Access the full text.
Sign up today, get DeepDyve free for 14 days.
Loading next page...
/lp/arxiv-cornell-university/an-isogeometric-boundary-element-method-for-soft-particles-flowing-in-rIKpvO6ef0
- ISSN
- 0045-7930
- eISSN
- ARCH-3341
- DOI
- 10.1016/j.compfluid.2020.104786
- Publisher site
- See Article on Publisher Site
Abstract
Understanding the ow of deformable particles such as liquid drops, syn- thetic capsules and vesicles, and biological cells con ned in a small channel is essential to a wide range of potential chemical and biomedical engineering applications. Computer simulations of this kind of uid-structure (mem- brane) interaction in low-Reynolds-number ows raise signi cant challenges faced by an intricate interplay between ow stresses, complex particles' in- terfacial mechanical properties, and uidic con nement. Here, we present an isogeometric computational framework by combining the nite-element method (FEM) and boundary-element method (BEM) for an accurate pre- diction of the deformation and motion of a single soft particle transported in micro uidic channels. The proposed numerical framework is constructed consistently with the isogeometric analysis paradigm; Loop's subdivision el- ements are used not only for the representation of geometry but also for the membrane mechanics solver (FEM) and the interfacial uid dynamics solver (BEM). We validate our approach by comparison of the simulation results with highly accurate benchmark solutions to two well-known examples avail- able in the literature, namely a liquid drop with constant surface tension in a circular tube and a capsule with a very thin hyperelastic membrane in a square channel. We show that the numerical method exhibits second-order convergence in both time and space. To further demonstrate the accuracy and long-time numerically stable simulations of the algorithm, we perform hydrodynamic computations of a lipid vesicle with bending sti ness and a Corresponding author Email address: gang.chen@univ-amu.fr (Paul G. Chen) Preprint submitted to Computers and Fluids February 13, 2020 arXiv:2002.04845v1 [physics.flu-dyn] 12 Feb 2020 red blood cell with a composite membrane in capillaries. The present work o ers some possibilities to study the deformation behavior of con ning soft particles, especially the particles' shape transition and dynamics and their rheological signature in channel ows. Keywords: Fluid-structure interaction, Viscous drops, Elastic capsules and vesicles, Red blood cells, Low-Reynolds-number ow, Loop subdivision 1. Introduction Micro uidics refers to devices and methods for the manipulation of uids and immersed objects inside channels with dimensions of tens to hundreds of micrometers [1]. Understanding the ow of soft or deformable micro- particles (e.g., viscous droplets, arti cial capsules and vesicles, and biological cells, etc.) in con ned channels is essential for many multiphase micro u- idic applications like control and application of droplets (or bubbles) [2], and particle/cell sorting [3, 4]. It is also fundamental to biological material trans- port through the microcirculation such as red blood cells (RBCs) [5, 6] and drug delivery [7]. It is therefore essential to study how soft particles deform under the action of ow stresses in a small (con ned) channel and what ef- fect these stresses can have on the transport of the deformable particles and associated processes. Of particular interest are the particle's translational velocity, the overall hydraulic resistance in a given channel containing sus- pended particles, and the disturbed ow eld induced by the presence of suspended particles [8, 9]. The dynamical behavior of these soft objects in an imposed ow of an external uid exhibits distinct characteristics. The mechanical properties of the interfacial (membrane) composition play a key role in this dynamics. The simplest deformable object is a clean, surfactant-free drop, which is charac- terized by its interfacial material property { namely interfacial tension. Other soft entities, however, have increased structural complexity leading to more complex mechanical properties. For instance, a synthetic, liquid- lled capsule can be thought of as a drop enclosed by a solid polymerized membrane that resists shear and area deformation, while a lipid vesicle is a drop enclosed by an inextensible, uidic membrane resisting bending. Shearing and stretching deformations of a vesicle are negligibly small compared to bending ones. The surface-area incompressibility of the lipid membrane is ensured dynamically by a Lagrange eld { namely the membrane tension which is not a material 2 property but is ow dependent, analogous to pressure for three-dimensional incompressible ows. Biological cells like RBCs have a more complex archi- tecture. The RBCs membrane consists of a lipid bilayer (vesicle-like) and an underlying membrane-associated cytoskeleton (capsule-like). As such, vesi- cles and capsules have often served as a model system to mimic RBCs [10]. These soft entities (i.e., drops, capsules, vesicles, and RBCs) are considered in the present work. To facilitate subsequent description, we shall make no di erence between interface and membrane. The soft particles owing in micro uidic channels is essentially a uid- structure interaction (FSI) problem involving highly deformable membranes (interfaces). The small size and the low speed of micro uidics mean that the viscous forces predominate the inertial forces and the linear Stokes equations can, therefore, be applied to describe such low-Reynolds-number ow (Re usually much less than unity) [1]. Despite the linearity of the ow, FSI prob- lems at small scales are highly nonlinear due to the highly nonlinear nature of the deformation of the soft objects and their constitutive models [11, 12]. Moreover, a signi cant diculty arises from the hydrodynamic interaction between the micro uidic wall and the object's surface, especially at condi- tions of high con nement where con nement-induced viscous friction plays a dominant role in the forces exerted on the soft object [13, 14]. It is, therefore, a computationally challenging task to solve this kind of FSI problems with delity. There are a variety of di erent computational strategies to solve FSI problems. Recently developed approaches can be broadly classi ed into three categories (see, e.g., the reviews [11] for capsules, [15] for vesicles, and [12] for RBCs), namely bulk mesh-based methods [16, 17, 18, 19], particle-based methods [20, 21, 22], and boundary-element methods (BEMs). The BEM o ers very high accuracy compared to other methods in the prediction of the dynamics of deformable particles immersed in inertialess Newtonian ows. BEM's theory and its practical implementation are well- described in [23]. Its eciency has been demonstrated in the simulation of drops [24, 25, 26, 27], capsules [28], vesicles [29, 30, 31, 32], and RBCs [33]. The BEM has a notable advantage over domain discretization methods as it leads to a reduction in dimensionality; the ow equations are solved only for the unknown stress and velocity elds at the domain boundaries and at evolv- ing interfaces. This restriction of the discretization only to the boundaries and interfaces greatly improve the computational eciency when studying wall-deformable object interactions. A prominent example of such bene ts 3 is to resolve the drainage uid of thin liquid lm between the object surface and the channel wall at high uidic con nement, where the object occupies a large proportion of the channel cross-section. Regardless of how close the object is to the channel wall, and how their distance changes over time, the lm thickness needs only to be considered when it comes to determining the size of the mesh elements on the membrane and at the channel wall. In their numerical studies on vesicle dynamics con ned in tube ow, Refs. [13, 14] provided an estimate of a typical element size which is about half of this thickness, in the region of the liquid lm. This very worthy property of the BEM eliminates any problem of volume mesh topology related to the move- ment of an interface or error induced by the interpolation of physical elds to it. However, it should be tempered by the fact that the matrix system to be solved in this method is characterized by full and not sparse matrices as in most other methods. In other words, an increase in the degrees of freedom in modeling is seriously more penalizing in terms of computation time. There- fore, any discretization method that increases the accuracy of the numerical representation of surfaces and domain boundaries, while also optimizing the degrees of freedom of this representation, is of great interest to the BEM. In structural design, the most convenient and widespread way to rep- resent a surface is to use a triangular mesh. It enables any surface shape to be represented and allows for the development of adaptation algorithms and local mesh re nement. The simplest triangular elements provide linear interpolation per piece of the surface, as well as all physical elds that use the same approximation (referred to as isoparametric in nite element lan- guage). That is the strategy adopted in [29]. The second strategy makes use of the quadratic triangular elements [34], which allows for an improved accu- racy without increasing the degrees of freedom, but at the cost of a slightly more complex numerical implementation. In both