An EventDriven Hybrid Molecular Dynamics and Direct Simulation Monte Carlo Algorithm
Abstract
A novel Stochastic EventDriven Molecular Dynamics (SEDMD) algorithm is developed for the simulation of polymer chains suspended in a solvent. The polymers are represented as chains of hard spheres tethered by square wells and interact with the solvent particles with hard core potentials. The algorithm uses EventDriven Molecular Dynamics (EDMD) for the simulation of the polymer chain and the interactions between the chain beads and the surrounding solvent particles. The interactions between the solvent particles themselves are not treated deterministically as in eventdriven algorithms, rather, the momentum and energy exchange in the solvent is determined stochastically using the Direct Simulation Monte Carlo (DSMC) method. The coupling between the solvent and the solute is consistently represented at the particle level, however, unlike full MD simulations of both the solvent and the solute, the spatial structure of the solvent is ignored. The algorithm is described in detail and applied to the study of the dynamics of a polymer chain tethered to a hard wall subjected to uniform shear. The algorithm closely reproduces full MD simulations with two orders of magnitude greater efficiency. Results do not confirm the existence of periodic (cycling) motion of the polymer chain.
I Introduction
Driven by nanoscience interests it has become necessary to develop tools for hydrodynamic calculations at the atomistic scale Nanohydrodynamics_Alder (); DSMC_MPCD_Gompper (); Microfluidics_Review (); PolymerTumbling_Review (); FluctuatingHydro_Coveney (). Of particular interest is the modeling of polymers in a flowing “good” solvent for both biological (e.g., cell membranes) and engineering (e.g., microchannel DNA arrays) applications PolymerDynamics_Review (); PolymerTumbling_Review (). The most widely studied polymer models are simple linear beadspring; freelyjointed rods; or wormlike chains. Such models have been parameterized for important biological and synthetic polymers. Much theoretical, computational, and experimental knowledge about the behavior of these models has been accumulated for various representations of the solvent. However, the multiscale nature of the problem for both time and length is still a challenge for simulation of reasonably large systems over reasonably long times. Furthermore, the omission in these models of the explicit coupling between the solvent and the polymer chain(s) requires the introduction of adjustable parameters (e.g., friction coefficients) to be determined empirically. The algorithm presented here overcomes this for a linear polymer chain tethered to a hard wall and subjected to a simple linear shear flow TetheredPolymer_Experiment_PRL (); TetheredPolymer_HybridMD (); TetheredPolymer_FullMD (); TetheredPolymer_Cyclic_PRL (); TetheredDNA_FullMD (). Of particular interest is the longtime dynamics of the polymer chain TetheredPolymer_Experiment_PRL (); TetheredPolymer_FullMD (); PolymerTumbling_PRL (); TetheredPolymer_Cyclic_PRL (); TetheredPolymer_Cyclic_AIP () and any effects of the polymer motion on the flow field.
Brownian dynamics is one of the standard methods for coupling the polymer chains to the solvent BrownianDynamics_DNA (); BrownianDynamics_OrderN (). The solvent is only implicitly represented by a coupling between the polymer beads and the solvent in the form of stochastic (whitenoise) forcing and linear frictional damping. The flow in the solvent is not explicitly simulated, but approximated as a small perturbation based on the Oseen tensor. This approximation is only accurate at large separations of the beads and at sufficiently small Reynold’s numbers. Even algorithms that do model the solvent explicitly via Lattice Boltzmann (LB) LatticeBoltzmann_Polymers (), incompressible (low Reynolds number) CFD solvers FluctuatingHydro_FluidOnly (); DNA_Laden_Flow (); FluctuatingHydroMD_Coveney (), or multiparticle collision dynamics PolymerCollapse_Yeomans (); Trebotich_HardRods (); MultiparticleDSMC_Polymers (); DSMC_MPCD_MD_Kapral (), typically involve phenomenological coupling between the polymer chain and the flowing fluid in the form of a linear friction term based on an effective viscosity (a notable exception being the algorithm described in Ref. DSMC_MPCD_MD_Kapral ()). Furthermore, solvent fluctuations in the force on the polymer beads are often approximated without fully accounting for spatial and temporal correlations. Finally, the reverse coupling of the effect of the bead motion on the fluid flow is either neglected or approximated with delta function forcing terms in the continuum fluid solver Trebotich_Penalty (). More fundamentally, continuum descriptions of flow at micro and nanoscales are known to have important deficiencies Microfluidics_Review (); Nanohydrodynamics_Alder () and therefore it is important to develop an allparticle algorithm that is able to reach the long times necessary for quantitative evaluation of approximate, but faster, algorithms.
The most detailed (and expensive) modeling of polymers in flow is explicit molecular dynamics (MD) simulation of both the polymer (solute) and the surrounding solvent PolymerShear_MD (); TetheredDNA_FullMD (). Multiscale algorithms have been developed to couple the MD simulation to NavierStokesbased computational fluid dynamics (CFD) calculations of the flow field TetheredPolymer_HybridMD (). However, the calculation time still remains limited by the slow molecular dynamics component. Thus the computational effort is wasted on simulating the structure and dynamics of the solvent particles, even though it is the polymer structure and dynamics (and their coupling to the fluid flow) that is of interest. Our algorithm replaces the deterministic treatment of the solventsolvent interactions with a stochastic momentum exchange operation, thus significantly lowering the computational cost of the algorithm, while preserving microscopic details in the solventsolute coupling.
Fluctuations drive the polymer motion and must be accurately represented in any model. Considerable effort has been invested in recent years in including fluctuations directly into the NavierStokes (NS) equations and the associated CFD solvers FluctuatingHydro_FluidOnly (); FluctuatingHydro_Coveney (); FluctuatingHydro_Garcia (). Such fluctuating hydrodynamics has been coupled to molecular dynamics simulations of polymer chains FluctuatingHydroMD_Coveney (), but with empirical coupling between the beads and the fluid as discussed above. To avoid the empirical coupling, the solvent region could be enlarged by embedding the atomistic simulations of the region around the polymer chain (such as pure MD or our combined MD/DSMC algorithm) in a fluctuating hydrodynamics region. The bidirectional coupling between the continuum and particle regions has to be constructed with great care so that both fluxes and fluctuations are preserved FluctuatingHydro_AMAR (). A wellknown problem with such multiscale approaches is that the finest scale (atomistic simulation) can take up the majority of computational time and thus slow down the whole simulation. By using DSMC the cost of the particle region can be made comparable to that of the continuum component.
