2 exhibits relaxation time scales that challenge even this extended integration time. To validate the present particle framework (Wang et al. Reference Wang, Meng, Tian and Zhao2026), the statistics of the ‘stationary’ particle velocity (Rouson & Eaton Reference Rouson and Eaton2001) in plane channel flow are compared with the reference data of Marchioli et al. (Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008), as shown in figure 3. The statistics are computed from the data collected over the last 1000 non-dimensional time units, including the mean axial particle velocity
$v^{+}_z$
and the rms particle velocity fluctuations in the axial, radial and azimuthal directions, denoted by
$v^{\prime +}_{z,{\textit{rms}}}$
,
$v^{\prime +}_{r,{\textit{rms}}}$
and
$v^{\prime +}_{\theta ,{\textit{rms}}}$
, respectively. The present plane channel DNS results agree well with the previous data, which supports the reliability of the current simulations.
Importantly, the near-wall particle statistics in the curved configurations exhibit noticeable differences between the two walls, particularly in the more strongly curved cases, with an additional dependence on the Stokes number. This behaviour indicates distinct particle dynamics in the vicinity of the two walls, motivating the subsequent analysis of particle transport. Note that the particle velocity statistics in cases 3 and 4 appear to show some unexpected trends near the inner wall, especially for
$v_z^+$
and
$v^{\prime +}_{z,{\textit{rms}}}$
shown in figures 3(a) and 3(b). These discrepancies may be due to insufficient sampling, since the inner wall in these high-
$St^+$
and strong-curvature cases shows significant particle depletion, as will be discussed later in figure 7(b). In fact, under the present set-up, the instantaneous number of particles within the inner-wall viscous sublayer (
$r^+\lt 5$
) for cases 3 and 4 is only of order
$\mathcal{O}(1{-}10)$
.
3. Results and discussion
3.1. Curvature-induced asymmetric particle transport and its statistical characterisation
To elucidate the radial particle transport within the duct, we employ the Shannon entropy
$\mathcal{S}$
(Picano et al. Reference Picano, Sardina and Casciola2009; Bernardini Reference Bernardini2014) as a statistical descriptor. To obtain the parameter
$\mathcal{S}$
, the computational domain is first divided radially into
$N_r-1$
slabs of equal cross-sectional area (Picano et al. Reference Picano, Sardina and Casciola2009), where
$N_r = 193$
is the number of radial grid nodes. For each instantaneous snapshot, the probability of finding a particle in the
$j$
th slab is defined as
$p_{\!j}=N_{\!j}/N_{{p}}$
, where
$N_{\!j}$
is the number of particles contained in that slab, and
$N_{{p}}$
is the total number of particles. Accordingly, the Shannon entropy is given by
\begin{align} \mathcal{S}=-\sum _{j=1}^{N_r-1}p_{\!j}\log p_{\!j}, \end{align}
whose maximum value
$\mathcal{S}_{\textit{max}}=\log (N_r-1)$
is attained for a uniform particle distribution. As shown in figure 4, the normalised entropy initially decreases before reaching a plateau, suggesting an increase in the spatial inhomogeneity of the particle concentration. The comparison among cases 1–3 indicates that the inhomogeneity is more pronounced in the strong-curvature case, as evidenced by the lower asymptotic entropy value, suggesting enhanced radial particle transport. As
$St^+$
decreases from cases 3 and 4 to cases 5 and 6, the entropy decays more slowly overall, and remains at a relatively higher level, indicating weaker radial particle transport and a more homogeneous radial distribution.
Temporal evolution of the Shannon entropy parameter.

Figure 4. Long description
The line graph displays the temporal evolution of the Shannon entropy parameter for six different cases over time. The x-axis represents time (t) ranging from 0 to 3000, while the y-axis represents the normalized Shannon entropy (S/S_max) ranging from 0 to 1. Each case is represented by a distinct color: black for Case 1, light blue for Case 2, light orange for Case 3, green for Case 4, orange for Case 5, and maroon for Case 6. All values are approximated.
Spatial and temporal evolution of the dimensionless particle concentration
$c(r,t)$
near the outer wall of
$(a)$
case 1,
$(b)$
case 2,
$(c)$
case 3,
$(d)$
case 4,
$(e)$
case 5 and
$(f)$
case 6. Here,
$r^+$
denotes the viscous-scaled radial distance, and contours represent the particle concentration
$c(r,t)$
, with black iso-lines denoting
$c(r,t)=1$
.

Figure 5. Long description
Six heat maps display the spatial and temporal evolution of dimensionless particle concentration near the outer wall for six different cases. The x-axis represents time, while the y-axis represents the viscous-scaled radial distance. Contours indicate particle concentration levels, with black iso-lines denoting specific concentration values. Each subplot corresponds to a different case, illustrating how particle concentration changes over time and distance. The color gradient ranges from red to blue, representing varying concentration levels.
To provide insight into the directionality of the transport process, we further investigate the evolution of the dimensionless concentration
$c(r,t)=c_{{slice}}/c_0$
as shown in figures 5 and 6, where
$c_{{slice}}$
and
$c_0$
denote the mean particle number densities in each radial slice and in the entire domain, respectively. Compared to the zero-curvature reference case, the concentration distributions in annular channels exhibit noticeable differences near both walls, indicating asymmetric radial transport of particles. Similar to the wallward particle accumulation that has been reported in plane channel flows (Sardina et al. Reference Sardina, Schlatter, Brandt, Picano and Casciola2012), particles show continuous accumulation near the outer wall of concentric annular ducts. This behaviour is evidenced by the monotonic growth of the dimensionless near-wall concentration
$c_{{w}}$
shown in figure 7(a), where
$c_{{w}}$
is defined as the mean concentration within the viscous sublayer (
$r^+\lt 5$
).
Spatial and temporal evolution of the dimensionless particle concentration
$c(r,t)$
near the inner wall of
$(a)$
case 1,
$(b)$
case 2,
$(c)$
case 3,
$(d)$
case 4,
$(e)$
case 5 and
$(f)$
case 6. Here,
$r^+$
denotes the viscous-scaled radial distance, and contours represent the particle concentration
$c(r,t)$
, with black iso-lines denoting
$c(r,t)=1$
.