cases, as with all Lagrange elements, the approximation can only be C continuous between elements, thus the spatial derivatives are discontinuous across them. The regularity of representation could be increased to C using Hermitian elements. Their implementation in 2D (or axisymmetric), though more complex, remains af- fordable, the extension to 3D is much more problematic, if not impossible. Moreover, we are not aware of any studies using this option. Being limited to approximations of C is particularly penalizing for other obvious reasons. A direct computation of the membrane bending force re- quires a C representation of the membrane geometry since the bending force contains the fourth-order derivative of the position vector [35]. That is why 4 some local surface reconstruction techniques have been developed to compute the Laplace-Beltrami operator [30, 36]. An interesting alternative approach involves the use of di erential geometry techniques, as demonstrated by [29]. Now, one of the advantages of representation by nite element is that the 4 2 mesh smoothness requirements can be eased from C to H if formulating the interfacial mechanical problem in weak form [37]. In the mathematical eld of functional analysis, the Sobolev space H represents square-integrable functions whose rst- and second-order derivatives are themselves square- integrable. It means that with weak formulation a C nite element ap- proximation is only required to compute the bending force since the second derivatives are then piecewise continuous. The representation of surfaces is a longstanding issue in computer graph- ics. Exploiting the highly accurate interpolation functions developed in this eld, like splines and NURBS (Non-Uniform Rational Basis Splines), repre- sents a recent breakthrough in nite element analysis. Not only, it eases the connection with computer-aided design (CAD), but it opens the way to in- creased regularity of nite element approximation in general. All the physical elds involved in a problem to be solved with nite elements bene t from the same highly accurate representation as to the geometry. This rapidly devel- oping trends in nite element analysis are thus named isogeometric analysis or IGA [38]. In the IGA framework, the surface subdivision is well suited when only the domain boundaries and evolving interfaces have to be considered, as in the BEM. Its use for nite element analysis with meshes made of triangular elements has been made possible thanks to the work of Loop [39], with a rst application to shell analysis by Cirak et al. [37]. IGA-Loop guarantees al- most everywhere a C approximation, except at a few irregular nodes where it's only C . An IGA-BEM model with Loop subdivision for soft particles in the Stokes ows is proposed in [31]. Its eciency has been demonstrated to simulate free-space FSI problems involving drops, capsules, and vesicles in a uni ed numerical modeling framework. Its superior accuracy as com- pared to previous methods to compute geometric properties like curvature was con rmed in [36]. A coupled IGA-BEM and shell approach is seen as a promising way to study the interaction between thin elastic structures and Stokes ows and is attracting growing attention in the soft particle's com- munity [40, 32, 41]. The algorithm developed in [31] is limited to free-space Stokes ows. In [32, 42] three-dimensional computations of a vesicle owing in a circular 5 and rectangular duct (without bending force) have been reported. However, whereas the vesicle surface bene ted from the increased accurate representa- tion of the Loop subdivision, the unknown physical elds did not. Instead, they were represented by a piecewise linear interpolation, like earlier models. In that regard, the IGA spirit cannot be advocated. In this paper, we extend the previous work [31] on soft particles in un- bounded Stokes ows to con ned soft particles transported in micro uidic channels. Both the surface shape of soft objects and the wall boundary of channels are discretized with Loop's subdivision scheme. Most importantly, Loop elements are used not only for the discretization of the interface (mem- brane) but also for the membrane mechanics solver (FEM { the membrane force density) and the uid dynamics solver (BEM { the interfacial veloc- ity and the disturbed wall traction). In this way, the proposed numerical framework is constructed consistently with the IGA paradigm. Compared to the unbounded uid- ow calculations, uidic con nement raises the consider- able diculty in dealing with the hydrodynamic interaction between the wall boundary and the particle surface, as the bounded uid- ow computations require highly re ned grids to accurately capture wall-particle interactions. Such a hydrodynamical computation frequently encounters numerical insta- bilities, particularly in the simulation of vesicle dynamics, which are associ- ated with the membrane's bending rigidity and incompressibility, as reported in, e.g., Ref. [32]. Thanks to the use of Loop elements in a consistent way, the present computational framework overcomes these challenges, enabling accurate and long-time numerically stable simulations. The rest of the paper is organized as follows. In Section 2, we describe the equations governing the ow motion and the membrane mechanics along with the interfacial conditions. In Section 3, we introduce Loop's subdivision scheme and discuss the numerical method implemented in the membrane and uid solvers, whose validation is provided in Section 4. While the current numerical model can deal with channels of arbitrary cross-sections, we focus herein on a single deformable particle in a circular or rectangular channel, for demonstration purposes only. We present additional simulation results to demonstrate the accuracy and stability of the method in Section 5, followed by conclusions and future directions in Section 6. 6 2. Problem statement and formulation As sketched in Fig. 1, we consider a soft particle freely transported in a microchannel of constant cross-section. The particle deforms in response to the ow stresses of bulk ows, as well as to the wall boundary-induced viscous friction, resulting in a change in the membrane forces (e.g., the bending force), which in turn alter the bulk ows. The modeling of such an FSI involves formulating the following three parts: hydrodynamics; membrane mechanics; coupling conditions at the interface. W W n e e y y 2` I O W y e e x z R ⌘ 2` 2` x z Fig. 1. Schematic representation of a freely suspended deformable particle (de ned by the boundary ) transported in a pressure-driven ow through a straight microchannel (either circular of radius R or rectangular of 2` by 2` in cross-section), with I, O and t y z i;e W denoting the inlet, outlet and wall boundaries, respectively. are the viscosities of the uids and n is the unit normal vector pointing into the suspending uid. 2.1. Hydrodynamics The motion of the internal (with superscript i) and external (with super- script e) uids is governed by the incompressible Newtonian Navier-Stokes equations. In a wide variety of micro uidic ows, the Reynolds number is small (less than unity) [1]. For example, in a channel of height 100 μm, a 3 1 ow with water (shear viscosity 10 Pa s) at a typical speed 1 mm s has the Reynolds number Re 0:1. The governing equations can therefore be reduced to the Stokes equations for creeping ow i;e i;e 2 i;e i;e rp + r u = 0; r u = 0; (1) 7 where u and p denote the uid velocity and pressure, respectively. The external velocity eld u satis es the no-slip condition on the boundary walls of the channel u (x) = 0; 8x 2 W (2) and vanishing far- eld ow perturbations, e 1 u (x) = u (x); 8x 2 I[ O: (3) 2.2. Membrane mechanics A membrane is a closed and deformable interface separating the internal and external uids. It is described by its position x(t) at time t. Under stresses, energy variation E is stored in the membrane through the elastic deformations (bending, shearing, and dilation/compression) or dissipated by viscous friction [10]. The surface force density exerted by the membrane f onto surrounding uids is given by the rst variation of its surface energy 1 E f (x) = p ; E = w dS; (4) a x where w is the surface energy per unit area which completely determines the mechanical properties of the membrane, and a is the determinant of the local metric. Below we describe di erent formulations for w (or f ), depending on the type of soft objects under study. For liquid drops, the surface energy per unit area is simply the interfacial tension (which is a material property), i.e., w = , leading to f = r 2 Hn; (5) where r = (I nn)r is the surface gradient operator with I the identity tensor, n is the outward pointing normal vector, and H is the local mean curvature (with the convention that H is positive for a sphere). For vesicles, the lipid membrane is modeled as a two-dimensional incom- pressible uid with bending sti ness. The local surface-area incompressibility of the membrane r u = 0 (6) 8 is enforced via the Lagrange eld (equivalent to the membrane tension, which is not a material property but is ow dependent). This membrane tension is added to the bending energy density w [43], giving the elastic energy density of a vesicle H H w = w + ; w = (2H ) ; (7) s s where is the bending modulus of the lipid bilayer. In principle, a spon- taneous (or reference) curvature and Gaussian curvature (K ) appear in the bending energy. For simplicity, we take the minimum energy reference state as a at sheet. The term with Gaussian curvature does not contribute to variation of the bending energy if the topology remains unchanged, which is the case of our study. Using Eq. (4), one obtains a formal expression of the surface force density [35] m 2 f = 2 H + 4H (H K ) n +r 2 Hn; (8) s s where = r r is the Laplace-Beltrami operator, which contains the s s s fourth derivative of the surface position, posing numerical challenges to com- pute the bending forces [36]. For capsules with a vey thin hyperelastic membrane, two popular mem- brane constitutive laws are used herein: the Neo-Hookean (NH) law and the Skalak (Sk) law [11]. For these laws, the surface density of membrane energy is de ned upon a reference con guration S as, NH w = I 1 + 2 I + 1 ; (9) : Sk 2 2 w = I + 2I 2I + CI 1 2 s 1 2 where is the surface shear modulus, C represents the relative importance of the resistance to surface dilation, and I and I are the two strain invariants. 