The Stochastic EventDriven Molecular Dynamics (SEDMD) algorithm presented here combines EventDriven Molecular Dynamics (EDMD) for the polymer particles with Direct Simulation Monte Carlo (DSMC) for the solvent particles. The polymers are represented as chains of hard spheres tethered by square wells. The solvent particles are realistically smaller than the beads and are considered as hard spheres that interact with the polymer beads with the usual hardcore repulsion. The algorithm processes true (deterministic, exact) binary collisions between the solvent particles and the beads, without any approximate coupling or stochastic forcing. However, for the purposes of the EDMD algorithm, the solvent particles themselves do not directly interact with each other, that is, they can freely pass through each other as for an ideal gas. Deterministic collisions between the solvent particles are replaced with stochastic DSMC collisions. Both asynchronous (eventdriven) and synchronous (timedriven) algorithmic ways of processing these stochastic collisions will be discussed in the next section. Note that our algorithm is similar to a recent algorithm developed for soft interaction potentials combining timedriven MD with multiparticle collision dynamics DSMC_MPCD_MD_Kapral ().
Ii Hybrid Components
In this section we briefly describe the two components of the SEDMD algorithm: The stochastic handling of the solvent and the deterministic handling of the solute particles. These two components are integrated (i.e., tightly coupled) into a single eventdriven algorithm in Section III.
ii.1 Solvent DSMC Model
The validity of the (incompressible) NavierStokes continuum equations for modeling microscopic flows has been well established down to length scales of Microfluidics_Review (). However, there are several issues present in microscopic flows that are difficult to account for in models relying on a purely PDE approximation. Firstly, it is not a priori obvious how to treat boundaries and interfaces well so as account for the nontrivial (possibly nonlinear) coupling between the flow and the microgeometry. Furthermore, fluctuations are not typically considered in NavierStokes solvers, and they can be very important at instabilities FluidMixing_DSMC () or in driving polymer dynamics. Finally, since the grid cell sizes needed to resolve complex microscopic flows are small, a large computational effort (comparable to DSMC) is needed even for continuum solvers. An alternative is to use particlebased methods, which are explicit and unconditionally stable and rather simple to implement. The solvent particles are directly coupled to the microgeometry, for example, they directly interact with the beads of a polymer. Fluctuations occur naturally with the correct spatiotemporal correlations. However, as in continuum descriptions, the structure of the fluid is lost and under certain conditions the high compressibility of the DSMC (ideal gas) fluid can cause difficulties.
Several particle methods have been described in the literature, such as MD PolymerShear_MD (), dissipative particle dynamics (DPD) DPD_DNA (), and multiparticle collision dynamics (MPCD) DSMC_MPCD_Gompper (); DSMC_MPCD_MD_Kapral (). Our method is similar to MPCD (also called stochastic rotation dynamics or the MalevanetsKapral method), and in fact, both are closely related to the Direct Simulation Monte Carlo (DSMC) algorithm of Bird DSMCReview_Garcia (). The key idea behind DSMC is to replace deterministic interactions between the particles with stochastic momentum exchange (collisions) between nearby particles. Specifically, particles are propagated by a fixed time step , as in MD, moving ballistically along straight lines during a timestep (advection step). At the end of each time step, the particles are sorted into cells, each containing on the order of ten particles, and then a certain number of random pairs of particles that are in the same cell are chosen to undergo stochastic collisions (collision step). These collisions do not take into account the positions of the particles other than the fact they are in the same cell (i.e., they are nearby). The collisions conserve momentum and energy (but not angular momentum) exactly. Formally, DSMC can be seen as a method for solving the Boltzmann transport equation for a lowdensity gas, however, it is not limited to gas flows DSMC_DenseFluids (); DSMC_CBA (); DSMC_CBATheory (). Our purpose for using DSMC is as a replacement for expensive MD, preserving the essential hydrodynamic “solvent” properties: local momentum conservation, and linear momentum exchange on length scales comparable to the particle size, and a similar fluctuation spectrum.
In the multiparticle collision variant of this algorithm originally proposed by Kapral, the traditional DSMC collection of binary collisions is replaced by a multiparticle collision in which the velocities of all particles in the cell are rotated by a random amount around the average velocity DSMC_MPCD_Gompper (); DSMC_MPCD_MD_Kapral (). This change improves efficiency but at the cost of some artificial effects such as loss of Galilean invariance. These problems can be corrected and the method has been successfully used in modeling polymers in flow by including the beads, considered as (massive) point particles, in the stochastic momentum exchange step MultiparticleDSMC_Polymers (); PolymerCollapse_Yeomans (); ShearThinning_Yeomans (). We will employ traditional DSMC in our algorithm in order to mimic the actual (deterministic) momentum exchange between solvent molecules (as it would be in an MD simulation) and in order to avoid any possible artifacts.
A fundamental deficiency of DSMC as a (micro or nano) hydrodynamic solver is the large (ideal gas) compressibility of the fluid. For subsonic flows this compressibility does not qualitatively affect the results as the DSMC fluid will behave similarly to an incompressible liquid, however, the (Poisson) density fluctuations in DSMC are significantly larger than those in realistic liquids. Furthermore, the speed of sound is small (comparable to the average speed of the particles) and thus subsonic (Mach number less than one) flows are limited to relatively small Reynolds numbers ^{1}^{1}1For a lowdensity gas the Reynolds number is , where is the Mach number, and the Knudsen number is the ratio between the mean free path and the typical obstacle length . This shows that subsonic flows can only achieve high flows for small Knudsen numbers, i.e., large numbers of DSMC particles.. The Consistent Boltzmann Algorithm (CBA) DSMC_CBA (); DSMC_CBATheory (), as well as algorithms based on the Enskog equation DSMC_Enskog (); DSMC_Enskog_Frezzotti (), have demonstrated that DSMC fluids can have densefluid compressibility. A similar algorithm was recently constructed for MPCD MPCD_CBA (). We are currently evaluating several DSMC variants in terms of their efficiency and thermodynamic consistency under high densities ^{2}^{2}2Note that the density fluctuations in the CBA fluid are identical to those in an ideal gas and thus thermodynamically inconsistent with the compressibility. and will report our findings in future work.