Figure 6. Long description
Six heat maps display the spatial and temporal evolution of dimensionless particle concentration near the inner wall for six different cases. Each heat map represents a different case, labeled as case 1 through case 6. The x-axis denotes time (t), and the y-axis represents the viscous-scaled radial distance (r+). Contours within each heat map indicate the particle concentration (c), with black iso-lines denoting a specific concentration level. The color gradient ranges from red to blue, representing different concentration levels, with red indicating higher concentrations and blue indicating lower concentrations. The heat maps illustrate how particle concentration evolves over time and space for each case, highlighting the competition between turbophoresis and the centrifugal force near the convex inner wall.
Dimensionless mean particle concentration within the viscous sublayer (
$r^+\lt 5$
) near
$(a)$
the outer wall and
$(b)$
the inner wall.

Figure 7. Long description
The image contains two line graphs labeled (a) and (b), each showing dimensionless mean particle concentration within the viscous sublayer near the outer wall and the inner wall. Graph (a) displays six cases with different colored lines, all converging to a similar value over time. Graph (b) also shows six cases with varying trends, where concentration decreases over time as a parameter labeled St+ increases. Both graphs have logarithmic scales on the y-axis representing concentration and linear scales on the x-axis representing time. The lines in both graphs represent different cases, with graph (a) showing cases that stabilize and graph (b) showing cases that decrease in concentration.
In contrast,
$c_{{w}}$
near the inner wall exhibits non-monotonic temporal behaviour, characterised by an initial increase followed by a subsequent decrease before reaching eventual saturation, as presented in figure 7(b). Note that
$c_{{w}}$
in case 2 shows no obvious sign of saturation, indicating that a statistical steady state has not yet been reached, even though the entropy shown in figure 4 eventually becomes nearly flat. The effects of this slow convergence issue will be further discussed in § 3.2. Importantly, the transient overshoot observed over a range of Stokes numbers suggests a complex interplay of transport mechanisms intrinsic to curved geometries, thereby challenging standard equilibrium models that predict monotonic accumulation. The final near-inner-wall concentration shows relative accumulation
$(c_{{w}}\gt 1)$
in case 2 with mild curvature, whereas in cases 3–6, corresponding to strong curvature, a depletion of particles is observed
$(c_{{w}}\lt 1)$
. As shown in figure 7(b), the steady-state
$c_{{w}}$
value in case 3 is slightly higher than that in case 4, but remains lower than values in cases 5 and 6. The trend indicates that the inner-wall particle depletion is weakened as
$St^+$
decreases. The transient evolution of
$c_{{w}}$
is also affected by the Stokes number, as lower-
$St^+$
cases generally exhibit both a less pronounced overshoot and a slower subsequent decrease in
$c_{{w}}$
. Nevertheless, the transient overshoot persists in all inertial-particle cases considered here. Note that the overshoot is expected to vanish in the tracer-particle limit
$\tau _{{p}}\rightarrow 0$
, for which the near-wall concentration remains constant at
$c_{{w}}=1$
in the incompressible limit.
To quantitatively characterise the asymmetric radial transport of particles, with emphasis on the non-monotonic temporal evolution of the near-inner-wall concentration, we establish a statistical formulation for the concentration field. Within the single-particle position–velocity probability density function formalism (Johnson, Bassenne & Moin Reference Johnson, Bassenne and Moin2020; Jiang et al. Reference Jiang, Lu, Wang, Meng, Shen and Chong2025), the conservation equations for the particle-phase number density and radial momentum are given by
where
$J=\langle v_r|r\rangle\, c$
denotes the radial particle flux, and the angle brackets
$\langle {\cdot }|r\rangle$
represent ensemble averages conditioned on the particle being located at the radial position
$r$
. Here, the coefficient
$D_1$
characterises the advection velocity, and
$D_2$
corresponds to the turbulent diffusivity (Zahtila et al. Reference Zahtila, Chan, Ooi and Philip2023), which are expressed as
\begin{align} \frac {D_1(r,t)}{\tau _{{p}}}&= \underbrace {\frac {1}{\tau _{{p}}}\langle u_r|r\rangle }_{\text{biased sampling}} -\underbrace {\frac {1}{r}\frac {\partial }{\partial r}(r\,\langle v_r^2|r\rangle )}_{\text{`turbophoresis'}} +\underbrace {\left \langle \frac {v_\theta ^2}{r}\,|\,r\right \rangle }_{\text{centrifugal force}}, \end{align}
Here,
$u_r$
denotes the radial fluid velocity evaluated at the particle location, while
$v_r$
and
$v_\theta$
represent the radial and azimuthal components of the particle velocity, respectively. It should be emphasised that the coefficient
$D_1$
reflects the combined contributions of biased sampling, curvature-modulated turbophoresis, and additional centrifugal effects relative to plane channels, extending the framework of Zahtila et al. (Reference Zahtila, Chan, Ooi and Philip2023), which focused primarily on turbophoresis.
By combining (3.2) and (3.3), and eliminating the radial particle flux
$J(r,t)$
, a transport equation for the particle concentration
$c(r,t)$
is derived:
Under the assumption of perfectly elastic particle–wall interactions, which implies
$J=0$
at the particle-accessible boundaries, the corresponding no-flux boundary conditions are given by
Non-integrable singularities in the concentration
$c$
are avoided, since
$D_2$
remains strictly positive over the radial domain
$r\in [R_1,R_2]$
. Owing to the initially random particle distribution, the initial conditions are approximated as
\begin{align} \begin{aligned} &c(r,0)\approx 1,\\ &\frac {\partial c(r,t)}{\partial t} =-\frac {1}{r}\frac {\partial }{\partial r}[r\,\langle v_r|r\rangle\, c] \approx -\frac {1}{r}\frac {\partial }{\partial r}(r\,\langle u_r|r\rangle ) \approx 0\quad \text{at }t=0. \end{aligned} \end{align}
Equation (3.6) along with the boundary conditions (3.7) and initial conditions (3.8) are thus used for analysis of the statistical steady state and transient evolution of particles in the following subsections.
3.2. Particle concentration at statistical steady state
For the statistical steady state, imposing the time-independence condition
$\partial /\partial t = 0$
, (3.3) indicates that
$J=0$
can be satisfied through a balance between the advection velocity
$D_1$
and turbulent diffusivity
$D_2$
. Equation (3.6) then reduces to an ordinary differential equation governing the steady-state concentration profile
$c_{\infty }(r)$
, which reads
\begin{align} c_\infty (r) =\mathcal{N}\exp \left [ \frac {1}{\tau _{{p}}}\int _{R_1}^{r}\frac {\langle u_r|\xi \rangle }{\langle v^2_r|\xi \rangle }\mathrm{d}\xi -\int _{R_1}^{r}\frac {\mathrm{d}\ln (\xi \langle v^2_r|\xi \rangle )}{\mathrm{d}\xi }\mathrm{d}\xi +\int _{R_1}^{r}\frac {\left \langle v^2_\theta /\xi |\xi \right \rangle }{\langle v^2_r|\xi \rangle }\mathrm{d}\xi \right ]\!, \end{align}
where
$\mathcal{N}$
is an integration constant, and the three phoresis integrals in the bracket are denoted by
$I_b$
,
$I_t$
and
$I_c$
, corresponding to the contributions of biased sampling, turbophoresis and centrifugal effects, respectively. As shown in figure 8(a), the numerical results of
$c_{\infty }(r)$
(based on the statistics collected over the last 1000 non-dimensional time units) are well reproduced by (3.9) for a majority of cases, demonstrating the applicability of the statistical approach. Note, however, that case 2 does not reach a fully steady state until the end of the simulation, as indicated by the continued presence of active radial particle transport shown in figure 7(b). Accordingly, the
$c_{\infty }(r)$
distribution given by (3.9) (blue circles in figure 8
a) shows non-negligible deviations from the DNS results. To account for the non-zero radial flux, the steady solution
$c_\infty (r)$
in (3.9) is modified by premultiplying it with the unsteady correction term
which follows from (3.3) under the approximation that the term
$\tau _{{p}}\,\partial J(r,t)/\partial t$
is neglected. It is observed that introducing the modification in (3.10) significantly refines the results (see the blue crosses in figure 8
a). As will be demonstrated rigorously in § 3.3, the improvement introduced by (3.10) corresponds to the dominant slowly decaying eigenmode of the transport equation.
$(a)$
Radial profiles of the steady-state concentration and contributions of phoresis integrals.
$(b)$
Biased sampling integral
$I_b$
and a zoomed-in view of its variations near the outer wall.
$(c)$
Turbophoresis integral
$I_t$
and a zoomed-in view of its variations near the outer wall.
$(d)$
Centrifugal integral
$I_c$
and a zoomed-in view of
$I_t+I_c$
, where the black dashed line marks the zero level.

