Simulating Copolymeric Nanoparticle Assembly in the Co-solvent Method: How Mixing Rates Control Final Particle Sizes and Morphologies

Simon Keßler and Friederike Schmid Institut f¨urPhysik, Johannes Gutenberg-Universit¨atMainz, D 55099 Mainz

Klaus Drese Fakult¨atf¨urAngewandte Naturwissenschaften, Hochschule Coburg, 96450 Coburg

The self-assembly of copolymeric vesicles and micelles in micromixers is studied by External Potential Dynamics (EPD) simulations – a dynamic density functional approach that explicitly ac- counts for the polymer architecture both at the level of thermodynamics and dynamics. Specifically, we focus on the co-solvent method, where nanoparticle precipitation is triggered by mixing a poor co-solvent into a homogeneous copolymer solution in a micromixer. Experimentally, it has been reported that the flow rate in the micromixers influences the size of the resulting particles as well as their morphology: At small flow rates, vesicles dominate; with increasing flow rate, more and more micelles form, and the size of the particles decreases. Our simulation model is based on the assump- tion that the flow rate mainly sets the rate of mixing of solvent and co-solvent. The simulations reproduce the experimental observations at an almost quantitative level and provide insight into the underlying physical mechanisms: First, they confirm an earlier conjecture according to which the size control takes place in the earliest stage of the particle self-assembly, during the spinodal decomposition of polymers and solvent. Second, they reveal a crossover between different morpho- logical regimes as a function of mixing rate. Hence they demonstrate that varying the mixing rate in a co-solvent setup is an effective way to control two key properties of drug delivery systems, their mean size and their morphology.

I. INTRODUCTION an initial solution containing the molecular constituents of the nanoparticles (e.g., the copolymers) in good sol- Nanoparticles are molecule aggregates with spatial ex- vent is mixed with a selective or bad solvent. The mixing tensions on scales of ∼ 10 to 100 nm. They have at- eventually triggers the precipitation of the constituents tracted growing interest during the last decades due to and the aggregation to particles. One way to tune the improved technologies for visualization and manipulation sizes of the particles is to vary the rate of solvent mixing of nanoscale structures that have revealed their great po- [12, 13, 17, 18]. In experiments, solvent mixing is often tential for a wide range of applications. Depending on implemented by continuous flow mixing devices, called the chemical composition of the nanoparticles, these ap- passive micromixers [12, 13, 17, 18]. Mixing rates in plications include mesoscopic models for atomic systems passive micromixers increase with increasing flow rates [1, 2], optoelectronic devices [3], nanoreactors, models v [19–21], and in turn, particle sizes are found to de- for biological cells [4, 5], and most prominently, trans- crease with increasing v. Thiermann et al. [12, 13] have port vehicles for medication, i.e. drug delivery systems recently carried out an extensive study on the relation [6–8]. between flow rates and particle size in such a micromixer Most commercially available drug delivery systems are setup. The constituents of their nanoparticles were am- liposomes [9, 10], which are vesicular particles made of phiphilic diblock-copolymers with a hydrophobic polybu- amphiphilic lipids. Biocompatible amphiphilic (diblock) tadiene block and a hydrophilic polyethylene-oxide block, copolymers represent promising alternatives to the lipids, and they used tetrahydrofuran as good solvent and wa- since they are more stable and can be synthesized and ter as selective co-solvent. They found that variations of modified more easily [11–13]. In analogy to their lipid v for symmetric flow conditions in passive micromixers counterparts, vesicles built from amphiphilic diblock-co- enable a reproducible control over the mean size R of polymers are often called polymersomes [11]. In the a particle population with relatively low polydispersity. context of drug delivery systems one key property of The experimental data for R could be described reason- α nanoparticles is their size, since it not only determines ably well by scaling laws R ∝ v , where the exponent their loading capacity, but also the composition of their α differs from measurement to measurement, but scat- protein corona in blood [14], which in turn affects reten- ters around a mean value of α = −0.158 with a standard arXiv:1804.02875v1 [cond-mat.soft] 9 Apr 2018 tion times in the circulatory system. In addition, the deviation σα = 0.058 [22]. nanoparticle size plays a critical role in passive target- In a recent publication we proposed an explanation ing of tumors based on the Enhanced Permeability and for the experimentally observed nanoparticle size de- Retention effect [15]. pendence on flow rates or equivalently, mixing times A method to manufacture nanoparticle populations [23]. We combined a simple Cahn-Hilliard equation with with a specific mean size is the co-solvent method, also a Flory-Huggins-de Gennes free energy functional for known as flash nanoprecipitation [16]. In this method, homopolymers and implemented solvent mixing by a time 2 dependent interaction parameter. The simulation results rates, vesicles dominate, while at larger flow rates, mi- for the size of homopolymer aggregates at the (relatively celles become more frequent. Motivated by these obser- well-defined) crossover between spinodal decomposition vations, we here present a numerical study of copoly- and subsequent coarsening were in good semiquantita- mer aggregation in a co-solvent setup for different mix- tive agreement with the experimentally determined sizes ing rates, using a dynamic density functional approach for stabilized copolymer particles from Thiermann et al that explicitly accounts for the molecular architecture of [12, 13] (apart from a factor of two). This lead us to copolymers. To this end, we implement time dependent hypothesize that particle sizes in the co-solvent method interaction parameters into the established ’External Po- are determined during the very early stages of phase sep- tential Dynamics” (EPD) model [28], which has been aration. The description also provided an explanation successfully used to study spontaneous self-assembly of for the typical scaling laws observed in experiments and amphiphilic diblock-copolymers to nanoparticles [29, 30]. predicted an analytical expression, R ∼ v−α with the ex- We focus specifically on the effect of mixing rates on ponent α = 1/6, which is in good agreement with the particle morphologies, and how the existence of different experimentally observed exponents. According to this particle morphologies affects the typical scaling behavior theory, the particle sizes at different v result from a com- R ∝ vα. petition of the repulsion between solvent molecules and The article is organized as follows. In section II we monomers of the solvent-phobic block with the interfa- present the theoretical model. In section III we specify cial tension of diffuse interfaces in the very early stages of the input parameters used in the current article. The phase separation, during spinodal decomposition. With results and the discussion is presented in section IV, and respect to the entire particle growth process the descrip- the summary is given in V. In the A, we discuss technical tion in terms of the Cahn-Hilliard model is, however, issues and describe a novel integration scheme which was incomplete. The Flory-Huggins-de Gennes free energy used to perform the simulations. functional for homopolymers can only be applied in the early segregation regime, it does not describe the inter- nal reorganization of copolymers inside the particles at later stages. In particular, it does not include a mecha- II. THEORETICAL MODEL nism that physically stabilizes particles of finite size, and cannot be used to describe particles with more complex We consider a solution of an amphiphilic AB-diblock morphologies as can be observed in copolymeric systems. copolymer P and a single solvent S in a volume V at tem- Amphiphilic molecules may form particles of different perature T . To describe the phase separation dynamics morphologies in solution, including spherical micelles or we apply the EPD model [28, 29, 31], which is based on vesicles [5]. Apart from the size, particle morphology is the free energy functional of the popular Self Consistent another key property of drug delivery systems because it Field (SCF) Theory for polymers [32, 33]: affects the loading possibilities. Spherical micelles con-     Z sist of a hydrophobic core surrounded by a shell of hy- β F QS fP QP N 1 h = − fS ln − ln + − ωAφA drophilic blocks and they can only be loaded with hy- n V fS N V fP V V drophobic substances. Vesicles allow both hydrophilic − ωBφB − ωSφS + χABφAφB + χASφAφS and hydrophobic loading, which makes them appealing κ i canditates for multifunctional drug delivery systems: Hy- + χ φ φ + H (φ + φ + φ − 1)2 d~r. BS B S 2 A B S drophilic substances can be enclosed in their solvent con- (1) taining core, while the hydrophobic part of the bilayer Here fP and fS are the mean polymer and solvent vol- shell can be loaded with hydrophobic substances. These ume fractions, N is the number of monomers per polymer loading possibilities are important because some thera- chain, κH a mean compressive modulus of the solution peutic substances are hydrophilic and others hydropho- [33, 34], χij the Flory-Huggins interaction parameter be- bic. An example for a hydrophilic substance is the toxic tween species i and j, and β = 1/kBT the Boltzmann anti-cancer drug camptothecin [24]. Hydrophilic mate- factor. The fields ωi for i = A, B, S are potentials in rials are, for instance, the anti-cancer drug doxorubicin −1 units of β = kBT that act on the respective monomer [25] or the dye pholoxine B, which can be used to trace ρ n species i, and φi = i/ρ0 with ρ0 = /V are normalized particles in in vitro cell binding studies or studies on (hy- number densities, where n = nS + NnP is the total num- drophilic) loading efficiencies [26]. Although vesicles are ber of monomers and solvent molecules in the system. more versatile when it comes to loading possibilities, mi- nS is the number of solvent molecules and nP the num- celles have the advantage that, due to their smaller size ber of polymer chains. QP is the partition function of compared to vesicles, they may enable ways of cellular a polymer chain subject to ωA and ωB, while QS is the uptake that bypass the drug efflux mechanism of cancer partition function of a single solvent molecule exposed to cells in order to treat multiresistant cancer [10, 27]. ωS. These single chain partition functions are given by The experiments by Thiermann et al. [12, 13] clearly Z Z show that the flow rates in micromixers also affect the −ωS (~r) QP = g(~r, 1) d~r and QS = e d~r. (2) morphologies of the resulting particles. At small flow V V 3 g(~r, s) is the end-segment distribution function and de- the density-field relations from the SCF Theory and that scribes the probability that one end of a polymer chain QP = QP [ωA, ωB] as well as QS = QS[ω], the chemical segment of length s is located at position ~r. It obeys the potentials µi can be calculated as inhomogeneous diffusion equation µA = χABφB + χASφS + κH (φA + φB + φS − 1) − ωA, ∂g (~r, s) = ∆g(~r, s) − Nωg(~r, s) with g(~r, 0) = 1 (3) (10) ∂s [35, 36], where omega is defined by µB = χABφA+χBSφS +κH (φA+φB +φS −1)−ωB, (11)  ωA 0 < s < cA ω = . (4) ωB cA < s < 1 µS = χASφA +χBSφB +κH (φA +φB +φS −1)−ωS. (12) ~r denotes the position in units of the polymer’s radius of gyration Rg, s ∈ [0, 1] represents the position along a Equations 3, 6 – 8, and 9 constitute a SCF-EPD model polymer chain in units of N, and cA specifies the frac- that has been used to successfully study self-assembly of tion of A-monomers in the diblock copolymer. If the particles with various morphologies [29, 30]. polymer chain contains NA monomers of type A, one has In our previous publication we have shown that spin- N cA = A/N. Since the potential fields in the SCF The- odal decomposition under time-dependent quenches into ory mimic mean interactions between different species, the spinodal area reproduce experimental trends [23]. they depend on the densities. Introducing a second dis- Therefore, solvent mixing in the present work is described tribution function g0, which also solves equation (3) with in a very similar manner. We model solvent mixing by a time dependent interaction parameter between the sol-  vent and the B-block, which is from now on considered ωB 0 < s < cB ω = (5) to be the solvent-phobic one. If not stated otherwise, its ωA cB < s < 1 time dependence is linear and given by