ii.2 Polymer MD Model
Polymer chains in a solvent are modeled using continuous pair potentials and timedriven MD (TDMD), in which particles are synchronously propagated using a time step , integrating the equations of motion along the way. For good solvents, the polymer beads are represented as spherical particles that interact with other beads and solvent particles with (mostly) repulsive pair potentials, such as the positive part of the LennardJones potential. Additionally, beads are connected via (usually finitelyextensible FENE or wormlike) springs in order to mimic chain connectivity and elasticity PolymerShear_MD (). Additionally, stochastic forces may be present to represent the solvent. The time steps required for integration of the equations of motion in the presence of the strongly repulsive forces is small and TDMD cannot reach long time scales even after parallelization. An alternative is to use hard spheres instead of soft particles, allowing replacement of the FENE springs with squarewell tethers, thus avoiding the costly force evaluations in traditional MD. Hard sphere MD is most efficiently performed using eventdriven molecular dynamics (EDMD) EventDriven_Alder (); Event_Driven_HE (); EDMD_Polymers_Hall (); PolymerCollapse_EDMD (). If the detailed structure and energetics of the liquid is not crucial, such EDMD algorithms can be just as effective as TDMD ones but considerably faster. The essential difference between EDMD and TDMD is that EDMD is asynchronous and there is no time step, instead, collisions between hard particles are explicitly predicted and processed at their exact (to numerical precision) time of occurrence. Since particles move along simple trajectories (straight lines) between collisions, the algorithm does not waste any time simulating motion in between events (collisions).
Hardsphere models of polymer chains have been used in EDMD simulations for some time EDMD_Polymers_Hall (); EDMD_Polymers_Aggregation (); EDMD_Polymer_Fibrils2 (). These models typically involve, in addition to the usual hardcore exclusion, additional square well interactions to model chain connectivity. The original work by Alder et al. on EDMD developed the collisional rules needed to handle arbitrary square wells EventDriven_Alder (). Infinitely high wells can model tethers between beads, and the tethers can be allowed to be broken by making the square wells of finite height, modeling soft shortrange attractions. Recent studies have used square well attraction to model the effect of solvent quality PolymerCollapse_EDMD (). Even more complex square well models have been developed for polymers with chemical structure and it has been demonstrated that such models, despite their apparent simplicity, can successfully reproduce the complex packing structures found in polymer aggregation EDMD_Polymer_Fibrils2 (); EDMD_Polymers_Aggregation (). Recent work on coupling a Kramer beadrod polymer to a NS solver has found that the use of hard rods (instead of soft interactions) not only rigorously prevents rodrod crossing but also achieves a larger time step, comparable to the time step of the continuum solver Trebotich_HardRods ().
This study is focused on the simplest model of a polymer chain, namely, a linear chain of particles tethered by unbreakable bonds. This is similar to the commonlyused freely jointed beadspring FENE model model used in timedriven MD. The length of the tethers has been chosen to be , where is the diameter of the beads ^{3}^{3}3Note that the hardsphere model rigorously prevents chain crossing if the tether length is less than since two tethers shorter than this length cannot pass through each other without violating impenetrability.. The implementation of squarewell potentials is based on the use of nearneighbor lists (NNLs) in EDMD, and allows for the specification of squarewell interactions for arbitrary pairs of near neighbors. In particular, one can specify a minimal and maximal distance (tether length) for arbitrary pairs of near neighbors ^{4}^{4}4A value can be used to emulate chain rigidity (i.e., a finite persistence length) by using second nearestneighbor interactions between chain beads..
Iii Details of Hybrid Algorithm
In this section the hybrid EDMD/DSMC algorithm, which we name Stochastic EDMD (SEDMD), is described in detail. Only a brief review of the basic features of EDMD is given and the focus is on the DSMC component of the algorithm and the associated changes to the EDMD algorithm described in detail in Ref. Event_Driven_HE (). A more general description of asynchronous eventdriven particle algorithms is given in Ref. AED_Serial ().
Asynchronous eventdriven (AED) algorithms process a sequence of events (e.g., collisions) in order of increasing event time . The time of occurrence of events is predicted and the event is scheduled to occur by placing it an event queue. The simulation iteratively processes the event at the head of the event queue, possibly scheduling new events or invalidating old events. One impending event per particle , , is scheduled to occur at time with partner (e.g., another particle ). The particle position and velocity are only updated when an event involving particle is processed and the time of last update is recorded (we will refer to this procedure as a particle update). We note that traditional synchronous timedriven (STD) algorithms with a time step are a trivial variant of the more general AED class. In particular, in an STD algorithm events occur at equispaced times and each event is a time step requiring an update of all of the particles. The AED algorithm processes a mixture of events involving single particles or pairs of particles with time steps that involve the simultaneous (synchronous) update of a large collection of particles.
Every particle belongs to a certain specie . Particles with species and may or may not interact with each other (i.e., they may not be subject to the hardparticle nonoverlap condition). We focus on a system in which a large fraction of the particles belong to a special specie representing DSMC particles (e.g., solvent molecules). These DSMC particles do not interact with each other (i.e., they freely pass through each other), but they do interact with particles of other species. We focus on the case when the nonDSMC particles are localized in a fraction of the simulation volume, while the rest of the volume is filled with DSMC particles. This will enable us to treat the majority of DSMC particles sufficiently far away from nonDSMC particles more efficiently than those that may collide with nonDSMC particles.
Before describing the SEDMD algorithm in detail, we discuss the important issue of efficiently searching for nearby pairs of particles.
iii.1 Near Neighbor Searches
When predicting the impending event of a given particle , the time of potential collision between the particle and each of its neighbors (nearby particles) is predicted Event_Driven_HE (); AED_Serial (). The DSMC algorithm also requires defining neighbor particles, that is, particles that may collide stochastically during the DSMC collision step. For efficiency, geometric techniques are needed to make the number of neighbors of a given particle instead of .