1 2 Finally, for red blood cells having a composite membrane, we use our re- cent RBC membrane model [44] to compute the surface force density. Brie y, the RBC membrane is modeled as a composite network, which consists of a dynamically triangulated surface as in a uid vesicle model. The membrane is then coupled to an additional network of springs with xed connectivity, representing the cytoskeleton. We explicitly compute the mechanical interac- tion between the bilayer and the cytoskeleton by considering normal elastic spring and tangential friction force. Speci cally, the FENE-POW spring 9 model is used to describe the elastic cytoskeleton, which yields a spring force at node n by an edge np 4 1 x x e 0 0 f = f = (x x ); (10) np p n 2 +1 2x 1 x x np np 3 + + 1 1 x where x (x ) is the position of vertex p (n). The normalized spring length p n max x = l =l 2 (0; 1] (the ratio of the spring length and maximum spring np np np length), x = x denotes the normalized spring length of edge np in the np reference shape, and is a constant repulsive parameter. This force, em- bodied on each spring edge, is transmitted to the lipid bilayer in the normal direction directly and in the tangential plane indirectly via drag forces. In this way, interfacial viscosity is added (see Ref. [44] for details). 2.3. Coupling conditions The interface conditions need to be imposed to complete the problem formulation. First, the uid motion is coupled with the interface motion via the kinematic boundary condition, i.e., continuity of the velocities at the interface e i u (x) = u (x) = u ; 8x 2 ; (11) where u is the velocity of uids at the interface. Second, assuming an impermeable membrane, at least on typical experi- mental time scales, the membrane is advected by the interface ow dx = u ; 8x 2 ; (12) dt where x is the membrane position. Finally, the dynamic boundary condition at the interface establishes a nonlinear interaction between bulk ows and membrane mechanics, f + f = 0; (13) wherein we assume the membrane is in quasi-static mechanical equilibrium; m e the membrane force density f balances the traction jump f ( ( ) n) exerted on the membrane by bulk uids, with the stress tensor pI + ru + (ru) . 10 We also compute several qualities of interest to show the simulation re- sults, such as the particle's shape and mobility. The translational velocity of the particle's center of mass in the streamwise direction is given by Z Z 1 1 i 3 U = (u e ) d x = x(u n) dS(x); (14) x x V V where V is the enclosed volume of the particle, which is calculated from Z Z V = d x = (x n)dS(x): (15) Its derivation from the initial given volume during simulations provides an indication of the accuracy of the computations. For vesicles, the relative surface area variation is also an indicator of the accuracy. The coordinates of the particle's center of mass are given by Z Z 1 1 3 2 X e = x d x = (x e ) (n e ) dS(x): (16) g i i i V 2V 3. Numerical method We use Loop's subdivision elements [39] to represent every quantity/ eld of interest: meshes (of the surface of objects and the wall surface of channels) and the unknown density elds { the interfacial velocity, the membrane force, and the disturbed wall traction. We begin with Loop's subdivision scheme, followed by a description of how the membrane forces are calculated with a uni ed formalism. The thus-obtained membrane forces are then used to compute the interfacial velocity using Green's function. Finally, we describe the interface advection schemes. 3.1. Isogeometric analysis 3.1.1. Subdivision surfaces Loop's subdivision surface is an assembly of linear triangle elements gen- erated through a limiting procedure of repeated re nement starting from an initial coarse mesh, called the control mesh of the surface. For contin- uously deformed particles, an icosahedron containing 20 equilateral triangle faces with ve meeting at each of its 12 vertices can be used as the initial control mesh (S in Fig. 2). The control mesh and all re ned meshes (by quadrisection) consist of triangles only. 11 A 6 G 4 G 0 1 2 S S S Fig. 2. An illustration of Loop's subdivision rule from the initial control mesh S (an 1 2 icosahedron) to S mesh (after one re nement) and S mesh (after two re nements). In Loop's subdivision scheme, each triangle of the coarse mesh is quadri- sected by introducing a new vertex at each edge midpoint, as illustrated in Fig. 2. The coordinates of the newly generated vertices (level k + 1) on the edge of the previous mesh (level k) are computed as k k k k p + 3p + 3p + p A B C D k+1 p = ; (17) and the old vertices are updated to get new nodal positions at the mesh k + 1 k+1 k k p = (1 q$)p + $ p : (18) G G G i=1 Here G (i 2 [1; q]) are the one-ring neighbours (at level k) of the vertex G, i.e., those vertices which share an edge with it, and q denotes the valence of a vertex, the number of element edges attached to a vertex [37]. The value of $, proposed by Loop [39], is given by " # 1 5 3 1 2 $ = + cos : (19) q 8 8 4 q Note that almost all newly generated vertices are regular (with valence q = 6), except for the twelve vertices (with valence q = 5) updated from the initial icosahedron mesh, which remain irregular. Loop's subdivision scheme produces limit surfaces which are globally C except at those irregular points where they are only C . However, the surfaces obtained by this scheme are H , i.e., have nite bending energy. 12 (a) (b) Fig. 3. An example of closed wall meshes generated by Loop's subdivision scheme. (a) on the inlet and outlet (circular and square) sections and (b) on the wall surface of (circular and square) microchannels. The mesh comprises N = 3360 elements and 1682 nodes for the circular tube and N = 7427 elements and 3714 nodes for the square channel. Two typical microchannel meshes generated by Loop subdivision are shown in Fig. 3, one for cylindrical channel and the other for a square chan- nel. Since the soft particle is kept at the center of the channel, this region has a more re ned mesh on the wall surface. The intersections (of the wall and the inlet/outlet surfaces) and the corners of the square channel are rounded with an arc-circle (radius 2 [O(R =6); O(R =4)] to avoid corner e ects when t t solving the ow with the boundary element method [23, 28]. 3.1.2. Isogeometric representation Stam [45] shows that the limit position of any point inside a triangle element e may be expressed in terms of box-spline shape functions, e p 1 2 x = X N (s ; s ); (20) p2one-ring 1 2 where the sum is taken over one-ring vertices, as shown in Fig. 4, (s ; s ) is a local parametrization of this point on the element, and N are the shape functions spanning over all one-ring elements [37, 45]. The nodal values p e X are the approximation parameters of the limit position x in the space expanded by the shape functions N . The parameterization in Eq. (20) may also be used for any scalar function f de ned on the membrane or at the 13 2 1 2 Fig. 4. A Loop element (shaded triangle) with its local parametrization s = (s ; s ) and its one-ring elements (bounded by solid lines), forming a regular Loop's patch containing 12 control vertices. channel wall, e.g., a Cartesian component of the membrane force f , the interfacial velocity u, and the disturbed wall traction f , e e 1 2 p 1 2 f (x) = f (s ; s ) = F N (s ; s ); (21) p2one-ring where F is the p-th nodal value. Equation (21) is used to evaluate f (i.e., the limit value) at any position x on element e if the nodal values F are known. Inversely, we also need to convert the limit value of a eld f into its nodal values F , that is, given the approximation of f under the form (21) such that the approximation error is minimized. Using the collocation formulation, in n n which the known eld f is collocated at vertices, i.e., f = f (x = x ) is known at vertex x n p 1 n 2 n f = F N (s (x ); s (x )) 8n 2 f1; ; N g; (22) p v p2one-ring 1 n 2 n where N (s (x ); s (x )) are the shape functions evaluated in the local pa- 1 2 n rameter space (s ; s ) corresponding to the vertex x , and N is the total number of vertices. Assembling the linear system (22) in matrix form ac- cording to the index of vertices, we then have n n ff g = CfF g; (23) n 1 2 N T n 1 2 N T v v where ff g = ff ; f ; ; f g , fF g = fF ; F ; ; F g , and C is the collocation matrix that transforms between the limit values f and the 14 n nodal values F . The regular surface patch is a quadratic spline [45], the derivatives of rst- and second-order can thus be realized by directly deriving on the shape functions, such as e 1 2 p 1 2 f (s ; s ) = F N (s ; s ); (24) p; p2one-ring where here and henceforth Greek indices takes the values 1 and 2, and a comma is used to denote partial di erentiation. For irregular Loop elements, the irregular patch must be subdivided until 1 2 the parameter value (s ; s ) of interest is within a regular patch, and then the canonical regular-patch evaluation routine works again [37, 46]. 