N and cB = B/N = 1−cA, the relation between the poten- ( (0) χBS + sχt, t ≤ tmax tial fields ωi and the densities φi from the SCF Theory χ (t) = . (13) BS (max) can be cast into the form χBS , t > tmax

Z cA   V fP 0 χ(max)−χ(0) φA(~r)[ωA] = g(~r, s)g (~r, 1 − s) ds, (6) The cutoff time tmax = BS BS /sχ can be inter- QP 0 preted as a mixing time, and the parameter

(0) 1 1 cA Z 1 χ = + +χABcA −χAS (14) V fP 0 BS 2Nc f 2c (1 − f ) c φB(~r)[ωB] = g(~r, s)g (~r, 1 − s) ds, (7) B P B P B QP cA is the spinodal solvent-phobic interaction for which and 2 2 ∂ FFH /∂fP = 0, where FFH is the Flory-Huggins ap- V f proximation to F from equation (1) [29]. S −ωS (~r) φS(~r)[ωS] = e . (8) QS In the EPD formalism the dynamical equations for the III. SIMULATION SETUP potential fields are given by Simulations start from randomly perturbed homoge- ∂ωi neous initial states. All simulation results are averaged (~r, t) = −Di∆ (µi(~r, t) + ηi(~r, t)) with i = A, B, S, ∂t over five simulation runs with different initial conditions.  D 0 We consider a volume fraction fP = 0.1 of a model poly- 1 δβF DP = i = A, B µi(~r) = , and Di = N , mer with a solvent-philic block length NA = 3 and an ρ0 δφi(~r) DS = D0 i = S incompatible solvent-phobic block containing NB = 14 (9) monomers. The incompatibility is described by an inter- action parameter χAB = 1.05, and to keep φA + φB + φS δβF where ηi is a random noise and is the variational close to 1, the compressive modulus is set to κH = 100. δφi derivative of βF with respect to φi. Equation 9 is equiv- The mean volume fraction of selective solvent is fS = alent to the dynamical master equation for monomer den- 1 − fP = 0.9. The diffusion coefficient D0 in equation sitiy fields derived by Kawasaki and Sekimoto [37] with a (9) can be substituted by 1 without loss of generality as non-local kinetic coupling in a copolymer solution. The lengths are given in units of l0 = Rg and times in units R2 EPD formalism dramatically reduces the computational of t0 = g/D0. The solvent-philic interaction is kept at a cost of the non-local coupling model and was first intro- constant value χAS = −0.15, and the solvent-phobic one (0) (Spin) duced by Maurits and Fraaije [28]. Taking into account is varied from its spinodal value χBS = χBS = 1.249 to 4