In SEDMD we use the socalled linked list cell (LLC) method for neighbor searching in both the MD and DSMC components. The simulation domain is partitioned into cells as close to cubical as possible. Each particle stores the cell to which its centroid belongs, and each cell stores a list of all the particles it contains, as well as the total number of particles in the cell. For a given interaction range, neighbors are found by traversing the lists of as many neighboring cells as necessary to ensure that all particles within that interaction range are covered. In traditional DSMC, only particles within the same cell are considered neighbors and thus candidates for collision. There are also variants of DSMC in which particles in nearby cells are included in order to achieve a nonideal equation of state DSMC_Enskog (); DSMC_Enskog_Frezzotti (). A more general implementation would use different cell meshes for MD and DSMC neighbor searches, however, that would significantly complicate the implementation.
iii.1.1 Cell Bitmasks
In addition to the list of particles , each cell stores a bitmask consisting of bits (bitfields). These bits may be one (set) or zero (not set) to indicate certain properties of the cell, specifically, what species of particles the cell contains, whether the cell is event or time driven, and to specify boundary conditions. In order to distinguish the cells that contain nonDSMC particles (i.e., particles of specie other than ) from those that contain only DSMC particles, bit is set if the cell may contain a particle of specie . The bit is set whenever a particle of specie is added to the cell, and all of the masks are reset and then rebuilt (i.e., refreshed) periodically. When performing a neighbor search for a particle , cells not containing particles of species that interact with specie are easily found (by OR’ing the cell masks with a specie mask) and are simply skipped. This speeds up the processing of DSMC particles since cells containing only DSMC particles will be skipped without traversing their lists of particles.
For the purposes of the combined MD/DSMC algorithm we will also need to distinguish those cells that are nearby nonDSMC particles, that is, that contain particles within the interaction range of some nonDSMC particle. Such cells will be treated using a fully eventdriven (ED) scheme, while the remaining cells will be treated using a timedriven or mixed approach. We use one of the bits in the bitmasks, bit , to mark eventdriven (ED) cells whenever a neighbor search is performed for a nonDSMC particle. Specifically, bit is set for a given cell whenever the cell is traversed during a neighbor search for a nonDSMC particle. This scheme correctly masks the cells by only modifying the neighbor search routines without changing the rest of the algorithm, at the expense of a small overhead. We also mark the cells near hardwall boundaries as ED cells. Cell bitmasks should be refreshed (rebuilt) periodically so as to prevent the fraction of ED cells from increasing. As will be seen shortly, it is necessary to introduce at least one “sticky” bit that is not cleared but rather persists (has memory), and is initialized to zero (not set) at the beginning of the simulation.
iii.1.2 Near Neighbor Lists
The cell size should be tailored to the DSMC portion of the algorithm and can become much smaller than the size of nonDSMC particles. The LLC method becomes inefficient when the interaction (search) range becomes significantly larger than the cell size because many cells need to be traversed. In this case the LLC method can be augmented with the nearneighbor list (NNL) method, and in particular, the bounding sphere complexes (BSCs) method, as described in detail for nonspherical hard particles in Ref. Event_Driven_HE (). We have implemented the necessary changes to the algorithm to allow the use of NNLs and BSCs (in addition to LLCs), and we used NNLs in our simulations of polymer chains in solution. The use of BSCs is not necessary for efficient simulations of polymer solutions if the size of the polymer bead is comparable to the size of the cells, which is the case for the simulations we report. We do not describe the changes to the algorithm in detail; rather, we only briefly mention the essential modifications.
For the purposes of DSMC it is important to maintain accurate particle lists for all cells , so that it is known which particles are in the same cell (and thus candidates for stochastic collisions) at any point in time. Therefore, transfers of particles between cells need to be predicted and processed even though this is not done in the NNL algorithm described in Ref. Event_Driven_HE (). Nearneighbor lists are only built and maintained for DSMC particles that are in eventdriven cells (essentially exactly as described in Ref. Event_Driven_HE ()). For a DSMC particle that is not in an ED cell we consider the smallest sphere enclosing cell to be the (bounding) neighborhood (see Ref. AED_Serial ()) of particle and only update the (position of the) neighborhood when the particle moves to another cell. This ensures that neighbor searches using the NNLs are still exact without the overhead of predicting and processing NNL update events for the majority of the DSMC particles.
iii.2 The SEDMD Algorithm
We have developed an algorithm that combines timedriven DSMC with eventdriven MD by splitting the particles between ED particles and TD particles. Roughly speaking, only the particles inside eventdriven cells (i.e., cells for which bit is set) are part of the AED algorithm. The rest of the particles are DSMC particles that are not even inserted into the event queue. Instead, they are handled using a timedriven (TD) algorithm very similar to that used in classical DSMC.
It is also possible to implement DSMC as a fully asynchronous eventdriven (AED) algorithm and thus avoid the introduction of an external time scale through the time step . The algorithm introduces a novel type of event we term stochastic (DSMC) collisions, and it is discussed in more detail in Appendix A. Asynchronous processing has a few advantages over the traditional (synchronous) timedriven approach, notably, no errors due to time discretization DSMC_TimeStepError () and improved efficiency at low collision rates. For high densities (i.e., high collision rates) we have found that these advantages are outweighed by the (implementation and runtime) cost of the increased algorithmic complexity. Additionally, timedriven handling has certain important advantages in addition to its simplicity, notably, the synchrony of the DSMC portion of the algorithm allows for parallelization and easy incorporation of algorithmic alternatives (e.g., multiparticle or multicell collisions, adaptive open boundary conditions, etc.).
The main types of events in the SEDMD algorithm are:
 Update

Move particle to the current simulation time if .
 Transfer

Move particle from one cell to another when it crosses the boundary between two cells (this may also involve a translation by a multiple of the lattice vectors when using periodic BCs).
 Hardcore collision

Collide a particle with a boundary such as a hard wall or another particle with which it interacts.
 Tether collision

Bouncing of a pair of tethered particles in a polymer chain when the tether stretches (processed exactly like usual hardparticle collisions EventDriven_Alder (); EDMD_Polymers_Hall ()).
 Time step

Move all of the timedriven particles by and process stochastic collisions between them.
The position and time as well as the impending event prediction of particle are updated whenever an event involving the particle is processed.
Both the eventdriven and the timedriven DSMC algorithms process stochastic binary trial collisions. Processing a trial collision consists of randomly and uniformly selecting a pair of DSMC particles and that are in the same cell. For hard spheres in the lowdensity limit, the probability of collision for a particular pair is proportional to the relative velocity , and therefore the pair is accepted with probability . If a pair is accepted for collision than the velocities of and are updated in a random fashion while preserving energy and momentum DSMCReview_Garcia (). If a real collision involving an ED particle occurs then that particle is updated to time , its previous event prediction is invalidated (this may involve updating a thirdparty particle ), and an immediate update event is scheduled for (and possibly ).