Figure 8. Long description
The image contains four graphs labeled (a), (b), (c), and (d), each depicting different aspects of radial profiles related to steady-state concentration and phoresis integrals. Graph (a) shows the steady-state concentration (c_infinity) as a function of the radial distance (r - R_i) for six different cases, with both DNS (Direct Numerical Simulation) and theoretical steady-state data. The cases are distinguished by different colors and symbols. Graph (b) illustrates the biased sampling integral (I_b) as a function of the radial distance, with an inset zooming in on the variations near the outer wall. Graph (c) presents the turbophoresis integral (I_t) as a function of the radial distance, with an inset showing a zoomed-in view near the outer wall. Graph (d) displays the centrifugal integral (I_c) as a function of the radial distance, with an inset highlighting the variations near the outer wall and a black dashed line marking the zero level. Each graph includes multiple lines representing different cases, and the insets provide detailed views of specific regions of interest. All values are approximated.
The relative contributions of individual mechanisms are quantified by the phoresis integrals plotted in figures 8(b), 8(c) and 8(d). Overall, the biased sampling integral
$I_b$
remains small compared with the turbophoresis integral
$I_t$
and the centrifugal integral
$I_c$
. As the curvature increases from case 1 to case 3, the turbophoretic contribution near the inner wall is significantly reduced, as evidenced by the diminished near-wall variation of
$I_t$
at
$r-R_i\lt 0.1$
in figure 8(c). This is consistent with the decreasing peak radial velocity fluctuation shown in figure 3(c), presumably related to the suppression of turbulent fluctuations by the convex curvature near the inner wall in figure 2(c). In contrast, the variations of
$I_t$
near the outer wall, measured by
$I_t-I_t(R_2)$
, remain largely unchanged across different cases, as shown in the inset of figure 8(c). This behaviour is consistent with the limited influence of curvature on the turbulence statistics in the present concave boundary layer, as evidenced by figure 2. As shown in figure 8(d), comparison between cases 2 and 3 further reveals that the centrifugal effect is more pronounced in the case with stronger curvature, driving particles towards the outer wall. The turbophoresis and centrifugal effects contribute synergistically near the outer wall, while competing near the inner wall. Accordingly, particles are expected to accumulate near the outer wall, while the inner-wall boundary layer shows particle depletion in the strong-curvature case.
We further examine the effects of varying Stokes number by considering cases 3–6. As shown in figure 8(a), cases with smaller Stokes number feature a more uniform distribution accompanied by a relatively higher inner-wall concentration, consistent with figure 7(b). Further quantifying the phoresis integrals, the biased sampling contribution becomes slightly stronger as the Stokes number decreases, as reflected by the increased amplitude of
$I_b$
shown in figure 8(b). The inner-wall turbophoretic contribution is also enhanced, whereas the outer-wall turbophoretic contribution remains largely unchanged, as indicated by figure 8(c). The centrifugal integral
$I_c$
also becomes stronger with decreasing Stokes number. Despite the increase in the individual turbophoretic and centrifugal contributions, the combined effect, measured by
$I_t+I_c$
, becomes weaker in magnitude with decreasing Stokes number, as shown in the inset of figure 8(d). This suggests that the net redistribution driven by turbophoresis and centrifugal effects is weakened at lower Stokes number. As a result, the relative role of biased sampling becomes more important, which is consistent with the more uniform spatial distribution observed for case 6 in figure 8(a).
3.3. Slowly decaying asymmetric particle transport modes
Asymmetric particle transport is analysed quantitatively based on (3.6). Given the much longer convergence time of particle concentration compared to particle velocity statistics (Rouson & Eaton Reference Rouson and Eaton2001; Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008; Picano et al. Reference Picano, Sardina and Casciola2009; Sardina et al. Reference Sardina, Schlatter, Brandt, Picano and Casciola2012), a stationary closure
$D_1(r,t)=D_{1,\infty }(r)$
and
$D_2(r,t)=D_{2,\infty }(r)$
is adopted for simplicity. The validity of this time-independent assumption is assessed through the temporal evolution of the short-time-averaged particle-sampled velocity statistics, which are shown in figure 9. As discussed in § 3.2, the near-wall particle concentration approaches its statistical steady state only for
$t\gt 2000$
, with pronounced temporal evolution during the transient stage, as also reflected by its much slower convergence in figures 5 and 6. By contrast, the averaged radial and azimuthal particle velocity fluctuations in the first window,
$t\in [0,75]$
, are already in close agreement with their statistical steady-state values. The particle-sampled radial fluid velocity
$\langle u_r|r\rangle$
requires a relatively longer time to converge, while its amplitude varies only weakly over time. This separation of time scales suggests that
$D_1$
and
$D_2$
can be treated as approximately constant during the slow evolution of particle concentration, thereby supporting the adopted stationary closure.
Temporal evolution of the short-time-averaged particle-sampled velocity quantities associated with
$D_1$
and
$D_2$
: radial fluid velocity
$\langle u_r|r\rangle$
(solid lines), radial rms particle velocity
$\langle v_r^2|r\rangle ^{1/2}$
(dashed lines) and azimuthal rms particle velocity
$\langle v_{\theta }^2|r\rangle ^{1/2}$
(dotted lines) for
$(a)$
case 1,
$(b)$
case 2,
$(c)$
case 3,
$(d)$
case 4,
$(e)$
case 5 and
$(f)$
case 6. The red lines denote the results at statistical steady state. In the plane channel case,
$\langle u_r|r\rangle$
,
$\langle v_r^2|r\rangle ^{1/2}$
and
$\langle v_{\theta }^2|r\rangle ^{1/2}$
reduce to particle-sampled normal fluid velocity, radial rms particle velocity, and spanwise rms particle velocity, respectively.