(max) χBS = 2.25, which corresponds to a maximum quench 2 (a) and (b) provide a direct comparison of identically depth of approximately 1 like in reference [23]. As in prepared polymer solution samples for two different flow reference [23], the random noise is turned off, i.e. ηi = 0. rates [12] in the Caterpillar Micromixer [21], i.e. for two The number of spatial grid points is set to m = different mixing rates. The particle morphologies resem- 256 × 256 with a lattice constant ∆l = 0.25Rg. The ble the simulation results from figure 1 (a) and (c), except time step h varies during the simulation as described in that the TEM images lack cylindrical micelles while the A. Unless stated otherwise, we use an initial (and maxi- simulation results contain a few. In other experimental mum) time step of h = 0.05t0. The segment length in a work [17], such cylindrical micelles have also been found polymer chain is ds = 1/N. Simulations are restricted to to coexist with spherical micelles and vesicles. Whether two dimensions (2D) because the qualitative size depen- or not cylindrical micelles appear in simulations likely de- dence was shown to be independent of the dimension in pends on the interaction parameters χAB and χAS. We [23]. In 2D, much larger systems can be simulated over conclude that qualitatively, the dependence of morpholo- longer time periods. gies on mixing rates is in good agreement with exper- The set of equations (3) and (6 – 8) is solved with the iments. Furthermore, the simulations indicate that the same solvers as in Ref. [29]. The evolution equations for enrichment of micelles only marks the onset of a complete the potential fields (9) is solved by a newly developed morphological transition from vesicular to micellar. semi-implicit integration scheme, which is presented in A. B. Minkowski measures and determination of particle sizes IV. RESULTS AND DISCUSSION The snapshots in figure 1 are taken at the so-called A. Morphological transition and qualitative transition time ttr [42]. In the Cahn-Hilliard model, the discussion of particle size dependencies on mixing transition time separates a regime of spinodal decompo- speeds sition and initial droplet nucleation and growth, where polymer aggregates form and concentrate, from a com- Figure 1 depicts simulated polymer particles for differ- paratively slow ripening regime, where small aggregates ent mixing speeds sχ. Strictly speaking, only the density grow at the expense of smaller ones until only one single of solvent-phobic B-monomers φB is shown, but since polymer aggregate remains. The transition time ttr can the A-monomers accumulate approximately at green to be determined by inspecting, for instance, the Minkowski yellow φB-values, these colors can be imagined to rep- measure C [42]. Here C denotes the total boundary resent the solvent-philic A-block. It is evident that an length of the union over all subsets where φB exceeds (th) increase of the mixing rate leads to a decrease of the a predefined threshold φB . The rapid concentration of typical particle size and induces a morphological vesicle- solvent-phobic monomers in the spinodal decomposition to-micelle transition: From figure 1 (a) to (d), the num- stage leads to a very fast temporal increase of C, while ber of vesicles decreases while the number of micelles in- ripening is characterized by a slow but steady decrease creases until only spherical micelles are left. For every of C. This opposite behavior leads to a clear maximum sχ, the φB-profiles in the early stages of phase separa- in C(t), which marks the transition time. Time series of tion (not shown) closely resemble the profiles obtained C in the SCF-EPD copolymer model are shown in figure in our earlier work based on the Cahn-Hilliard equa- 3. They look very similar to the Cahn-Hilliard model in tion for homopolymers [23]. Once the polymer con- our earlier studies [23]. We still find the distinct maxi- tent inside polymer aggregates is sufficiently high, the mum that allows an analogous definition of the transition block incompatibility leads to an internal rearrangement time ttr. As in the Cahn-Hilliard model, the steep rise of copolymer chains inside the aggregates, which even- at t < ttr is caused by a rapid initial polymer aggre- tually leads to the formation of different morphologies. gation, which is already associated with the formation In the literature, different pathways of vesicle formation of different morphologies (see figure 1). The subsequent have been described theoretically and observed experi- slow decrease of C(t) is not caused by Ostwald Ripening, mentally [29, 30, 38–41]. In the present simulations, they [? ] but mainly by a shrinking of particles. In the simu- form by nucleation and growth (mechanism II according lations from figure 1, this shrinking is most pronounced −6 to Ref. [30]). The solvent-philic A-block, ultimately ori- at sχ = 4.4×10 . To give an impression of the extent of ented towards the solvent, sterically stabilizes the parti- the shrinking, figure 3 (a) shows the density profiles from cles by suppressing further ripening (where large parti- figure 1 (a) at the end of the simulation run, t = tend. cles grow and smaller particles dissolve) and preventing The shrinking and the increase of φB-values between ttr macrophase separation [29]. and tend is most likely caused by the increase of χBS dur- An enrichment of micelles with increasing flow rate ing this time interval. Finally, C(t)reaches a plateau at comparable to figure 1 (a) to (c) is also observed ex- late times (see figure 3 (b) and (c)). This is in contrast to perimentally. Transmission Electron Microscopy (TEM) the Cahn-Hilliard model, where C(t) continues to decay images from experiments are shown in figure 2. Figure at late times. The plateau corresponds to a state where 5

−6 Figure 1. Color coded solvent-phobic density profiles φB . (a) shows simulation results at a mixing rate of sχ = 4.4 × 10 (b) −5 −5 −3 at sχ = 4.4 × 10 , (c) at sχ = 8.8 × 10 , and (d) at sχ = 4.4 × 10 . An increase of mixing rates does not only affect particle sizes but also induces a morphological vesicle-to-micelle transition. multiple stabilized particles are present. are estimated by the mean values

We measure particle sizes at transition time because p p 1 X 1 X the shrinking is not very pronounced, and because a max- R = R and R = R . (16) imum of C can be defined more precisely than a plateau. s p s,i v p v,i i=1 i=1 To this end, pictures such as those shown in figure 1 are converted into binary images with a threshold value of As a measure for polydispersity of a nanoparticle popu- (th) lation we use the standard deviation φB . Then a standard 4-connected-component image labeling algorithm [43] is used to count and isolate sin- v u p gle particles. Afterwards, the area Ai and circumference u1 X 2 ∆Rj = t (Rj − Rj,i) . (17) Ui is determined for every single particle i = 1, ..., p in p a picture with the Minkowski functional algorithm from i=1 Mantz et al. [44]. We define the sphere equivalent radius for j = s, v. Rs,i and the vesicle equivalent radius Rv,i of particle i by