It is important to note that the division of the DSMC particles between ED and TD handling is dynamic and does not necessarily correspond to the partitioning of the cells into ED and TD cells (based on the cell bitfield ). As nonDSMC particles move, timedriven cells may be masked as eventdriven. This does not immediately make the DSMC particles in such cells eventdriven. Rather, timedriven DSMC particles are moved into the event queue only when a collision with a nonDSMC particle is scheduled for them, when they move into a TD cell following a time step, or when restarting the event handling. Eventdriven particles are removed from the event queue when they undergo cell transfer events into timedriven cells.
iii.2.1 Time Step Events
The hybrid ED/TD algorithm introduces a new kind of event (not associated with any particular particle) called a time step event. This event is scheduled to occur at times , where is an integer. When such an event is processed, all of the DSMC particles not in the event queue are moved ^{5}^{5}5Note that this update may involve moving some particles by less than since the time of the last update for such particles does not have to be a time step event but could be, for example, a cell transfer. to time and are then resorted into cells (recall that the ED particles are already correctly sorted into cells). Particles that change from ED to TD cells and viceversa are removed or inserted into the event queue accordingly. Then, in each cell trial DSMC collisions are performed, where
(1) 
is the DSMC collision rate. Here in three dimensions and in two dimensions is the collisional crosssection, is the volume of the cell, and is an upper bound for the maximal particle velocity ^{6}^{6}6More precisely, is an upper bound on the maximal relative velocity between a pair of particles. In our implementation we maintain the maximal encountered particle velocity and update it after every collision and also reset it periodically..
In order to ensure correctness of the AED algorithm, a TD particle must not move by more than a certain distance when it undertakes a time step. Otherwise, it may overlap with a nonDSMC particle that could not have anticipated this and scheduled a collision accordingly. Specifically, recall that the eventdriven cells are marked whenever a neighbor search is performed for a nonDSMC particle. Our simulation uses
where the masking width is the minimal number of cells covered by any neighbor search in any direction, is the (minimal) cell length, and is the diameter of the DSMC particles. Any DSMC particle whose velocity exceeds is inserted into the event queue at the end of a time step, and similarly, any particles that would have been removed from the event queue are left in the queue if their velocity exceeds the maximum safe velocity. Typically, only a small (albeit nonzero) fraction of the DSMC particles falls into this category and the majority of the particles that are not in ED cells are not in the event queue. In fact, we choose the time step to be as large as possible while still keeping the number of dangerously fast DSMC particles negligible. This typically also ensures that DSMC particles do not jump over cells from one time step to the next (given that typically ).
iii.3 Adaptive Open Boundary Conditions
In three dimensions, a very large number of solvent particles is required to fill the simulation domain. The majority of these particles are far from the polymer chain and they are unlikely to significantly impact or be impacted by the motion of the polymer chain. It therefore seems reasonable to approximate the behavior of the solvent particles sufficiently far away from the region of interest with that of a quasiequilibrium ensemble in which the positions of the particles are as in equilibrium and the velocities follow a local Maxwellian distribution (the mean of which is equal to the macroscopic local velocity). These particles do not need to be simulated explicitly, especially for a DSMC liquid which has no spatial structure (ideal gas). Rather, we can think of the polymer chain and the surrounding DSMC fluid as being embedded into an infinite reservoir of DSMC particles which enter and leave the simulation domain following the appropriate distributions.
Such open (Grand Canonical) boundary conditions (BC) are often used in multiscale (coupled) simulations. It is not trivial to implement them when coupling the “reservoir” to an MD simulation, especially at higher densities. An example of an algorithm that achieves such a coupling for softparticle systems is USHER TetheredPolymer_HybridMD (). It is also nontrivial to account for the velocity distribution of the particles entering the simulation domain DSMC_InflowDistribution (), as would be needed in a purely eventdriven algorithm in which particles are inserted at the surface boundary of the domain. However, the combination of a partially timedriven algorithm and an unstructured (ideal gas) DSMC fluid makes it very easy to implement open BCs by inserting DSMC particles in the cells surrounding the simulation domain only at timestep events, based on very simple distributions.
iii.3.1 Cell Partitioning
For the purposes of implementing such nontrivial BCs, we classify the cells as being interior, boundary, and external cells. Interior cells are those that are in the vicinity of nonDSMC particles, specifically, cells that are within a window of halfwidth cells around the centroid of a nonDSMC particle. The interior cells are divided into eventdriven and timedriven and are handled as described previously. If a boundary or external cell is marked as an eventdriven cell the simulation is aborted with an error, ensuring that ED cells are always interior. Boundary cells surround the interior cells with a layer of cells of thickness cells, and they represent cells in which particles may be inserted during time step events ^{7}^{7}7Since ED cells are never boundary cells such insertions cannot lead to overlaps with nonDSMC particles.. External cells are noninterior cells that are not explicitly simulated, rather, they provide a boundary condition around the interior and boundary cells. This layer must be at least cells thick, and the cells within a layer of cells around the simulation domain (interior together with boundary cells) are marked as both external and boundary cells. All of the remaining cells are purely external cells and simply ignored by the simulation. Our implementation uses bits in the cell bitmasks to mark a cell as being eventdriven (bit ), boundary (bit ), or external (bit ). Note that a cell may be a combination of these, for example, cells near hard walls might be both interior and boundary, and some cells may be both external and boundary.
Figure 1 provides an illustration of this division of the cells for the simulation of a tethered polymer in two and three dimensions. Note that we do not require that the domains of interior or nonexternal cells form a rectangular domain: The final shapes and even contiguity of such domains depends on the positions of the nonDSMC particles ^{8}^{8}8If this is not appropriate one can always make the simulation regions (unions of disjoint) rectangular domains simply by padding with interior cells.. Our implementation traverses each of the nonDSMC particles in turn and masks the cells in a window of halfwidth cells around the cell containing the nonDSMC particle as:
 Interior

representing cells where the nontrivial flow occurs ()
 Boundary

representing cells where particles may be inserted or propagated during time step events.
 External

representing cells that are not explicitly simulated but rather only provide appropriate BCs.