Figure 9. Long description
Six line graphs display the temporal evolution of short-time-averaged particle-sampled velocity quantities for six different cases. Solid lines represent radial fluid velocity, dashed lines indicate radial rms particle velocity, and dotted lines show azimuthal rms particle velocity. The red lines denote results at statistical steady state. In the plane channel case, these quantities reduce to particle-sampled normal fluid velocity, radial rms particle velocity, and spanwise rms particle velocity, respectively. The x-axis represents the radial distance from the inner wall, while the y-axis shows the velocity values. Each graph corresponds to a different case, illustrating how these velocities evolve over time and space. All values are approximated.
The concentration is then expressed as
$c(r,t)=c_\infty (r)\,\zeta (r,t)$
, where
$\zeta (r,t)$
represents the transient modulation describing the relaxation towards the steady-state profile
$c_\infty (r)$
. Accordingly, the evolution of
$\zeta (r,t)$
is governed by the equation derived from (3.6) and (3.7),
Applying separation of variables,
$\zeta (r,t)=\mathcal{R}(r)\,\mathcal{T}(t)$
, the radial function
$\mathcal{R}(r)$
satisfies the self-adjoint Sturm–Liouville eigenvalue problem
while the temporal function
$\mathcal{T}(t)$
is governed by
Based on the steady-state concentration profile
$c_{\infty }(r)$
evaluated from (3.9), with the coefficients
$D_{1,\infty }(r)$
and
$D_{2,\infty }(r)$
extracted a posteriori from the DNS statistics, a complete set of orthonormal eigenfunctions
$\{\mathcal{R}_n(r)\}$
and the corresponding eigenvalues
$\{\lambda _n\}\geqslant 0$
can be obtained numerically from (3.12). For simplicity, each eigenfunction
$\mathcal{R}_n(r)$
is normalised such that
where
$\delta _{mn}$
denotes the Kronecker symbol. The temporal function admits two exponential branches of the form
$\mathcal{T}_n(t)\sim \exp (s_n^{\pm }t)$
, with decay rates
$s_n^{\pm }=(-1\pm \sqrt {\varDelta _n})/(2\tau _{{p}})$
, where
$\varDelta _n=1-4\tau _{{p}}\lambda _n$
. Therefore, the particle concentration field can be expressed as a modal expansion of the form
$\mathcal{R}_n(r)\,\mathcal{T}_n(t)$
, given by
\begin{align} \frac {c(r,t)}{c_\infty (r)} =\sum _{n=0}^{n_c}\mathcal{R}_n(r)\, (\underbrace {a_n\,\mathrm{e}^{s_n^{+}t}}_{\substack {\text{slowly}\\ \text{decaying}}} + \underbrace {b_n\,\mathrm{e}^{s_n^{-}t}}_{\substack {\text{rapidly}\\ \text{decaying}}}) +\sum _{n=n_c+1}^{\infty }\!\mathcal{R}_n(r)\, \underbrace {c_n\, \mathrm{e}^{-{t}/({2\tau _{{p}}})} \cos \left ( \frac {\sqrt {|\varDelta _n|}}{2\tau _{{p}}}t+\theta _n \right )}_{\text{oscillatory decaying}}, \end{align}
where the coefficients
$a_n$
,
$b_n$
,
$c_n$
and phase
$\theta _n$
are determined by the initial conditions (3.8), and
$n_c$
denotes the cut-off index separating slowly and rapidly decaying modes (
$\varDelta _n \geqslant 0$
) from oscillatory decaying modes (
$\varDelta _n \lt 0$
).
The decoupled modal expansion quantitatively describes the multi-time-scale relaxation of the concentration field towards its steady state. The slowly decaying modes correspond to particle redistribution with larger time scales
$(\gt 2\tau _{{p}})$
, while the rapidly decaying modes correspond to redistribution with smaller time scales
$(\lt 2\tau _{{p}})$
. Both types of modes reflect the relaxation of the particle concentration towards the steady state along a fixed direction. In contrast, the oscillatory decaying modes correspond to oscillatory relaxation components whose direction varies with time, whereas the associated decay time scale remains constant
$(=2\tau _{{p}})$
. In particular, for the zero eigenvalue
$\lambda _0=0$
, the slowly decaying solution degenerates into the zeroth-order mode (
$n=0$
). This mode is referred to as the constant mode, because its amplitude remains unchanged over time, and the corresponding eigenfunction
$\mathcal{R}_0(r)$
is spatially uniform obtained from (3.12). This mode represents the equilibrium particle partition, considering that all other modes decay to zero for
$t\to \infty$
.
Characteristics of the modes:
$(a)$
the time scales normalised by
$\tau _{{p}}$
of the slowly decaying modes (circles), rapidly decaying modes (squares), and oscillatory decaying modes (crosses);
$(b)$
the initial modal energy contribution
$E_{k,n}/E_{k}$
;
$(c)$
the cumulative modal energy contribution
$\sum _{i=0}^{n}E_{k,i}/E_k$
;
$(d)$
the relative amplitudes of slowly decaying modes
$\alpha _n=|a_n|/(|a_n|+|b_n|)$
(circles) and rapidly decaying modes
$\beta _n=|b_n|/(|a_n|+|b_n|)$
(squares).