C. Simulation results for particle sizes and r Ai Ai Ui morphological regimes Rs,i = and Rv,i = + , (15) π Ui 4π Figure 4 shows simulation results for particle sizes and respectively. The vesicle equivalent radius is the outer ra- transition times[? ]. Particle sizes are given in units dius of the spherical shell with area Ai and total perime- of the polymer’s radius of gyration l0 = Rg and times R2 ter Ui. In case a spherical micelle is processed, Rs,i and in units of t0 = g/D0. The dependence of particle Rv,i are equal to its radius R, which can be verified by sizes and transition times on quench rates also resem- 2 insertion of Ai = πR and Ui = 2πR. Mean particle sizes bles the results from the Cahn-Hilliard model for ho- mopolymers: There is an asymptotic regime, where par- ticle sizes converge to results for a constant quench depth (max) χBS(t) ≡ χBS , and there is a non-asymptotic regime, α where particle sizes follow scaling laws R ∝ sχ with α ≈ −1/6, while transition times can be approximated −2/3 by ttr ∝ sχ [23]. As in reference [23], the asymptotic and the non-asymptotic regime are separated by the time when tmax in figure 4 intersects ttr. The similarity of the data curves in figure 4 to our earlier Cahn-Hilliard results for homopolymers [23] confirms the hypothesis that par- ticle sizes are determined during the very early stages of phase separation, where the specific sequence of polymers Figure 2. Transmission Electron Microscopy images of cross- (block copolymers vs. homopolymers) is not yet impor- linked nanoparticles made of polybutadiene-polythyleneoxide tant. In particular, the predictions of the simple Cahn- diblock copolymers obtained in the Caterpillar Micromixer Hilliard model [23] can still be applied. at two different flow rates as indicated and otherwise iden- For diblock-copolymers, however, the non-asymptotic tical experimental parameters under symmetric flow condi- regime splits into three different morphological regimes. tions. After Ref. [12] (samples Cd10 and Cd11), courtesy of We call a morphological regime an interval of mixing R. Thiermann. rates with specific predominant particle morphologies. In 6

−6 −3 Figure 3. Vesicles at tend > ttr (a) and time series of the Minkowski measure C(t) for sχ = 4.4 × 10 (b) and sχ = 4.4 × 10 (c). The transition time ttr is labeled by a circle. tend denotes the end of a simulation run and is marked by a rectangle. The insets in (b) and (c) show a magnification of the respective part framed by the rectangle.

figure 4 (a) the morphological regimes are separated by throughout the system. Specifically, we assume the vertical lines. At low mixing rates only vesicles form. Figure 1 (a) shows a corresponding snapshot of parti- χBSS − χBGS χBS(t) = χBGS + ϕSS(t), (18) cles. At intermediate mixing rates one gets a mixture 1 − fP of vesicles, cylindrical micelles and spherical micelles. In the intermediate regime the number of vesicles decreases with increasing mixing rates while the number of micelles increases. Corresponding snapshots of particles can be found in figure 1 (b) and (c). At large mixing rates there are only micelles as seen in figure 1 (d). Because every morphological regime covers a certain interval of particle sizes, particle morphologies are directly coupled to their size. With respect to the co-solvent method this means that it should be possible to produce pure vesicular or mi- cellar populations at ’extremely’ large or low flow rates, but intermediate ones typically result in heterogeneous morphologies. As long as the micelles are not exclusively spherical, Rv is larger than Rs because the formation of vesicles from a polymer aggregate increases the perime- ter Ui of a spherical shell, while it keeps the volume Ai constant. It should be noted though that the simulated ”mi- celles” from the rightmost morphological regime in fig- ure 4 (a) are, strictly speaking, actually no ’real’ mi- celles. Real micelles are equilibrium structures with a well-defined size distribution and composition that does not depend on the history of the system. Here, we consider ’micelle-like’ spheres with the typical core-shell structure that would be characteristic for a micelle, but they are not equilibrated. Figure 4. Simulation results for particle sizes (a) and transi- tion times (b) in dependence on the quench rate sχ. The data points in (a) represent the simulation results for the sphere D. Comparison of particle sizes to experimental and vesicle equivalent radius from equation (16) and the er- results ror bars are the standard deviation from equation (17). Trend line equations are given in the legend. The vertical lines sep- arate three different morphological regimes, where the corre- After having studied the idealized situation where the sponding morphologies above the dashed line appear. Empty solvent mixing is described by the simple linear increase circles symbolize vesicles, solid black dots spherical micelles of χBS(t) with time, (Eq. (13)), we will now introduce and the dumbbell cylindrical micelles. The data points in (b) a more general Ansatz which allows us to compare sim- are the transition times ttr (cp. figure 3) and the dashed line ulation results directly to experiments: We assume that is tmax from equation (13). The trend line equation refers to χBS(t) is a linear function of the volume fraction of se- the transition time. Error bars are standard deviations over lective solvent ϕSS(t), which we take to be homogeneous five different simulation runs. 7 where χBGS is the interaction parameter between the B- the simulations, micelles dominate already at flow rates monomers and the good solvent, while χBSS describes above approximately 0.8 ml/min. The specific flow rate the interaction of the B-monomers and the selective sol- values of the vesicle to micelle transition as well as the vent. The Ansatz ensures that χBS = χBGS if no selec- final particle sizes are likely to depend strongly on χAB tive solvent is present and χBS = χBSS if only selective and χAS. Since the material parameters of the exper- solvent is present, i.e. if ϕSS = 1 − fP . imental systems studied by Thiermann et al. [12, 13] We assume that ϕSS(t) can be calculated indepen- were not determined, a direct quantitative comparison dently of the particle self-assembly simply by consider- between model and experiments is difficult. ing the mixing of simple fluids in a given micromixer Nevertheless, our results show that the simulations re- geometry. This Ansatz allows a direct coupling of mixer produce the experimental trends and the order of magni- geometries and flow rates into the description of parti- tude of simulated particle sizes matches the experimen- cle growth. Here we specifically consider the case of the tally observed sizes. This suggests that the main mecha- Caterpillar Micromixer (CPMM), where analytical ex- nism determining particle sizes is captured by our mean pressions for ϕSS(t) have been derived by Schoenfeld et field model, even though it does not account for Brownian al. [21]. Simulation results for particle sizes with the mix- motion of particles and thereby induced collision induced ing profile ϕSS(t) from the Caterpillar micromixer are coagulation and growth of particles. shown in figure 5. Even with this Ansatz, the data still To conclude, the results show that the combination show scaling behavior. of established methods for calculating solvent mixing in If we normalize the experimental data with 15 nm, complex geometries on large scales with mean field the- i.e., we assume the gyration radius of the polymers to be ories for aggregation on mesoscales, coupled through an around Rg =15 nm, we get almost quantitative agree- Ansatz like equation (18), represent a promising multi- ment between simulation data and experiments. In re- scale approach for describing flow controlled particle as- ality, the radius of gyration is probably smaller, in the sembly in micromixers. range of 5-10 nm [22, 23], hence the simulations prob- ably underestimate the particle size. The simulations also reproduce the experimentally observed morphologi- V. SUMMARY AND OUTLOOK cal transition from vesicles to micelles upon reducing the flow rate, but here again, the absolute values of flow rates We have implemented time dependent interaction pa- where the morphological transition is observed in exper- rameters into a Dynamic Self Consistent Field Theory for iments and simulations do not match quantitatively. In copolymer simulations in order to describe the nonequi- the experiments pure vesicular populations are observed librium assembly of amphiphilic diblock-copolymers in up to flow rates of approximately 1.8 ml/min, while in micromixers. Experimental observations show that the final mor- phologies of particles are mostly vesicular at low mix- ing rates, and mostly micellar at high mixing rates [12, 13, 17]. Our simulations indicate that these changes in morphologies are the signature of an incomplete vesicle- to-micelle transition, and that nanoparticle populations should be purely micellar if the flow rates are sufficiently large. We conclude that it should be possible to produce both pure vesicular and pure micellar nanoparticle popu- lations with the co-solvent method. In the morphological regime at low flow rates, one can tune the size of vesi- cles, while in the regime at high flow rates the size of micelles can be adapted. The intermediate regime, how- ever, excludes certain mean sizes if one wants to produce particles with only one morphology. The rate of solvent mixing qualitatively affects parti- Figure 5. Particle sizes in units of Rg versus flow rates v cle sizes in the same way in the SCF-EPD model as in in ml/min in the Caterpillar Micromixer. The open sym- bols represent simulation results and the solid symbols exper- the Cahn Hilliard model for homopolymers. The present imental results from Thiermann et al. [13]. A ( ˆ=H) and work hence confirms our conjecture in Ref. [23], that the B ( ˆ=CO–CH2–CH2–COOH ) denote different end groups at- sizes of particles that are aggregated from amphiphilic tached to the polymer. Trend line equations are shown be- copolymer solutions are determined during the very early neath the diagram at the corresponding symbols. For the sake stages of phase separation by a competition between the of simplicity only the regression line to Rs is shown (thick light interfacial tension of diffuse interfaces and the decreas- grey line). Error bars mark the polydispersity as determined ing solvent quality for the solvent-phobic block. Further- from Eq. 17 and are again only shown for Rs for the sake of more, figure 4 shows that the rate-dependent morpho- clarity. ∆Rv looks very similar. logical changes do not affect the typical scaling behavior 8