The division of the cells into eventdriven, interior, boundary and external cells is rebuilt periodically during the simulation. This rebuilding may only happen at the beginning of time steps, and requires a synchronization of all of the particles to the current simulation time, a complete rebuilding of the cell bitmasks, and finally, a reinitialization of the event processing. Importantly, particles that are in purely external cells are removed from the simulation and those that are in eventdriven cells are reinserted into the event queue scheduled for an immediate update event. During the process of rebuilding the cell bitmasks cells that are masked as purely external cells are also marked with the sticky bit . This indicates that these cells need to be refilled with particles later if they enter the simulation domain again (due to the motion of the nonDSMC particles). Once the cell bitmasks are rebuilt, a time step event is executed as described next.
iii.3.2 Reservoir Particles
At the beginning of a time step event, after possibly rebuilding the cell masks, the timedriven DSMC particles are propagated as usual. If there are external cells, (trial) reservoir particles are then inserted into the cells that are both external and boundary, and also in cells whose sticky bit is set (i.e., cells that have not yet been filled with particles), after which the bit is reset. The use of the sticky bit to mark such cells ensures that subsequent rebuilding of the masks will not erase the flag (the sticky bit is only reset once the cell is filled with particles). The trial particles are thought to be at a time , and are propagated by a time step to the current simulation time. Only those particles that move into a nonexternal cell are accepted and converted into real particles. If the acceptance would insert the particle into a nonboundary cell (i.e., the particle moved by at least cells), the insertion is rejected and a count of the number of rejected particles reported to aid in choosing sufficiently large so as to ensure that the tails of the velocity distribution are not truncated (in our experience suffices for reasonable choices of ). Following the insertion of reservoir particles stochastic collisions are processed in each cell as usual.
For improved efficiency, it is possible to replace the volumebased particle reservoir with a surface reservoir, and insert particles only at the surface of the simulation domain DSMC_InflowDistribution (). However, we have not implemented such an approach since the boundary handling is not critical for the overall efficiency.
iii.3.3 Boundary Conditions
In our current implementation the reservoir particles follow simple localequilibrium ideal gas distributions. The number of particles to insert in a given cell is chosen from a Poisson distribution with the appropriate density, the positions are uniformly distributed inside the cell, and the velocities are drawn from a biased (local) Maxwellian distribution. The mean velocity and temperature for the local Maxwellian are chosen according to the specified boundary conditions (presently only uniform linear gradients are implemented). For example, if a uniform shear in the plane is to be applied, , where is the position of the centroid of the cell and is the shear rate. Using such biased local insertions allows one to specify a variety of boundary conditions (for example, a free polymer chain in unbounded shear flow) without resorting to hardwall boundaries or complicating LeeEdwards conditions.
It should be noted that in principle we should not use a local Maxwellian velocity distribution for a system that is not in equilibrium. In particular, for small velocity, temperature, and density gradients the ChapmanEnskog distribution is the appropriate one to use in order to avoid artifacts near the open boundaries at length scales comparable to the mean free path AMAR_DSMC (). We judge these effects to be insignificant in our simulations since our boundary conditions are fixed externally and are thus not affected by the possible small artifacts induced in the DSMC fluid flow, and since is small.
In the future, we plan to replace the particle reservoir with a PDEbased (NavierStokes) simulation coupled to the DSMC/MD one. Such a fluxpreserving coupling has been implemented in the past for coupled DSMC/Euler hydrodynamic simulations AMAR_DSMC (); AMAR_DSMC_SAMRAI (). It is however important for the coupling to also correctly couple fluctuations. This requires the use of fluctuating hydrodynamics in the coupled domain. Such solvers and associated coupling techniques are only now being developed FluctuatingHydro_Garcia (); FluctuatingHydro_AMAR ().
iii.4 Further Technical Details
In this section we discuss several technical details of the SEDMD algorithm such as hardwall boundary conditions and the choice of DSMC parameters.
iii.4.1 Slip and Stick Boundary Conditions
We have already discussed the open boundary conditions and their use to specify a variety of “farfield” flow patterns. Additionally, there can also be hardwall boundaries, i.e., flat impenetrable surfaces. These surfaces can have a velocity of their own and here we discuss how particles reflect from such walls in the frame that moves with the hard wall. Regardless of the details of particle reflections, the total change in linear momentum of all the particles colliding with a hard wall can be used to estimate the friction (drag) force acting on the wall due to the flow. This can give reliable and quick estimates of the viscosity of a DSMC fluid, for example. We use the classical noslip BCs (i.e., zero normal and parallel velocity) for smooth hardwall surfaces. Molecular simulations have found some slip; however, at lengthscales significantly larger than the mean free path and/or the typical surface roughness one may assume noslip boundaries if the hardwall boundary position is corrected by a slip length Microfluidics_Review ().
Our simulations of tethered polymers use thermal walls (kept at ) DSMCReview_Garcia () to implement noslip hard walls at the boundaries of the simulation cell. Following the collision of a particle with such a wall, the particle velocity is completely randomized and drawn from a half MaxwellBoltzmann distribution (other biased distributions may be used as appropriate). This automatically ensures a zero mean velocity at the wall boundary and also acts as a thermostat keeping the temperature constant even in the presence of shear heating. Noslip boundaries can also be implemented using (athermal) rough walls which reflect incoming particles with velocity that is the exact opposite of the incoming velocity DSMC_MPCD_CylinderFlow (). Similarly, slip boundary conditions (zero normal velocity) can be trivially implemented by using specular walls that only reverse the normal component of the velocity (relative to the wall). A mixture of the two can be used to implement partially rough walls, for example, a roughness parameter can be used as the probability of randomly selecting a rough versus a specular collision.
Similar considerations apply to the boundary conditions at the interface of a hard particle such as a polymer bead. Most particlebased methods developed for the simulation of particle suspensions consider the solvent particles as point particles for simplicity, and only MD or certain boundary discretization schemes SuspensionsDSMC_Reflection () resolve the actual solventsolute interface. Specular BCs are typical of MD simulations and assume perfectly conservative collisions (i.e., both linear momentum and energy are conserved). However, if the polymer beads are themselves composed of many atoms, they will act as a partially thermal (and rough) wall and energy will not be conserved exactly.