Figure 10. Long description
The image contains four graphs that analyze the characteristics of different modes in particle transport. The first graph shows the time scales normalized by tau_p for slowly decaying modes, rapidly decaying modes, and oscillatory decaying modes. The second graph illustrates the initial modal energy contribution for different cases. The third graph presents the cumulative modal energy contribution over time. The fourth graph displays the relative amplitudes of slowly decaying modes and rapidly decaying modes. Each graph uses different symbols to represent the cases, with circles for slowly decaying modes, squares for rapidly decaying modes, and crosses for oscillatory decaying modes. The x-axis represents the mode number n, while the y-axes represent different normalized values depending on the graph. The graphs collectively provide a detailed statistical formulation for the concentration field in particle transport.
As evidenced by the normalised decay time scales shown in figure 10(a), the slowest decaying mode evolves on a time scale much larger than
$\tau _{{p}}$
, indicating a very slow transport process. The comparison among cases 1–3 clearly demonstrates that the corresponding time scale decreases with increasing curvature, consistent with the enhanced asymmetric particle transport revealed by comparison of figures 5 and 6. The increase in time scale with decreasing Stokes number quantitatively explains the slower concentration evolution in case 6 compared with cases 3–5, as shown in figure 7(b). The initial modal energy contribution
$E_{k,n}/E_{k}$
is presented in figure 10(b). Here, the initial energy associated with the
$n$
th-order mode
$E_{k,n}$
is defined within the Sturm–Liouville framework as
\begin{align} E_{k,n}&=\left (\int _{R_1}^{R_2}\zeta (r,0)\,[r\,c_\infty (r)]\,\mathcal{R}_n(r)\,\mathrm{d}r\right )^2= \begin{cases} (a_n+b_n)^2,& n\leqslant n_c,\\ c_n^2\cos ^2(\theta _n), &n\geqslant n_c+1, \end{cases} \end{align}
while
$E_{k}=\sum _{n=0}^{\infty }E_{k,n}$
denotes the total modal energy. It is observed that the decaying modes (
$n\leqslant n_c$
) account for most of the initial energy, with the first- and second-order modes
$(n=1\text{ and }n=2)$
making the largest contributions. By contrast, the higher-order oscillatory decaying modes (
$n\geqslant n_c+1$
) contribute only negligibly. The dominant role of the leading modes is further illustrated by the cumulative initial modal energy contribution shown in figure 10(c). The results indicate that the modes of the first three orders account for more than 90 % of the total initial energy. Since the decaying mode at each order comprises both a slowly decaying component and a rapidly decaying component, as given by (3.15), it is useful to further distinguish their respective contributions. To this end, the relative amplitudes of the slowly decaying and rapidly decaying modes, defined as
$\alpha _n=|a_n|/(|a_n|+|b_n|)$
and
$\beta _n=|b_n|/(|a_n|+|b_n|)$
, are shown in figure 10(d). It is observed that the slowly decaying modes possess larger initial amplitudes than the rapidly decaying ones, suggesting that the transient concentration evolution is governed primarily by the slowly decaying modes.
To elucidate the physical role of the dominant first- and second-order modes, the corresponding modal flux
$J_n(r,t)$
is examined, where
$J_n(r,t)$
is defined according to (3.2) as
The flux associated with the slowly decaying modes,
$J^+_n(r,t)$
, satisfies
$J^+_n(r,t) = J^+_n(r,0)\exp (s_n^{+}t)$
. This form corresponds to selecting the branch
$\mathcal{T}_n(t)\sim \exp (s_n^+t)$
in (3.17) and implies exponential decay from the initial flux
$J^+_n(r,0)$
. We therefore examine its radial profile, as shown in figure 11. The first slowly decaying mode represents an outer-wall-directed transport mode arising from the wall curvature, preferentially transporting particles from the inner-wall region towards the outer wall. This mode is specific to curved geometries, since it is activated only in annular ducts, as evidenced by its vanishing amplitude in the plane channel case shown in figure 11(a). This observation is also consistent with the zero energy contribution of the first-order mode,
$E_{k,1}=0$
, in case 1, as shown in figure 10(b). In contrast, the second slowly decaying mode corresponds to a wallward transport mode that drives particles towards both walls simultaneously, analogous to the classical turbophoretic mechanism, as reflected by the sign of its associated initial flux shown in figure 11(b). Except for case 1, the profiles of
$J^+_n(r,0)$
in the other cases exhibit pronounced asymmetry, highlighting the intrinsic effect of geometric curvature.
Radial profile of initial flux
$J^+_n(r,0)$
for
$(a)$
the outer-wall transport mode and
$(b)$
the wallward transport mode.