∝ −1/6 R ∼ v in the non-asymptotic regime, which is ob- the future, it will be particularly challenging to use our served here in copolymer solutions as well as in Ref. [23] methods to model the self-assembly of complex struc- in homopolymer solutions. tured nanoparticles. As we have mentioned above, vesicles form via a We should however, also point out some limitations of nucleation-and-growth pathway in the simulations con- the approach. The main advantage of using mean field sidered here (pathway II in Ref. [29]), which is typical continuum theories like the SCF-EPD model rather than for the aggregation of particles from solutions with rel- detailed particle models is the possibility to simulate self- atively low copolymer densities and/or relatively weak assembly on realistic mixing times in micromixers. This incompatibilities [30]. At higher copolymer concentra- comes at the prize that the simulations lack detail on tions [30, 41], one observes an alternative aggregation- a molecular level. The SCF-EPD model makes a local and-bending pathway (pathway I), where micellar parti- equilibrium assumption and thus cannot capture situa- cles first aggregate to platelets and these platelets then tions where chains freeze or crystallize partially and chain bend around, driven by a competition of bending energy conformations can no longer equilibrate. The scaling law −1/6 and line tension, to form closed vesicles. In future work, R ∝ sχ is typical for liquid-liquid demixing as par- it will be interesting to study the influence of the mixing ticle sizes are reported to depend differently on mixing rate on particle aggregation in the regime where path- rates when particle growth interferes with solidification way I is dominant. (These simulations would have to be [46]. Hence the SCF-EPD model should only be applied done in three dimensions, since in two dimensions, the if particles initially form during liquid-liquid demixing. final step of platelet bending does not take place). In Furthermore, so far, we have not investigated the direct particular, it will be interesting to investigate whether effect of flow on self-assembly. Although the size control ∝ −1/6 the scaling law for the particle sizes, R ∼ sχ , still per- appears to be dominated by solvent mixing rates, flow sists in that regime. If the scaling law is destroyed, this effects might still affect the particle size dependence on would provide indirect evidence that the vesicle forma- flow rates to some degree. This is particularly relevant tion in the micromixers that have studied experimentally in shear flows at high shear rates. In a recent study it proceeds predominantly via mechanism II. was found that shear flow in micro channels may delay To implement mixer geometries into solvent mixing, we nucleation [45] and thus, growing shear rates could di- have coupled an established description for the Caterpil- rectly increase the transition times ttr. Furthermore, it lar Micromixer (CPMM) [21] into the SCF-EPD model. was found that strong shear flows may induce irreversible Here, the interaction between the solvent and the solvent- changes in the final shapes of particles. Studying the ef- phobic monomers was assumed to be a linear function fect of solvent mixing in the presence of shear flow by of the volume fraction of selective solvent (cf. Eq. 18). coupling solvent mixing into the hybrid EPD-LB model This Ansatz allowed us to perform simulations of parti- by Heuser et al. [45] will be an interesting topic for fu- cle assembly with realistic mixing times and to compare ture work as well. If the shear rate is so high that it directly the particle sizes obtained in experiments and becomes comparable to the inverse Zimm time of poly- simulations. The scale of the simulated particle sizes mers, polymers will deform, which will further influence matches experimentally determined particle sizes, and the self-assembly. However, in micromixers, such high ∝ −1/6 shear rates are usually not applied. the scaling behavior R ∼ v was also reproduced, which implies that the mean field theories are suited to capture the self-assembly process in the co-solvent method. ACKNOWLEDGMENTS More generally, our Ansatz represents a computation- ally very efficient multiscale approach to describe the We thank R. Thiermann for allowing us to reproduce nanoparticle formation in mixer geometries on millime- the experimental images in figure 2 from his thesis, [12]. ter or centimeter scales over realistic mixing times. It is We acknowledge funding by EFRE (Europ¨aischer Fonds not restricted to the CPMM geometry, other micromix- f¨urregionale Entwicklung) and by the German Science ers can be implemented as well - even though, in most Foundation within the Collaborative Research Center cases, the resulting mixing behavior would have to be cal- TRR 146. The simulations were carried out on the high culated numerically if approximate analytical solutions performance computing center MOGON at JGU Mainz. are not available. The mesoscale model can also be re- FS wishes to thank Axel H.E. M¨ullerand Andr´eGr¨oschel fined. For example, it can can be extended such that it for their inspiring work and enjoyable discussions on com- allows for spatial variations of the selective solvent vol- plex nanoparticle assembly. ume fraction φSS(t) [22]. In the present work, we have assumed φSS = ϕSS to be constant everywhere; in re- ality, one would expect that the selective solvent prefer- Appendix A: NUMERICAL INTEGRATION ably accumulates outside of the particles, and this may SCHEME influence the particle assembly. Extending the model to more complex copolymer architectures, and to copolymer The value of the compressive modulus κH affects the mixtures, is also possible and quite straightforward. In number, the mean size, and the polymer content of sim- 9