In the simulations reported here we have used rough walls for collisions between DSMC and nonDSMC particles. This emulates a nonstick boundary condition at the surface of the polymer beads. Using specular (slip) conditions lowers the friction coefficient ^{9}^{9}9The Stokes friction force has a coefficient of for slip BCs instead of the wellknown for noslip BCs., but does not appear to qualitatively affect the behavior of tethered polymers.
iii.4.2 Constant Pressure Flows
We note briefly on our implementation of constant pressure boundary conditions, as used to simulate flow through open pipes. A constant pressure is typically emulated in particle simulations via a constant acceleration for the DSMC particles DSMC_PlanePoiseuille () together with periodic BCs along the flow (acceleration) direction. In timedriven algorithms, one simply increments the velocity of every particle by and the position by at each timestep (before processing DSMC collisions). In SEDMD this is not easily implemented, since the trajectory of the DSMC particles becomes parabolic instead of linear and exact collision prediction between the DSMC and the nonDSMC particles is complicated. We have opted to implement constant pressure BCs by using a periodic deltafunction forcing on the DSMC particles. Specifically, the velocities of all DSMC particles are incremented ^{10}^{10}10Recall that the event prediction for any ED particle whose velocity is changed must be updated, typically by scheduling an immediate update event. at the beginning of each time step by , and then stochastic collisions are processed.
iii.4.3 Choice of DSMC Collision Frequency
The viscosity of the DSMC fluid is determined by the choice of collision frequency and cell size . Classical DSMC wisdom DSMCReview_Garcia () is that cell size should be smaller than the mean free path, , but large enough to contain on the order of particles (in three dimensions). It is obvious that both of these conditions cannot be satisfied for denser liquids, where is only a fraction of the particle size. It is now wellknown that it is not necessary to have many particles per cell, so long as in Eq. (1) we use instead of the traditional (but wrong) . Coupled with the Poisson distribution of this gives a constant average total collision rate. However, using very small cells leads to very large variability of collision rates from cell to cell and thus spatial localization of momentum transfer during each time step. Namely, with very small cells one rarely has two particles and thus most of the collisions will occur in the few cells that happen to be densely populated.
We have aimed at trying to mimic what would happen in an MD simulation in the DSMC one. In an MD simulation particles collide if their distance is equal to the particle diameter . Therefore, we have aimed at keeping the cell size at a couple of diameters, . At typical hardsphere liquid densities this leads to , which seems appropriate in that it allows enough collision partners for most of the particles but still localized the momentum transfer sufficiently. For very small mean free paths DSMC does not distinguish velocity gradients at length scales smaller than the cell size and, in a longtime average sense, localizes the velocity gradients at cell interfaces DSMC_CellSizeError (). We will assume that, for problems of interest to us, the structure of the fluid and flow at lengthscales comparable to (and thus ) is unimportant, and verify this by explicit comparisons to MD.
When the cell size is chosen such that and the time step is reasonable, , Eq. (1) gives collision frequencies that are sufficiently high so that almost all particles suffer at least one collision every time step, and typically more than one collision. The effect of such repeated collisions is to completely thermalize the flow to a local equilibrium (i.e., local Maxwellian), and we have observed that further increasing the collision frequency does not change the effective viscosity (we do not have a theoretical understanding of this behavior DSMC_CellSizeError ()). For greater efficiency, we have chosen to use the lowest collision rate (for a given timestep) that still achieves a viscosity that is as high as using a very high collision rate. We find that this is typically achieved when each particle suffers about half a collision or one collision each timestep DSMC_GammaBounds (). Appendix B describes some multiparticle collision variants that may be more appropriate under different conditions.
iii.4.4 DSMC without Hydrodynamics
The solvent exerts three primary effects on polymers in flow: (1) stochastic forces due to fluctuations in the fluid (leading to Brownianlike motion), (2) (local) frictional resistance to bead motion (usually assumed to follow Stokes law), and (3) hydrodynamic interactions between the beads due to perturbations of the flow field by the motion of the beads. Brownian dynamics, the most common method for simulating the behavior of polymers in flow, coordinates the first two effects via the fluctuationdissipation theorem, and potentially adds the third one via approximations based on the Oseen tensor (neglecting the possibility of large changes to the flow field due to the moving beads). By turning off local momentum conservation one can eliminate all hydrodynamic interactions, and thus test the importance of the coupling between polymer motion and flow.
Yeoman’s et al. PolymerCollapse_Yeomans (); ShearThinning_Yeomans () have implemented a nohydrodynamics variant of the MPCD algorithm by randomly exchanging the velocities between all particles at each time step (thus preserving momentum and energy globally, but not locally). In the presence of a background flow, such as shear, only the components of the velocities relative to the background flow are exchanged. We have implemented a nohydrodynamics variant of DSMC by neglecting momentum conservation in the usual stochastic binary collisions ^{11}^{11}11In this implementation switching hydrodynamics off becomes an alternative branch localized in the binary collision routine and the algorithm is otherwise unchanged.. Specifically, if particles and collide, the postcollisional velocity of is set to be the same magnitude as that of but with a random orientation, and vice versa (this conserves energy but not momentum). If the boundary conditions specify a background flow such as a uniform shear the flow velocity is evaluated at the center of the DSMC cell and the collisions are performed in the frame moving with that velocity. This forces the average velocity profile to be as specified by the boundary conditions, but does not allow for perturbations to that profile due to hydrodynamic effects.
Iv Performance Improvement
It is, of course, expected that the DSMC algorithm will give a performance improvement over MD. However, to make an impact on realworld problems this performance gain must be an order of magnitude or more improvement. Indeed, we find that SEDMD with adaptive boundary conditions can be up to two hundred times faster than EDMD under certain conditions. Note also that it is wellknown that EDMD is already significantly faster than TDMD (depending on the density), although such a comparison is somewhat unfair since the hardcore interaction potentials are very simple by design. It is important to note that this algorithm is serial, and we do not consider or use any parallelization. Because of the inherent simplicity and thus efficiency of the algorithm, however, it is possible to study time scales and system sizes as large or larger than parallel simulations described in the literature so far. The combined timedriven DSMC with eventdriven MD algorithm can be parallelized using traditional techniques from TDMD if proper domain partitioning is constructed, so that each eventdriven region is processed by a single processor ^{12}^{12}12Achieving good load balancing will be easiest for systems containing multiple polymer chains..
As model problem we study a tethered polymer in three dimensions. The solvent density was chosen to be typical of a moderately dense hardsphere liquid. The performance and optimal choice of parameters depends heavily on the size of the beads relative to the size of the solvent particles for both MD and the hybrid algorithm. Realistically, beads (meant to represent a Kuhn segment) should be larger than the solvent molecules^{13}^{13}13For example, in Ref. FluctuatingHydroMD_Coveney () an appropriate bead size for polyethylene is estimated at , and for DNA (a much stiffer molecule with large persistence length) at .. This of course dramatically increases the computational requirements due to the increase in the number of solvent particles (and also makes neighbor searching more costly). For this reason, most MD simulations reported in the literature use solvent particles that are equivalent, except for the chain connectivity, to the solute particles.