Figure 11. Long description
Two line graphs compare radial profiles of initial flux for outer-wall and wallward transport modes. The x-axis represents the distance from the inner radius, while the y-axis represents the flux values. Each graph contains six lines representing different cases. In graph (a), the flux values for Case 1 to Case 6 increase with distance, with Case 6 showing the highest values and Case 2 the lowest. In graph (b), the flux values for Case 1 to Case 6 show varying trends, with Case 1 increasing steadily, Case 2 decreasing initially then increasing, and the other cases showing moderate changes. The graphs illustrate the differences in flux profiles for different transport modes and cases.
As also indicated by the quantification of turbophoresis and centrifugal effects shown in figure 8, the outer-wall-directed transport mode is expected to dominate under strong-curvature conditions, as evidenced by the comparison between figures 11(a) and 11(b), which shows that
$|J^+_1(r,0)|$
is larger than
$|J^+_2(r,0)|$
in cases 3–6. This is also supported by the substantially larger energy contribution of the first-order mode shown in figure 10(b). In contrast, for case 2,
$|J^+_2(r,0)|$
is relatively larger near the outer wall, suggesting that the wallward transport mode becomes more important when the curvature is mild. This is likewise reflected in the pronounced energy contribution of the second-order mode for case 2 shown in figure 10(b). The amplitudes of both fluxes near the inner wall also decrease slightly with decreasing Stokes number, as can be seen from the comparison among cases 3–6. This behaviour suggests slower radial particle transport in the near-inner-wall region, in agreement with the trend shown in figure 7(b). The identical signs of
$J_1^+(r,0)$
and
$J_2^+(r,0)$
near the outer wall imply that the two leading modes contribute synergistically in this region, while their opposite signs near the inner wall indicate competition between them.
To demonstrate the dominant role of these two modes in the transient evolution of particle concentration, the concentration field is reconstructed as follows:
\begin{align} c(r,t)\sim c_\infty (r)\,[ \underbrace {\mathcal{R}_0(r)\,a_0}_{\substack {\text{constant}\\ \text{mode}}} + \underbrace {\mathcal{R}_1(r)\,a_1\,\mathrm{e}^{s_1^{+}t}}_{\substack {\text{outer-wall-directed}\\ \text{transport mode}}} + \underbrace {\mathcal{R}_2(r)\,a_2\,\mathrm{e}^{s_2^{+}t}}_{\substack {\text{wallward}\\ \text{transport mode}}} ]. \end{align}
In this reconstruction, only the slowly decaying components of the first three modal orders are retained. As discussed in figure 10, these three leading slowly decaying modes make the dominant contributions to the evolution of the concentration. The resulting near-wall concentration
$c_{{w}}$
is shown in figure 12. Remarkably, for cases 1–3, the first three slowly decaying modes are sufficient to capture the evolution of
$c_{{w}}$
, particularly its non-monotonic behaviour near the inner wall. This indicates that the concentration overshoot is driven by the phase lag between the outer-wall-directed transport mode and the wallward transport mode, which evolve on distinct time scales. During the initial transient, the wallward transport mode dominates and promotes particle accumulation at the inner wall. At later times, the more slowly decaying outer-wall-directed transport mode prevails, leading to subsequent particle depletion. When the concentration dynamics is predominantly governed by the leading slowly decaying modes, the term
$\tau _{{p}}\,\partial J/\partial t\sim \tau _{{p}}s_n^{+}\exp (s_n^{+}t)$
is of smaller order and therefore can be neglected relative to the remaining terms
$J$
,
$D_1c$
and
$D_2\,\partial c/\partial r\sim \exp (s_n^{+}t)$
in (3.3). This further justifies the proposed modification term (3.10).
Reconstructed near-wall concentration
$c_{{w}}$
for
$(a)$
case 1,
$(b)$
case 2,
$(c)$
case 3,
$(d)$
case 4,
$(e)$
case 5 and
$(f)$
case 6.

Figure 12. Long description
The image contains six line graphs labeled (a) through (f), each showing the concentration of particles (c_w) over time (t) in wall-bounded turbulent flows. The x-axis represents time (t) ranging from 0 to 3000 units, and the y-axis represents particle concentration (c_w) on a logarithmic scale from 10^-3 to 10^2. Each graph compares data from direct numerical simulations (DNS) with model predictions. Graph (a) shows a single dataset with DNS represented by circles and the model by a solid line, indicating a close match. Graphs (b) through (f) compare outer and inner regions using different colors: blue for outer and red for inner. The outer DNS data are shown with blue circles, inner DNS data with red circles, outer model predictions with a blue line, and inner model predictions with a red line. The graphs illustrate how particle concentration evolves over time, with the model generally capturing the trends observed in the DNS data. The concentration levels off or decreases depending on the case, highlighting the impact of turbophoresis and other factors on particle accumulation near walls.
Considering cases 3–6, the reduced-order model accurately describes the continuous particle accumulation near the outer wall over a range of Stokes numbers. The trend of the concentration evolution near the inner wall is also captured reasonably well, despite the less satisfactory quantitative agreement with the DNS data at small
$St^+$
. Importantly, the model successfully captures the tendency for radial transport to weaken as
$St^+$
decreases, and provides a good estimate of the eventual concentration. This thus suggests that the reduced-order model captures the dominant physical mechanisms governing the evolution of particle concentration.
The validity of the stationary closure for
$D_1$
and
$D_2$
, which constitutes the only modelling assumption introduced to obtain (3.11), is also re-examined in light of the remaining discrepancies at small
$\tau _{{p}}$
. As shown in figure 9, the particle-sampled radial fluid velocity, associated with biased sampling, converges more slowly than the radial and azimuthal particle velocity fluctuations associated with turbophoretic and centrifugal effects. This suggests that the stationary treatment may be less appropriate for the biased sampling term, which might cause noticeable deviations for smaller
$\tau _{{p}}$
, as the biased sampling contribution becomes more important as presented in figure 8. Accordingly, the reduced-order model is expected to perform better for cases with larger Stokes numbers, as shown in figure 12.
4. Conclusions and outlook
This study reports the first DNS of particle-laden turbulence in concentric annular ducts, revealing how transverse curvature fundamentally alters inertial-particle transport. The results show that curvature introduces a pronounced asymmetry in the radial transport. In particular, in the statistical steady state, the curvature-induced competition between turbophoresis and centrifugal effects leads to severe particle depletion at the inner wall under strong curvature, contrasting with the particle accumulation near the outer wall. The near-inner-wall concentration also exhibits a distinct non-monotonic evolution over a range of Stokes numbers, characterised by an initial overshoot followed by a subsequent decrease towards its statistical steady state.
To interpret these behaviours, a transport equation for the particle concentration was derived within a statistical framework. Through a spectral decomposition of the equation, we demonstrate that the concentration dynamics is governed by a limited number of asymmetric transport modes with distinct decay rates. In particular, the observed overshoot near the inner wall arises from the phase lag between the outer-wall-directed transport mode analogous to centrifugal drift and the wallward transport mode analogous to turbophoresis. A reduced-order model retaining only the dominant slowly decaying modes is able to capture the main dynamics of particle transport, successfully characterising the trend of near-wall concentration across different Stokes numbers. This modal framework not only elucidates the physics of curvature-induced transport, but also offers a basis for modelling non-equilibrium particle dynamics in complex geometries.
In the present study, cases with different Stokes numbers are included to discuss the effects of varying particle response time in the strong-curvature configuration. The range of Stokes number considered, however, is limited. Whether the trend observed with decreasing Stokes number – i.e. a more uniform particle distribution at steady state, a less pronounced overshoot, and a slower evolution of concentration – persists at even smaller Stokes number, e.g.
$St^+\lt 1$
, remains open for future investigation. Analyses of cases at higher Reynolds numbers are also of great interest, as the turbophoretic effect is expected to strengthen.
Acknowledgements
The authors thank Professor Sandberg for providing the licence to use HiPSTAR.
Funding
This work was supported by the National Natural Science Foundation of China (grant nos. 12588201, 12572247 and 12432010).
Declaration of interests
The authors report no conflict of interest.
Appendix A. Effects of compressibility
To assess the effects of compressibility, the mean temperature, mean density and their fluctuations are examined, with detailed profiles presented in figures 13(a) and 13(b), respectively. Both the mean temperature and density remain nearly constant, with deviations confined to within
$1\,\%$
in the near-wall region. The fluctuation levels of temperature and density stay below 0.15 %. Note that the present simulations only consider the drag force (Marchioli et al. Reference Marchioli, Soldati, Kuerten, Arcen, Tanière, Goldensoph, Squires, Cargnelutti and Portela2008) for the particle dynamics, whereas the direct temperature-gradient-induced effect, namely the thermophoretic force (Talbot et al. Reference Talbot, Cheng, Schefer and Willis1980; Guha Reference Guha2008), is excluded. Therefore, the influence of compressibility on particle dynamics is expected to arise predominantly through the slight variation in viscosity, which is considered negligible in the present study. This is further supported by the close agreement between the present flow and particle statistics and published results, as discussed in figures 2 and 3.
Profiles of
$(a)$
mean temperature
$T_{{f}}$
and density
$\rho _{{f}}$
, and
$(b)$
rms of the corresponding fluctuations
$T^\prime _{{f,rms}}$
and
$\rho ^\prime _{{f,rms}}$
.