Figure 6. Color coded two-dimensional solvent-phobic density profiles φB of stable micelles for different compressive moduli 1 κH = 1.176 (a), κH = 5 (b) and κH = 100 (c). Simulations were performed on a 128 × 128 grid with a lattice constant ∆l = /3 and a fixed time step of h = 0.002. The values of the remaining parameters are χBS = 1.6, fP = 0.1, NA = 2, NB = 15, 1 χAB = 1.05, χAS = 0.0375, DP = /N, and DS = 1. All snapshots are taken at time step n = 200000.

ulated nanoparticles as shown in figure 6. bilities at high values of κH , but they come at the cost Polymer solutions typically possess a liquid-like com- of a dramatic increase in computation times. To avoid pressive modulus [47], so simulations should be per- this problem while still enlarging the κH -region where the the algorithm runs stably, we have developed a semi- formed at high κH . If the chemical potentials µi from equations 10 – 12 are inserted into the dynamical equa- implicit integrator for equation (9) that does not require (n+1) tions 9, κH appears as a prefactor of Laplacian terms the computation of φi . The main equations are sum- ∆φi. Hence, the stiffness of the dynamical equations in- marized in figure 7 (equations (A1–A3)), the derivation creases with κH . Stiffness of differential equations can will be shown below. In this scheme, the potential fields (n+1) be tackled by applying implicit or semi-implicit numeri- ωi (~r) are decoupled from the other unkown variables cal integration schemes [38, 48–50]. (n+1) (n+1) 0(n+1) φi (~r), g (~r, s), and g (~r, s). This decoupling A standard approach [48–50] to cope with the allows one to selectively deal with κH -induced stiffness of stiffness caused by high κH in equation (9) is to the dynamical equations while keeping the computational carry out a direct implicit quadrature of the term effort per time step as close to efficient explicit schemes R tn+1 as possible. In other words, it guarantees that, in partic- t κH (∆φA(t) + ∆φB(t) + ∆φS(t)) dt upon integrat- n ular, g(n+1)(~r, s) and g0(n+1)(~r, s) can still be calculated ing the equation over a time step h = tn+1 − tn. Hence, the right hand side of the time discrete version of equa- by cost-efficient explicit schemes, whereas the iterative (n+1) methods are only applied to a 3m-dimensional system tion (9) depends on φi , which in turn depends on (n+1) of equations for ωi(~r, tn+1). At κH = 100, our semi- ωi via the (discretized) equations (3) and (6 – 8). implicit integration scheme allows us to use much larger (n) (n+1) Here φi ≈ φi(tn) and ωi ≈ ωi(tn+1) describe the time steps than, e.g., the explicit scheme used in earlier discrete evolution of φi, ωi in the time discrete dynamics. work [29], and as a result, the simulations are up to 100 (n+1) In order to solve equation (9) for ωi , one must thus times faster [22]. We solve equation (3) with the explicit solve the whole complexly nested non-linear system of the scheme from Tzemeres et al. [51] and to calculate φi, discretized equations (3), (6 – 8), and (9) by an intera- the integrals in equations (6 – 8) are approximated by a tive method. The unknowns of the SCF-EPD model are standard Euler method. (n+1) (n+1) (n+1) 0(n+1) ωi (~r), φi (~r), g (~r, s), and g (~r, s) with The first step in the derivation of this integrator from i = A, B, S at all m spatial grid points and discretized equation (9) is to extract explicit ωi-expressions from 1 positions s = n ds of distance ds = /N along a poly- ∆φi. Equation (8) directly yields the exact relation mer chain with n ∈ {0, ..., N}. It is evident that the number of unknowns and thus, the dimension of the dis- ∆φS = φS [∇ωS · ∇ωS − ∆ωS] . (A4) cretized system of equations, increases dramatically with the polymer chain length N. In fact, the iteration along To obtain analogous expressions for ∆φA and ∆φB we the polymer chain for every ~r to calculate g and g0 typi- apply the Feynman-Kac formula. It states that solutions cally consumes by far the most part of the computation to equation (3) can be defined recursively by time even for algebraic update rules from explicit inte- Z grators for equation (3). Therefore, solving the complete 0 0 0 set of equations (3), 6 – 8, and 9 by an iterative method g[ω](~r, s) = exp (−dsNω(~r)) Φ(~r − ~r )g[ω](~r , s − ds)d~r V would lead to a particularly dramatic increase of simula- =: exp (−dsNω(~r)) I[ω](~r), tion times with growing N. (A5) Hence, implicit iterative schemes that involve multiple where Φ is proportional to the bond transition probabil- (n+1) calculations of φi help to overcome stiffness insta- ity of a Gaussian chain [36] and ω is given by equation 10

h  (n) (n) (n) i (n+1) h (n)  (n) i (n+1) 1 − hDP (1 + 2˜κH φA )∆ − 4βAκ˜H φA ∇ωA · ∇ ωA − hDP κ˜H 2φB ∆ − 2βB ∇ωB · ∇ ωB

h (n) (n) i (n+1) (n) (n) − hDP κ˜H φS (∆ − βS ∇ωS · ∇) ωS = ωA − hDP EA , (A1)

h (n)  (n) i (n+1) h  (n) (n) (n) i (n+1) − hDP κ˜H 2φA ∆ − 2βA∇ωA · ∇ ωA + 1 − hDP (1 + 2˜κH φB )∆ − 4βB κ˜H φB ∇ωB · ∇ ωB

h (n)  (n) i (n+1) (n) (n) − hDP κ˜H φS ∆ − βS ∇ωS · ∇ ωS = ωB − hDP EB , (A2)

h (n)  (n) i (n+1) h (n)  (n) i (n+1) − hDS κ˜H 2φA ∆ − 2βA∇ωA · ∇ ωA − hDS κ˜H 2φB ∆ − 2βB ∇ωB · ∇ ωB

h  (n) (n) (n) i (n+1) (n) (n) + 1 − hDS (1 +κ ˜H φS )∆ − βS κ˜H φS ∇ωS · ∇ ωS = ωS − hDS ES . (A3)

Figure 7. Semi-implicit integrator for the potential field equations 9 to increase stability regions at high κH .