3.2. Membrane solver { FEM 3.2.1. Weak formulation of FSI The membrane solver is designed to calculate the membrane force density f using the nite element method. The solver is based on the principle of virtual work for a deformable body. A detailed description is provided in [31]. For the sake of completeness, we recall brie y some basic concepts below. To describe the deformation of a surface from a reference con guration x (s) to its current con guration x(s), we introduce the tangential vectors at a point on the surface, which are given by the covariant base vectors a @x(s) a (s) = = x (s): (25) @s The unit normal vector n can be written as a (s) a (s) 1 2 n(s) = : (26) ja (s) a (s)j 1 2 Contravariant base vectors a are obtained through the relation a a = , where is the Kronecker delta. With the tangential and normal vectors, we can write the rst and second fundamental forms of the surface a (s) = a (s)a (s); b (s) = a (s)n(s); (27) where a and b denote the metric and curvature tensors of the surface, 1 2 respectively. The di erential area element of the surface dS = ads ds , 15 with a = det(a ) the determinant of the metric tensor. The above de ni- tions and relations hold for the reference con guration as well, with x (s) replacing x(s). By virtue of the principle of virtual work, a membrane is in equilibrium if the sum of internal and external virtual work vanishes W + W = 0: (28) int ext The external virtual work is given by W = (f + g) xdS; (29) ext where (and in what follows) means that a variable derives from a virtual displacement x, and g is some additional body forces (e.g., buoyancy) acting on the membrane in addition to the term f representing the traction jump across the membrane. According to Ref. [37], the internal virtual work of the membrane can be written as W = (E ) + (B ) dS; (30) int where and are the e ective membrane and bending stress tensors, respectively. They are membrane dependent { namely its position and me- chanical properties. The Green-Lagrange strain tensor E = a a (31) represents in-plane deformation, i.e., stretching, while the bending strain tensor B = b b (32) describes out-of-plane deformation, i.e., the change in curvature or bending strains. Note that f = f (Eq. (13)), so for neutrally buoyant particles, Eq. (28) reads m 2 (a ) + (b ) + f x dS = 0; 8x 2 H ( ): (33) 16 This weak formulation of a uid-membrane interaction problem gives a gen- eral relationship between membrane force density f and its position x. Using isogeometric nite element which ensures H , this uni ed formalism makes it possible to study deformable objects spanning from a simple liquid drop to elastic capsules and vesicles with bending sti ness. The membrane forces are obtained once the mechanical properties of the membrane are spec- i ed via the membrane ( ) and bending ( ) stress tensors. 3.2.2. Membrane constitutive laws For liquid drops, the surface energy per unit area is the interfacial tension, w = , independent of curvature, thus the bending stress tensor = 0. Hence, the membrane stress tensor is given by = a : (34) For a capsule with a very thin elastic membrane, the surface energy den- sity w is given by Eq. (9) for the NH and Sk laws, we also have = 0. Following [47, 40], the membrane stress tensor can be written as 2 @w @w s s 0; = a + 2J a ; (35) J @I @I s 1 2 where J = a= a is the Jacobian of the transformation from the reference to the deformed con guration. For a lipid membrane satisfying the Helfrich bending energy subjected to the surface incompressibility constraint (Eq. (8)), the membrane and bending stress tensors are given by [31] > 2 @ aw > 2 = p = 4H a 8Hb + a a @a 2 : (36) @w = = 4Ha @b 2 Finally, for an RBC having a composite membrane, the cytoskeletal elas- tic forces f are computed directly based on a spring network (Eq. (10)). These forces are added to the lipid bilayer (Eqs. (36) and (33) in the normal direction directly and in the tangential plane indirectly via drag forces [44]. 17 3.2.3. Calculation of membrane force Since surfaces obtained by Loop's subdivision scheme are globally C ex- cept at some xed irregular points, the curvature tensor b at any quadrature points (12 Gauss quadrature points are used) can be computed by direct dif- ferentiation of Loop's shape functions. The metric tensor a and the unit normal vector n are readily obtained from the interpolation of the position. As such, the nite-element discretization of Eq. (33) using the Loop shape functions for the Cartesian components of membrane force and position leads to a matrix-vector form for the unknown nodal values of the membrane force Mff g = frhsg: (37) The mass matrix M and the right hand side vector frhsg are formed by nu- merical integration of Eq. (33) using 12 Gauss quadrature points, see Ref. [31] for details. 3.3. Fluid solver { BEM Under Stokes ow conditions, boundary integral equations for the inter- facial velocity and the disturbed wall traction can be expressed as [23, 48], on the membrane surface 1 + 1 m w a u(x ) = u (x ) +S f (x ) S f (x ) p S n(x ) 0 0 0 W 0 O 0 PV +(1 )D u(x ); x 2 (38) 0 0 and at the channel wall (the no-slip condition) w m a S f (x ) = S f (x ) p S n(x ) + (1 )D u(x ); x 2 W (39) W 0 0 O 0 0 0 where u is the interfacial velocity, u is the velocity of the ambient ow (i.e., the ow without the deformable object), f is the disturbed wall traction, i e and ( = ) is the viscosity ratio between the internal and external uids. The single-layer operator S and double-layer operator D are de ned as (S ) (x ) = (x)G (x;x )dS(x); (40a) 0 i ij 0 (D ) (x ) = (x)T (x;x )n (x)dS(x); (40b) 0 i ijk 0 k 18 where G (Stokeslet) and T (stresslet) are the Green's functions in the three- PV dimensional free space [23]. D indicates that the double-layer integral is evaluated in the principal-value sense when the point x lies on the integra- tion domain . The term p , called additional pressure drop, is due to the presence of a particle in the channel ow which causes an increase in the pressure drop across the channel. It can be calculated by the reciprocal theorem of Stokes ow [48] a m 1 1 p = [f u + (1 )f u] dS(x); x 2 (41) where Q is the total ow rate, which is assumed not disturbed by the presence of the deformable object, i.e., Q = Q . The equations (38) and (39), together with (41) allow to determine the interfacial velocity u and the disturbed wall traction f , as well as the additional pressure drop p . The last quantity has a direct implication in the rheological properties of a dilute suspension [48]. For the sake of simplicity, we performed computations in this paper only with unity viscosity ratio (i.e., = 1), unless speci ed otherwise. The singularity in the single- layer integrals is treated in two ways, depending on the integration domain; one consists in a singularity subtraction technique proposed in [30] when the integration domain lies on the membrane surface, and the other involves transforming the parametric triangular to polar coordinates as introduced in [23] when it lies at the channel wall. For a vesicle, an additional eld, namely the membrane tension , remains to be determined. It is the solution of the surface velocity incompressibility constraint (6), which is solved by an iterative method [31]. 3.4. Time-stepping schemes As have been implemented in [31], two time-stepping schemes, i.e., a high- order explicit scheme and a second-order implicit scheme, are used to update the new membrane position x at time t = t + t. n+1 n+1 n 3.4.1. Runge-Kutta-Fehlberg scheme The explicit time-stepping scheme consists of a Runge-Kutta Fehlberg fourth- fth (RKF45) stage scheme [49]. This high-order scheme allows dy- namically adapting the time step t = h and provides very good conser- vation of invariants such as the enclosed uid volume. The fourth and fth stage formulations are given by 19 25 1408 2197 1 (4) x = x + k + k + k k (42) n 1 3 4 5 n+1 216 2565 4101 5 16 6656 28651 9 2 (5) x = x + k + k + k k + k ; (43) n 1 3 4 5 6 n+1 135 12858 56430 50 55 where k correspond to the intermediate values [49]. The dynamic time step is realized by comparing the di erence between the fourth and fth stage results (4) (5) = maxjx x j with the two pre-setting tolerances " ( 10 ) n max n+1 n+1 and " ( 10 ) min 1=4 > max h = h if > " n n n max > 2 1=4 : (44) min h = h if < " > n+1 n n min > 2 h = h else n+1 n 3.4.2. Trapezoidal scheme While the RKF45 time scheme is usually used in the simulation of drops and capsules, the bending sti ness of a vesicle precludes its use in the sim- ulation of vesicle dynamics since the stability condition imposes very small e 3 time-steps, namely t O( x =), for an explicit time-stepping scheme to be numerically stable [33]. Hence, an implicit scheme is needed. The implicit time scheme is the trapezoidal rule { an implicit second-order Crank-Nicolson time integration. For a given position and tension (x ; ) at n n time t , the position and tension (x ; ) at t are nonlinearly coupled n n+1 n+1 n+1 such as x = x + [u(x ; ) + u(x ; )] n+1 n n n n+1 n+1 2 : (45) r u(x ; ) = 0 s n+1 n+1 These equations are solved iteratively using the Jacobian-free NewtonKrylov method [50] (see [31] for details). 