Figure 13. Long description
The image contains two line graphs side by side. The left graph shows the profiles of mean temperature and density across three different cases. The right graph displays the root mean square (rms) values of the corresponding fluctuations for the same cases. Each graph has a horizontal axis labeled ‘r - Ri’ ranging from 0 to 2. The left graph’s vertical axis is labeled ‘Tf, ρf’ and ranges from 0.98 to 1.03, while the right graph’s vertical axis is labeled ‘Tf_rms, ρf_rms’ and ranges from 0 to 2.5 times 10^-3. The lines are color-coded: black for case 1, blue for case 2, and red for case 3. The solid lines represent temperature profiles, and the dashed lines represent density profiles. The graphs illustrate how these values vary across different cases and positions.
Appendix B. Effects of Reynolds number
To assess the robustness of the reported phenomena at a higher Reynolds number, an additional simulation (case 7) is conducted at
$ {\textit{Re}}_{{f}}=3062$
for the strong-curvature configuration
$\eta = 0.1$
. To reduce the computational cost, the axial domain length is set to
$L_z=16\delta$
, consistent with the set-up adopted by Orlandi et al. (Reference Orlandi, Soldati and Pirozzoli2025). The grid resolution satisfies the requirements reported in previous DNS studies of annular pipe flow (Chung et al. Reference Chung, Rhee and Sung2002; Bagheri & Wang Reference Bagheri and Wang2020), as detailed in table 3. The increase in bulk Reynolds number yields an outer-wall friction Reynolds number
$ {\textit{Re}}_{\tau ,o}\approx 185$
, which is approximately 20 % higher than the value
$ {\textit{Re}}_{\tau ,o}\approx 153$
in case 3. Accordingly, case 7 represents a more turbulent flow and is therefore expected to exhibit stronger turbophoretic effects (Bernardini Reference Bernardini2014). The particle diameter and density values are chosen to be identical to those in case 3 for a direct comparison, and the corresponding parameters are summarised in table 4.
Parameters for case 7. Here,
$ {\textit{Re}}_\tau$
is the friction Reynolds number, and the
$N$
and
$\Delta ^+$
terms denote the number of grid nodes and the grid resolution in wall units in each direction, respectively. ‘Inner’ and ‘outer’ denote quantities normalised by the viscous scales at the inner and outer walls, respectively.

Table 3. Long description
The table presents parameters for case 7, focusing on friction Reynolds numbers and grid details. It includes columns for inner and outer friction Reynolds numbers, grid nodes in the radial, azimuthal, and axial directions, and grid resolution in wall units. The table also provides inner and outer normalized quantities for radial, azimuthal, and axial directions. Key data points include friction Reynolds numbers of 233 and 185, grid nodes ranging from 241 to 320, and grid resolutions such as 0.292 and 0.232 in wall units. Notable trends include variations in grid resolution and the number of grid nodes across different directions.
Parameters for case 7. Here,
$N_{{p}}$
denotes the total particle number,
$\varPhi _{{p}}$
denotes the particle volume fraction, and
$\eta _{{w}}$
denotes the smallest eddy scale near the wall.

Table 4. Long description
The table presents a comparison of various parameters for case 7, focusing on particle dynamics in a fluid flow. It includes columns for particle diameter to boundary layer thickness ratio, particle to fluid density ratio, particle response time to smallest eddy scale ratio, particle volume fraction, total particle number, and particle diameter to wall unit ratio for inner and outer walls. Additionally, it shows the dimensionless particle diameter for inner and outer walls and the Stokes number for inner and outer walls. The table has one row of data and ten columns of parameters.
Spatial and temporal evolution of the dimensionless particle concentration
$c(r,t)$
near the
$(a)$
outer and
$(b)$
inner walls for case 7, and
$(c,d)$
the corresponding dimensionless mean particle concentrations within the viscous sublayer (
$r^+\lt 5$
). Here,
$r^+$
denotes the viscous-scaled radial distance, and contours represent the particle concentration
$c(r,t)$
, with black iso-lines denoting
$c(r,t)=1$
. A zoomed-in view of inner-wall concentration
$c_{{w}}$
is added in figure 14(d).

Figure 14. Long description
The image contains four graphs showing the spatial and temporal evolution of dimensionless particle concentration near the outer and inner walls for case 7. The first two graphs (a and b) are heat maps with the y-axis labeled as r+ and the x-axis labeled as t. The color bar on the right indicates the Log10 of the concentration c, ranging from -3 to 2. The black iso-lines denote a specific concentration level. The third graph (c) is a line graph showing the dimensionless mean particle concentrations within the viscous sublayer for two cases, labeled as Case 3 and Case 7. The y-axis is labeled as cw and the x-axis is labeled as t. The fourth graph (d) is another line graph similar to the third one but includes a zoomed-in view of the inner-wall concentration in the inset. The trends and values in these graphs illustrate the transient evolution of particle concentrations over time.
Figure 14 presents the spatial and temporal evolution of the particle concentration near both walls for case 7. Noticeable differences are observed between the concentration distributions near the two walls in case 7, further confirming the asymmetric nature of the radial transport. Particles also continue to accumulate near the outer wall, as indicated by the monotonic increase of
$c_{{w}}$
in figure 14(c), whereas the inner-wall concentration exhibits the characteristic non-monotonic overshoot identified in § 3.1. Consistent with the stronger turbophoretic effect at the higher Reynolds number, the eventual near-wall concentration levels at both walls are higher than those in case 3, and the overshoot amplitude at the inner wall is also slightly increased, as shown in the inset of figure 14(d).
Characteristics of the modes:
$(a)$
the time scales normalised by
$\tau _{{p}}$
of the slowly decaying modes (circles), rapidly decaying modes (squares), and oscillatory decaying modes (crosses);
$(b)$
the initial modal energy contribution
$E_{k,n}/E_{k}$
;
$(c)$
the cumulative modal energy contribution
$\sum _{i=0}^{n}E_{k,i}/E_k$
;
$(d)$
the relative amplitudes of slowly decaying modes
$\alpha _n=|a_n|/(|a_n|+|b_n|)$
(circles) and rapidly decaying modes
$\beta _n=|b_n|/(|a_n|+|b_n|)$
(squares).