(4). Likewise, one has integrals according to

Z tn+1 n 0 0 g [ω](~r) = exp(−dsNω) I [ω](~r) (A6) ωi(tn+1) = ωi(tn) − Di ∆µi tn o with ω from equation (5). Inserting the recursive def- + ∆ωi +κ ˜H (φAXA + φBXB + φSXS) dt initions of the end segment distribution functions from Z tn+1 n o equations (A5) and (A6) into the respective equation (6) + Di ∆ωi +κ ˜H (φAXA + φBXB + φSXS) dt or (7) yields tn (A10) Since the first integral on the right hand side of equa- ∆φi = φi [4dsN∇ωi · ∇ωi − 2dsN∆ωi] + Ri (A7) tion (A10) contains the chemical potential minus the leading stiffness contributions, we approximate it by an explicit Euler formula, i.e. for i = A, B. The remainder Ri summarizes all non- leading stiffness contributions, which in this case are all Z tn+1 n o terms that contain derivatives of the integrals I and I0 ∆µi + ∆ωi +κ ˜H (φAXA + φBXB + φSXS) dt = T tn with respect to the entries rj of ~r = (r1, r2, r3) . Since n (n) (n)  (n) (n) (n) (n) (n) (n) o 1 N we use ds = / in our simulations, we set dsN = 1. In h ∆µi + ∆ωi +κ ˜H φA XA + φB XB + φS XS the following we use the short hand notations (n) =: hEi . (A11) XS = ∆ωS − βS∇ωS · ∇ωS (A8) The second integral on the right hand side of equation (A10) is treated semi-implicitly. To approximate the in- R and cluded integrals φiXi dt we use the quadrature formulas

Z tn+1 Xi = 2∆ωi − βi4∇ωi · ∇ωi (A9) (n) (n+1) φi∆ωi dt ≈ φ ∆ωi h (A12) tn for i = A, B, where β , β , β ∈ [0, 1] are damping A B S and coefficients to adjust the ’degree’ of the implicit treat- Z tn+1 ment and to regulate truncation errors: the smaller the (n) (n) (n+1) damping coefficients, the smaller are truncation errors. φi∇ωi · ∇ωi dt ≈ φi ∇ωi · ∇ωi h, (A13) t To prepare the dynamical equations (9) for the semi- n implicit integrator we defineκ ˜H = ακH with another which are first order time accurate. The first ∇ωi on the damping coefficient α ∈ [0, 1] and zero-pad their time right hand side of equation (A13) is treated explicitly to 11

(n+1) obtain a linear system of equations for ωi . Inserting robust [52]. In our particular implementation a Krylow the short hand notations from equations (A8), (A9), and iteration is considered to have converged once the resid- (A11) together with the quadrature formulas (A12) and ual drops below 10−10 or the maximum number of 50 (A13) into equation (A10) yields the semi-implicit inte- iterative steps is reached. In case the residual is still gration scheme shown in figure 7, equations (A1), (A2), above 10−8 after 50 steps, we decrease the width of sub- and (A3)). sequent time steps by multiplication with 1/1.5. Further details on the implementation of the algorithm can be found in [22]. We set the damping coefficients to α = 0.5 Spatial derivatives are discretized by a second order and βA = βB = βS = 0. For this choice, the simulations finite differences. To solve the resulting discrete linear in the present work are found to be sufficiently stable. (n+1) system of equations for ωi (~r) we use the Generalized Minimal Residual Method (GMRES), which is a Krylow subspace iteration method for linear systems with posi- REFERENCES tive semidefinite matrices and known to be efficient and

[1] D. M. Herlach, I. Klassen, P. Wette, and D. Holland- D. Metzke, P. L¨ob,V. Hessel, and M. Maskos. Size con- Moritz. Colloids as model systems for metals and alloys: trolled polymersomes by continuous self-assembly in mi- A case study of crystallization. Journal of Physics: Con- cromixers. Polymer, 53(11):2205–2210, 2012. densed Matter, 22(15):153101, 2010. [14] S. Tenzer, D. Docter, S. Rosfa, A. Wlodarski, J. Kuharev, [2] F. Smallenburg, N. Boon, M. Kater, M. Dijkstra, and A. Rekik, S. K. Knauer, C. Bantz, T. Nawroth, C. Bier, R. Roij. Phase Diagrams of Colloidal Spheres with a J. Sirirattanapan, W. Mann, L. Treuel, R. Zellner, Constant Zeta-Potential. Journal of Chemical Physics, M. Maskos, H. Schild, and R. H. Stauber. Nanoparti- 134:074505, 2011. cle Size Is a Critical Physicochemical Determinant of the [3] M. A. Bucaro, P. R. Kolodner, J. A. Taylor, A. Sidorenko, Human Blood Plasma Corona: A Comprehensive Quan- J. Aizenberg, and T. N. Krupenkin. Tunable Liquid Op- titative Proteomic Analysis. ACS Nano, 5(9):7155–7167, tics: Electrowetting-Controlled Liquid Mirrors Based on 2011. Self-Assembled Janus Tiles. Langmuir, 25(6):3876–3879, [15] H. Maeda, J. Wu, T. Sawa, Y. Matsumura, and K. Hori. 2009. Tumor vascular permeability and the epr effect in macro- [4] D. Lensen, D. M. Vriezema, and van Hest, J. C. M. Poly- molecular therapeutics: a review. Journal of Controlled meric Microcapsules for Synthetic Applications. Macro- Release, 65:271–284, 2000. molecular Bioscience, 8(11):991–1005, 2008. [16] C. Zhang, V. J. Pansare, R .K. Prud’homme, and [5] D. E. Discher and A. Eisenberg. Polymer Vesicles. Sci- R. D. Priestley. Flash nanoprecipitation of polysterene ence, 297:967–972, 2002. nanoparticles. Soft Matter, 8:86–93, 2012. [6] L. Zhang, F. X. Gu, J. Chan, A. Z. Wang, R. S. Langer, [17] W. M¨uller.Hydrophobe und hydrophile Beladung poly- and O. Farokhzad. Nanoparticles in Medicine: Thera- merer Vesikel. Dissertation, Johannes Gutenberg Univer- peutic Applications and Developments. Clinical Phar- sity Mainz, 2009. macology and Therapeutics, 83:761–769, 2008. [18] A. Nikoubashman, V. E. Lee, C. Sosa, R. K. [7] M. E. Gindy and R. K. Prud’homme. Multifunctional Prud’homme, R. D. Priestley, and A. Z. Panagiotopou- nanoparticles for imaging, delivery and targeting in can- los. Directed Assembly of Soft Colloids through Rapid cer therapy. Expert Opinion on Drug Delivery, 6(8):865– Solvent Exchange. ACS Nano, 10:1425–1433, 2016. 878, 2009. [19] L. Falk and J. M. Commenge. Performance comparison of [8] P. Vartholomeos, M. Fruchard, A. Ferreira, and micromixers. Chemical Engineering Science, 65(1):405– C. Mavroidis. MRI-Guided Nanorobotic Systems for 411, 2010. Therapeutic and Diagnostic Applications. Annual Re- [20] K. S. Drese. Optimization of interdigital micromixers view of Biomedical Engineering, 13(1):157–184, 2011. via analytical modeling—exemplified with the SuperFo- [9] T. M. Allen and P. R. Cullis. Drug Delivery Systems: cus mixer. Chemical Engineering Journal, 101(1-3):403– Entering the Mainstream. Science, 303:1818–1822, 2004. 407, 2004. [10] R. Bleul. Herstellung, Characterisierung und Funktion- [21] F. Sch¨onfeld, K. S. Drese, S. Hardt, V. Hessel, and alisierung polymerer Nanopartikel und Untersuchung der C. Hofman. Optimized distributive µ-mixing by ’chaotic’ Wechselwirkung mit biologischen Systemen. Dissertation, multilamination. NSTI-Nanotech, 1:378–381, 2004. Freie Universitt Berlin, 2014. [22] S. Keßler. Dissertation Universit¨atMainz, 2017. in [11] B. M. Discher, Y. Y. Won, D. S. Ege, J. C. M. Lee, F. S. preparation. Bates, D. E. Discher, and D. A. Hammer. Polymersomes: [23] S. Keßler, F. Schmid, and K. Drese. Modeling size con- Tough Vesicles Made from Diblock Copolymers. Science, trolled nanoparticle precipitation with the co-solvency 284:1143–1146, 1999. method by spinodal decomposition. Soft Matter, 12:7231 [12] R. Thiermann. Selbstorganisation amphiphiler Block- – 7240, 2016. Copolymere in Mikromischern. Dissertation, Berlin Uni- [24] M. E. Wall, M. C. Wani, C. E. Cook, K. H. Palmer, versity of Technology, 2014. McPhail A. T., and G. A. Sim. Plant Antitumor Agents. [13] R. Thiermann, W. M¨uller,A. Montesinos-Castellanos, I. The Isolation and Structure of Camptothecin, a Novel 12