4. Numerical setup and validation 4.1. Dimensionless groups Let V and A denote the enclosed volume and the surface area of a de- formable particle, respectively. The volume remains constant and de nes 20 1=3 a length scale R = (3V=4) . For a lipid vesicle, the surface area of the 3=2 membrane also remains constant. Then the reduced volume = 6 V A (0 < 1) measures the asphericity of the vesicle. In the present work, we consider only laminar viscous ow through a uniform channel of either circular or rectangular (square) in cross-section. Let R denote the charac- teristic dimension of the ow channel (the radius of a cylindrical tube or the half-width of the cross-section of a rectangular channel), then the ra- tio = R=R measures the ow con nement; the particle's motion is more signi cantly hindered by particle-wall interactions as the con nement in- creases. The length-to-diameter (width) ratio of the channel is de ned by = 2` =(2R ), with 2` being the total length of the channel. x t x In the absence of any particle or far from the particle, the ow approaches the unperturbed ow u in a channel. It is Poiseuille ow with a parabolic velocity pro le for a circular tube 2 2 y + z u = 2U 1 e ; (46) where U = Q=(R ) is the mean velocity with 2U representing the maximum undisturbed velocity at the centerline of the tube. The unperturbed ow in a rectangular channel is given in Ref. [51], see also Ref. [52], that is u e b y b z x m m 2 2 = ` z + B cosh cos ; (47) ` ` z z m=1 with m 2 1 dp (2m 1) ( 1) 4 ` = ; b = ; B = ; (48) m m e 3 2 dx 2 b cosh (b ` =` ) m y z where ` and ` denote the channel's half-height and half-width, respectively. y z Integrating over the channel's cross-section yields the volumetric ow rate Q, Q 8` ` 2` b ` y z m y = + B sinh sin (b ) : (49) m m 3 b ` m z m=1 The mean velocity is U = Q=(` ` ). Substituting Eq. (49) into (47) leads to a y z velocity pro le u that is proportional to the mean velocity U and depends on the aspect ratio of the channel's cross-section ` =` . In our simulations, we y z set m = 40, as in [52]. The maximum undisturbed velocity at the centerline of a square channel is approximately 2:1U . 21 In addition to the dimensionless geometrical parameters mentioned above ( and ), the interfacial mechanical property of a deformable particle im- mersed in a viscous ow introduces a dynamic dimensionless parameter, the capillary number Ca = = , which is the ratio of the characteristic shape sr f relaxation timescale to a viscous timescale . Speci cally, the capillary sr f number for surface tension dominant liquid drops, Ca = U= , measures the rel- ative importance of viscous forces to interfacial tension forces; e 2 bending dominant membrane of vesicles, Ca = U R =, is a ratio of viscous stress to resistive bending stress on the membrane; shearing dominant membrane of capsules, Ca = U= , determines the relative importance of viscous forces to resistive elastic forces on the membrane. In this study, we assume that uid ows at an imposed, constant volumetric ow rate Q driven by a pressure di erence between the channel's inlet and outlet. Hence, the dynamical behavior of a deformable particle owing in a tube or square channel is determined only by four independent dimensionless groups: the con nement , the reduced volume (only for vesicles), the capillary number Ca, and the viscosity ratio (which is set to unity, unless speci ed otherwise). Numerical solutions should be independent of the total length of the channel provided that it is suciently long for the disturbances to become negligibly small at the channel ends. So, we present in the next subsection two numerical examples to show the e ects of the channel's length and the minimum element size (` ) of the channel mesh, and then give a min general criterion for the choice of these two parameters. 4.2. E ect of the channel's length and wall mesh Figure 5 shows the numerical results of an initial prolate vesicle ( = 0:9, the membrane surface discretized with N = 320 Loop elements) owing in a cylindrical tube with di erent lengths. Fig. 5(a) shows the disturbed shear stress f along the tube wall in the xy-plane (see Fig. 1) for di erent tube lengths 2` by varying the ratio . As can be seen from this gure, the disturbed shear stress f decreases exponentially with distance from the vesicle and are vanishingly small towards the channel ends when 5. This observation is consistent with the analysis of Liron and Shahar [53], who 22 0.06 0.16 0 −3 ×10 (a) (b) ζ=3.0 ζ=3.0 5.0 0.04 4.0 0.15 -5 7.0 5.0 -1 6.0 0.02 0.14 -10 8 12 16 20 24 7.0 0 0.13 -15 θ 3 R κ/R -0.02 0.12 -20 -0.04 0.11 -25 1.35R -0.06 0.1 -30 -30 -20 -10 0 10 20 30 0 10 20 30 40 50 60 e 3 x/R κt/(η R ) Fig. 5. E ect of the channel length quanti ed by for a vesicle in a tube ow with ` = 0:18R . The parameters are = 0:9, = 0:25, and Ca = 1. (a) The disturbed min t shear stress on the tube wall f for di erent values of . The upper inset is a zoomed-in view. The bottom inset shows the steady vesicle pro le in the xy-plane with G being the vesicle's centroid. (b) Temporal evolution of the centroid (Y =R) and inclination angle (, in degree). showed the perturbation ow in tube generated by a point-force distribution decays exponentially with distance from the source point. Interestingly, the steady-state shape [Fig. 5(a)], as well as the temporal evolution curve of the centroid Y and the inclination angle [Fig. 5(b), the angle between the vesicle major axis and the ow direction e ], is insensitive to far- eld perturbations, which is important when studying the dynamical behavior of a deformable particle in a channel ow. Indeed, the numerical results are virtually indistinguishable when 3. Regarding the minimum element size ` that may be required to obtain min accurate results, we have set, in this example, ` = 0:18R , which is close to min t that used by Hu et al. [28] in highly con ned conditions. Moreover, our test of a liquid drop in a cylindrical tube with a wide range of ` 2 [0:02R ; 0:56R ] min t t shows that the resulting f matches very well at the same grid points in the wall mesh. Again the drop shape remains essentially una ected, as shown in Fig. 6. Even in the coarsest mesh tested, i.e., ` = 0:56R , in which min t the peak of the wall shear stress is not captured due to a lack of grid points there, the obtained values at the existing grid points are very close to those obtained by ner meshes. In this study, the general rule is that we set the ratio 5{7 and the dimensionless minimum element size in the wall mesh ` =R = min [O(h=R); O(0:2 )], where h=R is the dimensionless gap size min 23 0.5 ` /R = min t 0.4 0.56 3.386R 0.4 0.3 0.24 0.2 0.04 f 0.02 0.1 η U/R -0.1 -0.2 -0.3 -10 -8 -6 -4 -2 0 2 4 6 8 10 x/R Fig. 6. The disterbuted shear stress on the tube wall f for a liquid drop (N = 1280) for di erent ratios ` =R . The parameters are Ca = 0:5, = 0:8, and = 7:5. The inset min t shows the steady drop pro le in the xy-plane. between the particle surface and the channel wall. 4.3. Numerical validation We validate the coupled isogeometric FEM-BEM approach by comparing the simulation results with a well-know example of a (clean, surfactant-free) liquid drop in tube ow, for which very highly accurate numerical compu- tations are available in the literature (e.g., Ref. [24]). The motion of the drop for given con nement is determined only by the capillary number Ca (apart from the viscosity ratio , which is set to unity for this compar- ison). Of particular interest are the drop relative velocity U =U and the a e dimensionless additional pressure drop p =( U=R ), the latter is due to the presence of the drop in tube ow in order to maintain the volumetric ow rate Q = R U . Our 3D numerical results (with N = 320 elements) are compared with the axisymmetric simulations reported in [24]. The compari- son in Fig. 7 shows excellent agreement. Under weak con nement (i.e., small ), the simulation results are also in excellent agreement with the theoretical predictions for a vanishingly small droplet moving along the centerline of a tube [54, 55], given by 24 U 4 2 3 = 2 + O( ); (50a) U 5 R 24 a 5 10 p = + O( ): (50b) U 5 2.2 100 (a) Present simulations (b) Lac & Sherwood (2009) Ca = 0.05 Ca = 1.8 0.1 0.5 0.2 a Δp 0.4 1.6 0.4 x η U/R 0.2 t 0.5 1.4 0.1 0.1 0.05 1.2 0.01 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0.2 0.5 1 1.5 β β Fig. 7. (a) droplet relative velocity U =U and (b) dimensionless additional pressure drop a e p =( U=R ) as a function of the con nement compared to those reported in Ref. [24]. The dashed curves are the theoretical predictions of (50a) and (50b) for 1. (a) (b) −2/3 Ca −5/3 Ca 0.1 β=0.8 β=0.8 0.1 β=1.1 β=1.1 0.01 0.1 1 0.01 0.1 1 Ca Ca Fig. 8. (a) droplet relative velocity 2 U =U and (b) dimensionless additional pressure a e drop p =( U=R ) as a function of the capillary number Ca for = 0:8 and 1.1. The dashed curves are the scalings of (51a) and (51b) at high Ca. As an additional veri cation and validation of the numerical model, we plot, in Fig. 8, the e ect of the capillary number Ca on the drop velocity and 2 − U /U a e Δp R /(η U ) t the additional pressure drop for two values of con nement = 0:8 and 