Figure 15. Long description
The image contains four graphs that illustrate different characteristics of modes in a study of particle behavior near a convex wall. The first graph shows the time scales normalized by the slowly decaying modes, rapidly decaying modes, and oscillatory decaying modes. The second graph displays the initial modal energy contribution. The third graph presents the cumulative modal energy contribution. The fourth graph illustrates the relative amplitudes of slowly decaying modes and rapidly decaying modes. Each graph uses different symbols to represent the data points: circles for slowly decaying modes, squares for rapidly decaying modes, and crosses for oscillatory decaying modes. The graphs provide insights into the behavior of particles in an annular duct configuration, highlighting the competition between turbophoresis and centrifugal force.
Radial profile of initial flux
$J^+_n(r,0)$
for
$(a)$
the outer-wall transport mode and
$(b)$
the wallward transport mode.

Figure 16. Long description
The image contains two line graphs labeled (a) and (b), each comparing radial profiles of initial flux for different transport modes in curved conduits. Graph (a) shows the flux for the outer-wall transport mode, while graph (b) depicts the flux for the wallward transport mode. Both graphs plot the flux on the y-axis against the radial distance from the inner wall on the x-axis. The data series are represented by two lines, one for Case 3 in orange and another for Case 7 in blue. In graph (a), both cases show a similar trend, starting from zero, increasing rapidly, and then leveling off as the radial distance increases. In graph (b), the lines for both cases start from zero, dip slightly below zero, and then gradually increase, with Case 7 showing a slightly higher flux than Case 3 at larger radial distances. The graphs illustrate how inertial forces due to wall curvature affect particle transport in conduits, highlighting the differences in flux profiles for outer-wall and wallward transport modes.
The characteristics of the Sturm–Liouville modes are revisited in figure 15. As shown in figure 15(a), the increase in Reynolds number reduces the time scales of the slowly decaying modes, indicating enhanced radial particle transport. Figure 15(b) shows that the leading modes, especially the first- and second-order modes, retain larger initial energy than the higher-order modes. Notably, a higher
$ {\textit{Re}}_\tau$
is associated with a decrease in
$E_{k,1}$
but an increase in
$E_{k,2}$
, suggesting that the contribution of the curvature-induced centrifugal mechanism weakens, while that of turbophoresis strengthens, consistent with the overall physical picture. The modes of the first three orders account for more than 90 % of the total initial energy, as shown in figure 15(c). Most of this contribution is associated with the slowly decaying modes, as indicated by figure 15(d), confirming their dominant role in the transient evolution. Consistent with the conclusion drawn from figure 15(b), a moderate increase in
$ {\textit{Re}}_\tau$
slightly weakens the outer-wall-directed transport mode near the outer wall, while strengthening the wallward transport mode, as reflected by the initial flux profiles
$J^+_n(r,0)$
presented in figure 16.
The proposed three-mode reconstruction in (3.18) is then applied to case 7, and the corresponding results are presented in figure 17. The reduced-order model remains applicable at the higher Reynolds number. The model accurately captures the inner-wall overshoot, agreeing well with the DNS data. This is expected, as the turbophoretic effect is of great importance in this high Reynolds number case (see discussion in § 3.3).
Reconstructed near-wall concentration
$c_{{w}}$
for case 7.

Figure 17. Long description
The line graph presents the reconstructed near-wall concentration for case 7, featuring two sets of data points and model lines. The blue circles represent DNS outer data, while the red circles represent DNS inner data. The blue line indicates the model outer, and the red line indicates the model inner. The x-axis represents time (t) ranging from 0 to 3000, and the y-axis represents the concentration (c_w) on a logarithmic scale from 10^-3 to 10^2. The blue data points and model line show a rapid increase and then level off, while the red data points and model line show an initial increase followed by a gradual decrease. All values are approximated.



Reτ
N
Δ+
(a)
uz+
(b)
uz,rms′+
(c)
ur,rms′+
(d)
uθ,rms′+
+
r+
Ref=3300
Ref=2100
Ref=2225
Np
Φp
ηw
(a)
vz+
(b)
vz,rms′+
(c)
vr,rms′+
(d)
vθ,rms′+
+
r+
Ref=2100
c(r,t)
(a)
(b)
(c)
(d)
(e)
(f)
r+
c(r,t)
c(r,t)=1
c(r,t)
(a)
(b)
(c)
(d)
(e)
(f)
r+
c(r,t)
c(r,t)=1
r+<5
(a)
(b)
(a)
(b)
Ib
(c)
It
(d)
Ic
It+Ic
D1
D2
⟨ur|r⟩
⟨vr2|r⟩1/2
⟨vθ2|r⟩1/2
(a)
(b)
(c)
(d)
(e)
(f)
⟨ur|r⟩
⟨vr2|r⟩1/2
⟨vθ2|r⟩1/2
(a)
τp
(b)
Ek,n/Ek
(c)
∑i=0nEk,i/Ek
(d)
αn=|an|/(|an|+|bn|)
βn=|bn|/(|an|+|bn|)
Jn+(r,0)
(a)
(b)
cw
(a)
(b)
(c)
(d)
(e)
(f)
(a)
Tf
ρf
(b)
Tf,rms′
ρf,rms′
Reτ
N
Δ+
Np
Φp
ηw
c(r,t)
(a)
(b)
(c,d)
r+<5
r+
c(r,t)
c(r,t)=1
cw
(a)
τp
(b)
Ek,n/Ek
(c)
∑i=0nEk,i/Ek
(d)
αn=|an|/(|an|+|bn|)
βn=|bn|/(|an|+|bn|)
Jn+(r,0)
(a)
(b)
cw