Alkaloidal Leukemia and Tumor Inhibitor from Camp- [40] D. J. Adams, S. Adams, D. Atkins, M. F. Butler, and totheca acuminata. Journal of American Chemical Soci- S. Furzeland. Impact of mechanism of formation on en- ety, 88:3888 – 3890, 1966. capsulation in block copolymer vesicles. J. Contr. Re- [25] V. P. Torchilin. Recent advances with liposomes as phar- lease, 128(2):165–170, 2008. maceutical carriers. Nature Review Drug Discovery, 4:145 [41] J. Gummel, M. Sztucki, T. Narayanan, and M. Gradziel- – 160, 2005. ski. Concentration dependent pathways in sponta- [26] W. M¨uller,K. Koynov, S. Pierrat, R. Thiermann, C. Fis- neous self-assembly of unilamellar vesicles. Soft matter, cher, and M. Maskos. pH-change protective PB-b-PEO 7(12):5731–5738, 2011. polymersomes. Polymer, 52:1263 – 1267, 2011. [42] V. Sofonea and K. R. Mecke. Morphological characteri- [27] A. V. Kabanov, E. V. Batrakova, and V. Y. Alakhov. zation of spinodal decomposition kinetics. The European Pluronic((R)) block copolymers for overcoming drug re- Physical Journal B, 8(1):99–112, 1999. sistance in cancer. Advanced Drug Delivery Reviews, [43] T. T. Nielsen. An Implementation Of The 54:759 – 779, 2002. Connected Component Labelling Algorithm. [28] N. M. Maurits and J. G. E. M. Fraaije. Mesoscopic dy- https://www.codeproject.com/articles/825200/an- namics of copolymer melts: From density dynamics to implementation-of-the-connected-component-label, external potential dynamics using nonlocal kinetic cou- 2014. pling. The Journal of Chemical Physics, 107(15):5879, [44] H. Mantz, K. Jacobs, and K. Mecke. Utilizing Minkowski 1997. functionals for image analysis: A marching square algo- [29] X. He and F. Schmid. Dynamics of Spontaneous Vesi- rithm. Journal of Statistical Mechanics: Theory and Ex- cle Formation in Dilute Solutions of Amphiphilic Diblock periment, 2008(12):P12015, 2008. Copolymers. Macromolecules, 39(7):2654–2662, 2006. [45] J. Heuser, G. J. A. Sevink, and F. Schmid. Self-assembly [30] X. He and F. Schmid. Spontaneous Formation of Com- of polymeric particles in poiseuille flow: A hybrid lat- plex Micelles from a Homogeneous Solution. Physical tice boltzmann/external potential dynamics simulation Review Letters, 100(13), 2008. study. Macromolecules, 50(11):4474–4490, 2017. [31] M. M¨ullerand F. Schmid. Incorporating Fluctuations [46] J. J. van Franeker, G. H. L. Heintges, C. Schaefer, G. Por- and Dynamics in Self-Consistent Field Theories for Poly- tale, W. Li, M. M. Wienk, P. van der Schoot, and R. A. J. mer Blends. In Advances in Polymer Science, volume Janssen. Polymer Solar Cells: Solubility Controls Fiber 185, pages 1–85. Springer Verlag, 2005. Network Formation. Journal of the American Chemical [32] S. F. Edwards. The statistical mechanics of polymers Society, 137:11783–11794, 2015. with excluded volume. Proceedings of the Physical Soci- [47] K. F. Freed. Interrelation between density functional ety, 85:613 – 624, 1965. and self-consistent-field formulations for inhomogeneous [33] F. Schmid. Self-consistent-field theories for complex flu- polymer systems. The Journal of Chemical Physics, ids. Journal of Physics: Condensed Matter, 10:8105– 103(8):3230, 1995. 8138, 1998. [48] G. Dahlquist. A special stability problem for linear mul- [34] E. Helfand. Theory of inhomogeneous polymers: Funda- tistep methods. Communications of the ACM, 14(3):176– mentals of the Gaussian random-walk model. The Jour- 179, 1963. nal of Chemical Physics, 62(3):999, 1975. [49] J. Zhu, L. Q. Shen, J. Shen, and V. Tikare. Coarsening [35] E. Helfand and Y. Tagami. Theory of the Interface be- kinetics from a variable-mobility Cahn-Hilliard equation: tween Immiscible Polymers. II. The Journal of Chemical Application o fa semi-implicit Fourier spectral method. Physics, 56(7):3592, 1972. Physical Review E, 60(4):3564–3572, 1999. [36] G. H. Fredrickson. The Equilibrium Theory of Inhomo- [50] U. M. Ascher, S. J. Ruuth, and B. T. R. Wetton. geneous Polymers. Oxford University Press, 2006. Implicit-explicit methods for time-dependent partial dif- [37] K. Kawasaki and K. Sekimoto. Dynamical Theory of ferential equations. SIAM Journal on Numerical Analy- Polymer Melt Morphology. Physica, 143A:349–413, 1987. sis, 32(3):797–823, 1995. [38] T. Uneyama. Density functional simulation of sponta- [51] G. Tzeremes, K. O. Rasmussen, T. Lookman, and A. Sax- neous formation of vesicle in block copolymer solutions. ena. Efficient computation of the structural phase behav- The Journal of Chemical Physics, 126(11):114920, 2007. ior of block copolymers. Physical Review E, 65:041806 [39] T. M. Weiss, T. Narayanan, and M. Gradzielski. Dynam- 1–5, 2002. ics of spontaneous vesicle formation in fluorocarbon and [52] Y. Saad and M. H. Schlutz. GMRES: A generalized hydrocarbon surfactant mixtures. Langmuir, 24(8):3759– minimal residual algorithm for solving nonsymmetric lin- 3766, APR 15 2008. ear systems. SIAM Journal on Scientific and Statistical Computing, 7:856–869, 1986.