1.1. Fig. 8(a) shows how much the drop velocity exceeds the mean velocity of the suspending uid, approaching the axis velocity 2U and resulting in a dramatic decrease in the additional pressure drop [Fig. 8(b)] as Ca increases. At high a e capillary numbers, the dimensionless groups 2 U =U and p =( U=R ) x t exhibit a remarkable power law. Lac and Sherwood [24] provided via the asymptotic analysis for a long slender drop in tube ow the following scalings at high Ca 2=3 2 U =U Ca ; (51a) a 5=3 p Ca : (51b) As shown, the present 3D simulations captured these limiting behaviors. -2 0.02 (a) (b) -3 0.016 -4 -5 0.012 -6 D ε xy 10 v 0.008 -7 80 10 -8 0.004 -9 -10 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 e e γt/(η R) γt/(η R) Fig. 9. Temporal evolution of (a) the Taylor deformation parameter D and (b) the xy relative derivation of the enclosed volume " = V=V 1 of a drop in a capillary ow v 0 (Ca = 0:05 and = 0:8) as a function of the number of elements N used on the drop surface. We then proceed to conduct a convergence study on the spatial and tem- poral discretization for one of the above settings at Ca = 0:05 and = 0:8. The deformation behavior of the drop is characterized by the Taylor defor- mation parameter D = (L B)=(L +B), where L and B are the major and xy minor axis of the drop pro le in the xy-plane. The temporal evolution of D xy plotted in Fig. 9(a) shows that except for the coarsest mesh (80 elements), which gives a non-converged solution, the simulations from other numbers of mesh elements ranging from 320 to 20480 lead to very good agreement 26 results. We assess the e ect of mesh re nement by examining the enclosed volume V and its drift from the initial volume V : " = (V V )=V . It 0 v 0 0 is seen that the volume drift is only 0.4% after a long-time simulation (i.e., t=( R) = 30) with 320 elements. Its temporal evolution (for N ranging from 320 to 20480) displayed in Fig. 9(b) suggests that one more subdivision process (i.e., multiplying the number of elements by four) leads to at least one order of magnitude gaining in volume conservation. We performed these simulations from an initially spherical drop using the RKF45 time-stepping scheme (i.e., adaptive time step scheme), while the computations with the nest mesh (20480 elements) started from a drop shape obtained with 5120 elements. Using N = 1280 elements, we assess the e ect of temporal discretization by varying the time step from 0.1 to 5 10 . As can be seen from Fig. 10, the implicit time integration with these time steps leads to a consistent, very good agreement result. 0.02 0.016 0.1 7.5e-2 0.012 5e-2 0.0168 xy 2.5e-2 1e-2 0.008 5e-3 0.0164 5e-4 0.004 16 20 24 0 5 10 15 20 25 30 γt/(η R) Fig. 10. Temporal evolution of the Taylor deformation parameter D of a drop (N = xy 1280 elements) owing in a capillary ow (Ca = 0:05; = 0:8) as a function of time steps (scaled by R= ). Time-stepping scheme used is the trapezoidal rule with a xed time step. To conclude this validation subsection, we provide an estimate of the convergence rate of the numerical method. The convergence order of spatial and temporal discretization is evaluated via the relative error of the Tay- lor deformation parameter " = (D D )=D at a dimensionless time xy ref ref t=( R) = 30 (a steady state), where D is the reference Taylor deforma- ref 27 tion parameter. Fig. 11 makes it clear that the present algorithm preserves second-order convergence in both space and time for a liquid drop con ned in capillary ow. For the spatial convergence shown in Fig. 11(a), the RKF45 scheme is used, and the reference value is computed with 20480 elements. For the temporal convergence displayed in Fig. 11(b), the trapezoidal scheme is used with 1280 elements, and the reference value is computed with a time 4 e step t = 5 10 (scaled by R= ). The previous study [31] on unbounded soft particles shows that the second-order convergence is not a ected by the membrane's constitutive law, including bending sti ness, we may expect that the second-order convergence achieved for con ned liquid drops in tube ow is also applicable to other con ned soft particles. -2 -3 10 10 ε ε -3 −2 2 N Δt -4 -4 -5 -5 10 10 -6 -6 -7 (a) (b) -8 -7 10 10 2 3 4 -2 -1 10 10 10 10 10 Time step Δt Number of elements N Fig. 11. Relative error of the Taylor deformation parameter of a drop in a capillary ow (Ca = 0:05 and = 0:8) as a function of (a) the number of elements N and (b) the time step size t (scaled by R= ). 5. Numerical examples In this section, we present further simulation results to demonstrate the accuracy and stability of the numerical method. Having described drop dy- namics in Section 4, we focus now on the other three types of soft particles to illustrate potential applications. Speci cally, we simulate (i) an elastic capsule in a square channel, (ii) a vesicle in a cylindrical tube, and (iii) a single RBC in a capillary. Where it is possible, we compare the numerical results with previously published studies. Relative error ε 5.1. Elastic capsule in a square channel As shown in Fig. 12, the rst simulation example concerns the steady- state deformation of an initially spherical capsule moving through a square microchannel with the undisturbed ow in the channel u being given by (47). Fig. 12. An elastic capsule owing in a square microchannel. The channel wall mesh is generated by Loop's subdivision process and rounded with an arc-circle. Colors on the surface of the capsule represent the x-component of the membrane elastic force (f ). We performed computations with the strain-hardening Sk law (9) with C = 1, as used in [56], for Ca = 0.02, 0.05, and 0.1, and at = 0:85. Simulations are run for two capsule meshes consisting of 320 and 1280 Loop elements. The capsule pro les at steady-state are displayed in Fig. 13(a)-(c) in the xy-plane and in Fig. 13(d) in the yz-plane. Increasing the capillary number, or equivalently decreasing the membrane elasticity, leads to the cap- sule less able to retain its spherical shape and more elongated. The obtained results are compared with those reported in Ref. [56] in which the surface of the capsule is discretized by 1280 quadratic triangular element [34]. Thanks to the use of Loop elements for the representation of geometry and the two solvers in our simulations, even a coarse mesh of 320 elements reproduces the numerical results of [56], which represents a signi cant improvement. 5.2. Vesicle in a circular tube The second numerical example deals with a con ned vesicle owing through a circular tube, as illustrated in Fig. 14. In aqueous solution, lipid vesicles exhibit a large variety of shapes and shape transformations, in particular, they can exhibit a biconcave shape typical of red blood cells. When con ned 29 (a) (b) 1 1 0.5 0.5 Hu et al. y y 0 320 0 -0.5 -0.5 -1 -1 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 x x (c) (d) 1 1 0.5 0.5 y 0 0 -0.5 -0.5 -1 -1 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 Fig. 13. Comparison of the steady-state deformation of a capsule owing in a square channel with those obtained by Hu et al. [56] for = 0:85. Pro les in the xy-plane: (a) Ca = 0:02, (b) Ca = 0:05, and (c) Ca = 0:1. Pro le in the yz-plane: (d) Ca = 0:1. The blue lines indicate channel walls. Fig. 14. Deformation of a vesicle in tube ow for Ca = 10, = 0:9, and 1= = 1:2. Colors on the vesicle surface represent the local mean curvature. 30 in capillary tubes subjected to Poiseuille ow, however, vesicles assume com- plex shapes and behave in di erent ways due to an intricate interplay between ow stresses, membrane's bending rigidity, and con nement [57, 14]. y y x z β=0.125 β=0.2 Fig. 15. The steady-state slipper shape (top, = 0:125) and croissant shape (bottom, = 0:2) of a vesicle ( = 0:9, discretized by 1280 elements) in a con ned axial Poiseuille ow (Ca = 10). Left: side view in the xy-plane; right: rear view in the yz-plane. Colors on the vesicle surface represent the local mean curvature, and the arrows show the membrane ow. To illustrate two types of 3D vesicle shapes at steady state in a con- ned axial Poiseuille ow, we performed computations for an initially pro- late vesicle of the reduced volume = 0:9 at two con nement conditions, i.e., = 0:125 and 0.2. The vesicle, initially located at a height H (= 0:06, refer to Fig. 17) from the ow axis, rapidly changes its shape, becoming a slipper shape, which is characterized by a single mirror symmetry in the yz- plane due to the ow curvature. At very weak con nement (i.e., = 0:125), the slipper shape reaches a steady-state with the inward migration ending 2 2 at a certain position to the centerline H = y + z 0:028. This nal stationary shape, as well as its membrane ow structure, is illustrated in the 31 top panel of Fig. 15. The membrane ow is characterized by two unequal vortices both on the front and rear faces of the membrane. At slightly high con nement ( = 0:2), the transitional slipper shape is unstable, becoming a croissant shape characterized by two mirror symmetries in the yz-planes, and its radial position H decrease to zero. The membrane ow now consists of the two equal vortices both on the front and rear faces of the membrane, as illustrated in the bottom panel of Fig. 15. These are long-time simula- tions and steady state using the implicit time integration is reached around e 3 3 t=( R ) 300 with a dimensionless time step t = 4 10 . The rela- tive error in the enclosed volume and the total surface area are respectively " 0:1% and " 0:045% (for = 0:125), and 0:07% (for = 0:2), v A indicating the high accuracy and stability of the present algorithm. 1.2 λ=1 0.9 −2 ×10 0.6 -1 -2 0.3 -3 0 40 80 120 160 200 0 30 60 90 120 150 180 e 3 κt/(η R ) Fig. 16. Temporal evolution of the centroid Y of a vesicle ( = 0:9) in a weakly con ned ( = 0:1) Poiseuille ow (Ca = 100) for two di erent viscosity ratios: = 1 and = 6. The inset shows the evolution of the lateral migration velocity U (left) and vesicle shapes from side and rear views (right) with colors representing the local mean curvature. For the sake of simplicity, all these numerical examples are limited to a viscosity ratio of unity (i.e., = 1). To demonstrate the ability of the present 32 code to handle non-unity viscosity ratios, we show in Fig. 16 simulation results for a vesicle in a weakly con ned Poiseuille ow at = 1 and 6. In the absence of viscosity contrast, the vesicle migrates towards the center (i.e., Y ! 0 and U ! 0), and an axisymmetric parachute shape is obtained, as g y illustrated by the insets surrounded by red dashed lines. While the situation is di erent for = 6, starting from an initial position H = 0:5R, the vesicle migrates outwards, and an asymmetric shape is now produced, as shown by the insets surrounded by a purple dashed lines. These simulations results are consistent with a previous study [57], which has shown that depending on the viscosity contrast , a vesicle in an unbounded Poiseuille ow can migrate either inward towards the center, or outward of the ow at high Ca if the initial position is chosen suciently far from the centerline. We note that the e ect of the viscosity ratio on the dynamics of a con ned vesicle is the subject of a very recent study [58]. 5.3. RBC in capillary ow The last numerical example concerns a single RBC in capillary ows. The initial biconcave discoid shape of RBC is given by the following expres- sion [59] " # 2 2 2 2 2 2 4(x + z ) x + z (x + z ) y = D 1 a + a + a ; (52) 1 2 3 2 2 4 D D D where D = 7:82 μm is the cell diameter, a = 0:0518, a = 2:0026 and 1 2 a = 4:491. The initial shape is shown in Fig. 17 with a clip to better represent its three-dimensional structure. The volume and surface area of the 3 2 corresponding RBC are respectively 94 μm and 135 μm , giving a reduced volume = 0:64 and an e ective diameter D = 6V= = 5:64 μm. As in the case of a vesicle owing in capillary, Poiseuille ow is given by Eq. (46), where U = 0:016 cm s for all cases considered in this subsection, which lies 1 1 in the range of 0:001 cm s to 1 cm s exploited by Pozrikidis [48]. For the membrane with shape memory, the reference shape (i.e., unstressed shape) is another in uencing ingredient to compute the elastic force [Eq. (10)]. Here, the initial biconcave form (52) is used as the unstressed shape. First, we consider that RBCs are initially placed on the axis of the capil- lary (H = 0:0, = 0:5) and that their at surfaces are orthogonal to the ow ( = 90 ) with varying elastic moduli = 0:0; 0:5; 5:0 and 10:0 μN m . 0 s The steady shapes, shown in Fig. 18, suggest that the cell deformation is 33 θ flow direction (a) microchannel axis Fig. 17. Schematic representation of the initial con guration of a biconcave red blood cell (colors on the membrane represent jxj), where H and are respectively the initial o set 0 0 position of its centroid relative to the ow axis and the initial inclination angle measured from its at plane to the axis of the ow. signi cantly reduced at higher resistance to elastic forces since an RBC with a higher shear modulus has a greater ability to withstand hydrodynamic stresses. Fig. 18. The steady shapes of RBCs in a capillary for four di erent shear moduli, 0.0, 0.5, 5.0 and 10.0 μN m (from left to right). We then consider a case in which the at plane of RBCs is not placed orthogonally to the ow direction, but only with a small inclined angle 6 , and is placed at H = 0:3 in a capillary with = 0:4. By 0 0 varying the shear modulus of the membrane from 0 μN m (vesicle) to a relatively high value 4.0 μN m , we obtain a totally di erent shape evolution process, which depends upon the shear modulus, as shown in Fig. 19. time µ =0 µ N/m 0.1 0.5 4.0 Fig. 19. Temporal evolution of the RBCs shape in a capillary ow ( = 0:4) for ve di erent shear moduli = 0:0; 0:1; 0:5, and 4:0 μN m with H = 0:3 and 6 . s 0 0 Membranes are colored by the local mean curvature. The snapshots in the rst row of Fig. 19 show that an initially biconcave vesicle ( 0:64) placed at H = 0:3 evolves into a slipper shape. The overall evolution of the shape remains essentially unchanged up to 0:1 μN m . A di erent transition occurs by increasing the shear modulus to 0.5 μN m , as shown in the third row of Fig. 19; a biconcave shape undergoes a transition into a biconcave-croissant shape (a biconcave shape with two planes of symmetry like a croissant shape), for which the two dimples are preserved during the transition. The biconcave shape is rst stretched under the action of the ow stresses, but the subsequent stretching is mitigated by the cytoskeletal forces, which are characterized by the cytoskeletal shear modulus . 35 6. Conclusions We have presented a coupled isogeometric FEM and BEM computational framework to accurately predict the ow of a single soft particle con ned inside a micro uidic channel. An outstanding advantage of our numerical developments is that they integrate di erent types of soft objects in a unique framework, the only di erence being in the description of the interfacial me- chanics and the time integration schemes. Thanks to the use of Loop elements for the representation of the geometry and unknown elds, the uni ed for- malism established in the weak formulation of this kind of uid-membrane interaction problem allowed us to eciently study con ned soft objects span- ning from a simple liquid drop to a membrane-enclosed particle with shear or/and bending resistance, such as a capsule, a vesicle, and even a red blood cell. We have validated the numerical method by comparing the simulation results with highly accurate computations of a liquid drop moving through tube ow, showing a second-order convergence in both space and time. We have carried out several additional simulations to illustrate the possible ap- plications of the current numerical method, while also demonstrating the accuracy and stability of the algorithm. Taken together, the code developed here provides a solid base for making a reliable prediction of the dynami- cal behavior of a con ning deformable particle in channel ows, such as the phase diagram, the shape transition, and the lateral migration of 3D vesicles. These are indeed the subject of our ongoing investigation. Finally, we point out potential extensions that are closely related to the present work. One of the directions of interest is applying the Loop sub- division to the channel walls only without incorporating the inlet and out- let sections. Such an implementation with a periodic Green's function is highly useful to micro uidic applications and biological ows involving de- formable walls. Micro uidic channels are usually fabricated with soft materi- als which may experience substantial deformations due to the uid stresses. A deformable wall may be modeled as an elastic shell under the bending- dominated regime in the Kirchho Love equation. A C representation of the wall by Loop elements, as developed in the present study, is then a key to accurately simulating this kind of uid-structure interaction. Declaration of competing interest The authors declare that they do not have any con ict of interests. 36 Acknowledgments We acknowledge nancial supports from Labex MEC (grant no. ANR- 11-LABX-0092), from A*MIDEX (grant no. ANR-11-IDEX-0001-02), from ANR (grant no. ANR-18-CE06-0008-03), from the LabEx Tec21 (ANR-11- LABX-0030), from the PolyNat Carnot Institute (ANR-11-CARN-007-01) and from CNES. J.M. Lyu was sponsored by the China Scholarship Council (CSC). Centre de Calcul Intensif dAix-Marseille is acknowledged for granting access to its high performance computing resources. References [1] H. Stone, A. Stroock, A. Ajdari, Engineering ows in small devices: Micro uidics toward a lab-on-a-chip, Annu. Rev. Fluid Mech. 36 (1) (2004) 381{411. doi:10.1146/annurev.fluid.36.050802.122124. [2] L. Shui, J. C. Eijkel, A. van den Berg, Multiphase ow in micro uidic systems { Control and applications of droplets and interfaces, Adv. Col- loid. Interface Sci. 133 (1) (2007) 35{49. doi:10.1016/j.cis.2007.03. [3] T. M. Geislinger, T. Franke, Hydrodynamic lift of vesicles and red blood cells in ow | from F ahrus & Lindqvist to micro uidic cell sorting, Adv. Colloid. Interface Sci. 208 (2014) 161{176. doi:10.1016/j.cis. 2014.03.002. [4] C. Wyatt Shields IV, C. D. Reyes, G. P. L opez, Micro uidic cell sorting: a review of the advances in the separation of cells from de- bulking to rare cell isolation, Lab Chip 15 (2015) 1230{1249. doi: 10.1039/C4LC01246A. [5] T. W. Secomb, A. R. Pries, Blood viscosity in microvessels: Experiment and theory, C. R. Phys. 14 (6) (2013) 470{478. doi:10.1016/j.crhy. 2013.04.002. [6] G. Tomaiuolo, M. Simeone, V. Martinelli, B. Rotoli, S. Guido, Red blood cell deformation in microcon ned ow, Soft Matter 5 (2009) 3736{3740. doi:10.1039/B904584H. 37 [7] D. Liu, H. Zhang, F. Fontana, J. T. Hirvonen, H. A. Santos, Micro uidic- assisted fabrication of carriers for controlled drug delivery, Lab Chip 17 (11) (2017) 1856{1883. doi:10.1039/C7LC00242D. [8] C. N. Baroud, F. Gallaire, R. Dangla, Dynamics of micro uidic droplets, Lab Chip 10 (2010) 2032{2045. doi:10.1039/C001191F. [9] S. L. Anna, Droplets and bubbles in micro uidic devices, Annu. Rev. Fluid Mech. 48 (1) (2016) 285{309. doi:10.1146/ annurev-fluid-122414-034425. [10] P. M. Vlahovska, D. Barthes-Biesel, C. Misbah, Flow dynamics of red blood cells and their biomimetic counterparts, C. R. Phys. 14 (6) (2013) 451{458. doi:10.1016/j.crhy.2013.05.001. [11] D. Barth es-Biesel, Motion and deformation of elastic capsules and vesi- cles in ow, Annu. Rev. Fluid Mech. 48 (1) (2016) 25{52. doi: 10.1146/annurev-fluid-122414-034345. [12] J. B. Freund, Numerical simulation of owing blood cells, Annu. Rev. Fluid Mech. 46 (1) (2014) 67{95. doi:10.1146/ annurev-fluid-010313-141349. [13] R. Trozzo, G. Boedec, M. Leonetti, M. Jaeger, Axisymmetric boundary element method for vesicles in a capillary, J. Comput. Phys. 289 (2015) 62{82. doi:10.1016/j.jcp.2015.02.022. [14] P. G. Chen, J. M. Lyu, M. Jaeger, M. Leonetti, Shape transition and hy- drodynamics of vesicles in tube ow, https://arxiv.org/abs/1912.06937 (2019). [15] D. Abreu, M. Levant, V. Steinberg, U. Seifert, Fluid vesicles in ow, Adv. Colloid. Interface Sci. 208 (2014) 129{141. doi:10.1016/j.cis. 2014.02.004. [16] Z. Li, K. Ito, M.-C. Lai, An augmented approach for stokes equations with a discontinuous viscosity and singular forces, Comput. Fluids 36 (2007) 622{635. doi:10.1016/j.compfluid.2006.03.003. [17] S. Mendez, E. Gibaud, F. Nicoud, An unstructured solver for simula- tions of deformable particles in ows at arbitrary Reynolds numbers, J. Comput. Phys. 256 (2014) 465{483. doi:10.1016/j.jcp.2013.08.061. 38 [18] H. Ye, H. Huang, X. yun Lu, Numerical study on dynamic sorting of a compliant capsule with a thin shell, Comput. Fluids 114 (2015) 110{120. doi:10.1016/j.compfluid.2015.02.021. [19] H. Zhang, C. Misbah, Lattice Boltzmann simulation of advection- di usion of chemicals and applications to blood ow, Comput. Fluids 187 (2019) 46{59. doi:10.1016/j.compfluid.2019.04.018. [20] H. Noguchi, G. Gompper, Shape transitions of uid vesicles and red blood cells in capillary ows, Proc. Natl. Acad. Sci. USA 102 (40) (2005) 14159{14164. doi:10.1073/pnas.0504243102. [21] D. A. Fedosov, B. Caswell, G. E. Karniadakis, Systematic coarse- graining of spectrin-level red blood cell models, Comput. Methods Appl. Mech. Eng. 199 (2010) 1937{1948. doi:10.1016/j.cma.2010.02.001. [22] L. Lanotte, J. Mauer, S. Mendez, D. A. Fedosov, J.-M. Fromental, V. Claveria, F. Nicoud, G. Gompper, M. Abkarian, Red cells' dynamic morphologies govern blood shear thinning under microcirculatory ow conditions, Proc. Natl. Acad. Sci. USA 113 (47) (2016) 13289{13294. doi:10.1073/pnas.1608074113. [23] C. Pozrikidis, Boundary integral and singularity methods for lin- earized viscous ow, Cambridge University Press, 1992. doi:10.1017/ CBO9780511624124. [24] E. Lac, J. D. Sherwood, Motion of a drop along the centreline of a capillary in a pressure-driven ow, J. Fluid Mech. 640 (2009) 27{54. doi:10.1017/S0022112009991212. [25] M. Nagel, F. Gallaire, Boundary elements method for micro uidic two- phase ows in shallow channels, Comput. Fluids 107 (2015) 272{284. doi:10.1016/j.compfluid.2014.10.016. [26] A. Joneidi, C. Verhoosel, P. Anderson, Isogeometric boundary integral analysis of drops and inextensible membranes in isoviscous ow, Com- put. Fluids 109 (2015) 49{66. doi:10.1016/j.compfluid.2014.12. 39 [27] J. Gounley, G. Boedec, M. Jaeger, M. Leonetti, In uence of surface viscosity on droplets in shear ow, J. Fluid Mech. 791 (2016) 464{494. doi:10.1017/jfm.2016.39. [28] X.-Q. Hu, A.-V. Salsac, D. Barth es-Biesel, Flow of a spherical capsule in a pore with circular or square cross-section, J. Fluid Mech. 705 (2012) 176{194. doi:10.1017/jfm.2011.462. [29] G. Boedec, M. Leonetti, M. Jaeger, 3D vesicle dynamics simulations with a linearly triangulated surface, J. Comput. Phys. 230 (2011) 1020{ 1034. doi:10.1016/j.jcp.2010.10.021. [30] A. Farutin, T. Biben, C. Misbah, 3D numerical simulations of vesicle and inextensible capsule dynamics, J. Comput. Phys. 275 (2014) 539{ 568. doi:10.1016/j.jcp.2014.07.008. [31] G. Boedec, M. Leonetti, M. Jaeger, Isogeometric FEM-BEM simulations of drop, capsule and vesicle dynamics in Stokes ow, J. Comput. Phys. 342 (2017) 117{138. doi:10.1016/j.jcp.2017.04.024. [32] J. M. Barakat, E. S. G. Shaqfeh, Stokes ow of vesicles in a circular tube, J. Fluid Mech. 851 (2018) 606{635. doi:10.1017/jfm.2018.533. [33] H. Zhao, A. H. G. Isfahani, L. Olson, J. Freund, A spectral boundary integral method for owing blood cells, J. Comput. Phys. 229 (10) (2010) 3726{3744. doi:DOI:10.1016/j.jcp.2010.01.024. [34] S. Ramanujan, C. Pozrikidis, Deformation of liquid capsules enclosed by elastic membranes in simple shear ow: large deformations and the e ect of uid viscosities, J. Fluid Mech. 361 (1998) 117{143. doi:10. 1017/S0022112098008714. [35] O.-Y. Zhong-can, W. Helfrich, Bending energy of vesicle membranes: General expressions for the rst, second, and third variation of the shape energy and applications to spheres and cylinders, Phys. Rev. A 39 (1989) 5280{5288. doi:10.1103/PhysRevA.39.5280. [36] A. Guckenberger, M. P. Schraml, P. G. Chen, M. Leonetti, S. Gekle, On the bending algorithms for soft objects in ows, Comput. Phys. Commun. 207 (2016) 1{23. doi:10.1016/j.cpc.2016.04.018. 40 [37] F. Cirak, M. Ortiz, P. Schr oder, Subdivision surfaces: a new paradigm for thin-shell nite-element analysis, Int. J. Numer. Meth. Eng. 47 (12) (2000) 2039{2072. doi:10.1002/(SICI)1097-0207(20000430) 47:12<2039::AID-NME872>3.0.CO;2-1. [38] T. Hughes, J. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, - nite elements, NURBS, exact geometry and mesh re nement, Com- put. Methods Appl. Mech. Eng. 194 (39) (2005) 4135{4195. doi: 10.1016/j.cma.2004.10.008. [39] C. Loop, Smooth subdivision surfaces based on triangles, Master's the- sis, Department of Mathematics, University of Utah (1987). [40] J. Maestre, J. Pallares, I. Cuesta, M. A. Scott, A 3D isogeometric BE{ FE analysis with dynamic remeshing for the simulation of a deformable particle in shear ows, Comput. Methods Appl. Mech. Eng. 326 (2017) 70{101. doi:10.1016/j.cma.2017.08.003. [41] A. Bartezzaghi, L. Ded e, A. Quarteroni, Biomembrane modeling with isogeometric analysis, Comput. Methods Appl. Mech. Eng. 347 (2019) 103{119. doi:10.1016/j.cma.2018.12.025. [42] J. M. Barakat, S. M. Ahmmed, S. A. Vanapalli, E. S. G. Shaqfeh, Pressure-driven ow of a vesicle through a square microchannel, J. Fluid Mech. 861 (2019) 447{483. doi:10.1017/jfm.2018.887. [43] W. Helfrich, Elastic properties of lipid bilayers: theory and possible experiments, Z. Naturforsch. 28c (1973) 693{703. [44] J. Lyu, P. G. Chen, G. Boedec, M. Leonetti, M. Jaeger, Hybrid continuum{coarse-grained modeling of erythrocytes, C. R. Mec. 346 (2018) 439{448. doi:10.1016/j.crme.2018.04.015. [45] J. Stam, Evaluation of loop subdivision surfaces, SIGGRAPH 99 Course Notes (2001). [46] F. Cirak, Q. Long, Subdivision shells with exact boundary control and non-manifold geometry, Int. J. Numer. Meth. Eng. 88 (9) (2011) 897{ 923. doi:10.1002/nme.3206. 41 [47] J. Walter, A.-V. Salsac, D. Barth es-Biesel, P. Le Tallec, Coupling of nite element and boundary integral methods for a capsule in a stokes ow, Int. J. Numer. Meth. Eng. 83 (7) (2010) 829{850. doi:10.1002/ nme.2859. [48] C. Pozrikidis, Numerical simulation of cell motion in tube ow, Ann. Biomed Eng. 33 (2) (2005) 165{178. doi:10.1007/ s10439-005-8975-6. [49] E. Fehlberg, Low-order classical Runge-Kutta formulas with stepsize control and their application to some heat transfer problems, Tech. rep., NASA (1969). [50] D. A. Knoll, D. E. Keyes, Jacobian-free Newton{Krylov methods: a survey of approaches and applications, J. Comput. Phys. 193 (2) (2004) 357{397. doi:10.1016/j.jcp.2003.08.010. [51] C.-S. Yih, Fluid Mechanics, West River Press, 1979. [52] S. Kuriakose, P. Dimitrakopoulos, Motion of an elastic capsule in a square micro uidic channel, Phys. Rev. E 84 (2011) 011906. doi:10. 1103/PhysRevE.84.011906. [53] N. Liron, R. Shahar, Stokes ow due to a Stokeslet in a pipe, J. Fluid Mech. 86 (1978) 727{744. doi:10.1017/S0022112078001366. [54] H. Brenner, Pressure drop due to the motion of neutrally buoyant par- ticles in duct ows, J. Fluid Mech. 43 (1970) 641{660. doi:10.1017/ S0022112070002641. [55] G. Hetsroni, S. Haber, Wacholder, The ow elds in and around a droplet moving axially within a tube, J. Fluid Mech. 41 (4) (1970) 689{ 705. doi:10.1017/S0022112070000848. [56] X.-Q. Hu, B. S ev eni e, A.-V. Salsac, E. Leclerc, D. Barth es-Biesel, Char- acterizing the membrane properties of capsules owing in a square- section micro uidic channel: E ects of the membrane constitutive law, Phys. Rev. E 87 (2013) 063008. doi:10.1103/PhysRevE.87.063008. [57] A. Farutin, C. Misbah, Symmetry breaking and cross-streamline migra- tion of three-dimensional vesicles in an axial Poiseuille ow, Phys. Rev. E 89 (2014) 042709. doi:10.1103/PhysRevE.89.042709. 42 [58] D. Agarwal, G. Biros, Stable shapes of three-dimensional vesicles in uncon ned and con ned Poiseuille ow, Phys. Rev. Fluids 5 (1) (2020) 013603. doi:10.1103/PhysRevFluids.5.013603. [59] E. A. Evans, R. Skalak, Mechanics and thermodynamics of biomem- branes, CRC Press, 1980.
Journal
Physics – arXiv (Cornell University)
Published: Feb 12, 2020