Our first test problem is for a chain of large beads, each about times larger (in volume and in mass) than the solvent particles, in a box of size polymer lengths, for a total of about particles. For the DSMC simulations, we did not use BSCs (bounding sphere complexes Event_Driven_HE ()), and therefore the neighbor search had to include nextnearest neighbor cells as well (i.e., ). For the corresponding MD simulations, BSCs were used. Under these conditions, DSMC outperformed MD by a factor of . If adaptive open BCs were used with , giving about particles (the exact number changes with polymer conformation), the speedup was . While this may seem an unfair comparison, it is important to point out that we do not even know how to implement an adaptive simulation domain in pure EDMD.
The second test problem was for a chain of beads which were identical to the solvent particles, except for the added chain tethers. The number of particles in the simulation cell was thus much smaller, , and . Adaptive BCs with reduce the simulation domain to particles. For these parameters DSMC with adaptive BCs was about times faster than full MD. Table 1 summarizes the large performance gains of SEDMD relative to traditional EDMD.
Standard BCs  Adaptive BCs  

Large beads  35  180 
Small beads  20  30 
One of the fundamental problems with multiscale modeling is that typically the majority of the simulation time is spent in the finest model since it is difficult to match the time scales of the coupled components Trebotich_Penalty (). For example, MD simulations are so expensive that coupling them to almost any meso or macroscopic solver leads to simulation times limited by that of MD simulations (albeit of a much smaller system). By virtue of the fast microscopic algorithm (EDMD instead of TDMD) and the efficient coupling, our method spends comparable amounts of computation on the solute (and immediately surrounding solvent) and the solvent particles. For the DSMC run with adaptive open BCs and large beads, about of the time was spent in manipulation of near neighbor lists. Most of the remaining time was spent inside the routine that takes a DSMC timestep, and actual processing of DSMC collisions (both trial and real) occupied about of the computation time. For small beads, the majority of the time, , was spent in the DSMC timestep routine, and processing of DSMC binary collisions occupied about of the computation time.
V Tethered Polymer in Shear Flow
In this section results are presented for a tethered polymer chain in uniform shear in three dimensions. The linear chain is in a good solvent and is attached at one end to a hard wall, as represented by the plane . A linear velocity profile along the axis is imposed sufficiently far from the chain. This problem was first studied experimentally by Doyle et al. TetheredPolymer_Experiment_PRL () and since then numerous computational studies have investigated various aspects of the problem TetheredPolymer_HybridMD (); TetheredPolymer_FullMD (); TetheredPolymer_Cyclic_PRL (); TetheredPolymer_Cyclic_AIP (); TetheredDNA_FullMD (). We will focus on the dynamics of the chain at low to medium flow rates (Weissenberg numbers) because we wanted to verify that our polymer and solvent model can correctly reproduce nontrivial dynamics.
v.1 Background
The properties of a linear polymer in shear flow can be related to the dimensionless Weissenberg number , where is the relaxation time of the polymer chain when there is no shear. When the flow barely affects the polymer, contrary to when . Different models have given similar properties for the same Weissenberg number.
The original experimental study of tethered polymers TetheredPolymer_Experiment_PRL () observed what was termed “cyclic dynamics” of the chains. Specifically, the following cycle was proposed. When the polymer moves too far from the wall, presumably by an unusual fluctuation, it experiences a stronger flow and is stretched. A torque develops that then pushes the chain closer to the wall, where it can contract again due to the weaker flow near the wall. The cycle then repeats. Experiments TetheredPolymer_Experiment_PRL () did not identify clear periodicity of this motion. Subsequent computational studies have looked for such a characteristic period for this cycling motion.
The MD study in Ref. TetheredPolymer_FullMD () examined the crosscorrelation function , where measures the extension of the polymer along the flow, and measures the angle of the chain with respect to the hard wall. No exact definitions of or were given even though there are several possibilities. One can use the difference between the maximal and the minimal bead positions as a measure of the extension along a given axes. Optionally, one can simply use the maximal position, or one can use the position of the last bead. Similarly, the angle of the polymer can be based on a linear fit to the shape of the chain, on the position of the center of mass, the asymmetry of the gyration tensor PolymerTumbling_PRL (), or the position of the last bead. We have examined various choices and have found little qualitative difference between the different choices. We have found the position of the end bead to be the best option and will also measure the angle .
The authors of Ref. TetheredPolymer_FullMD () found that develops a peak at positive time for sufficiently large Wi numbers (). This was interpreted as supporting the existence of a critical Weissenberg number Wi where the flow effect on the polymer dynamics changes qualitatively. It was also found that decreases with increasing Wi and the height of the peak increases. It is important to note that was found to be comparable to the relaxation time of the polymer . Additionally, the internal relaxation time was found to decrease with increasing Wi, in agreement with theoretical predictions.
A subsequent study which used a hybrid MD/CFD model, and also a (freedraining) Brownian dynamics model, claimed to observe periodic oscillations in the crosscorrelation function between the extensions along the flow and along the shear direction (i.e., perpendicular to the wall), TetheredPolymer_Cyclic_PRL (); TetheredPolymer_Cyclic_AIP (). However, the period of oscillation was found to be an order of magnitude larger than the internal relaxation time, as revealed by a small peak in the power spectral density of . A similar claim was made in Ref. PolymerTumbling_PRL () based on of the polymer angle autocorrelation function ^{14}^{14}14The PSD is equivalent to the Fourier spectrum power of the angle trace based on the convolution theorem. for both a free polymer in unbounded shear flow and a tethered polymer in shear flow. No results for the shorttime crosscorrelation functions were reported in either of these studies making it difficult to reconcile the results obtained from PSDs with those in Ref. TetheredPolymer_FullMD ().
Most experimental and computational studies of the dynamics of polymers in shear flow have been for free chains in unbounded flow PolymerTumbling_Review (). In that problem, for , it is possible to identify a welldefined “tumbling” event as the polymer rotates. The frequency of such tumbling times can be measured by visual inspection and have been compared to the computed location of the peak in the PSDs PolymerTumbling_PRL (); PolymerTumbling_PDFs (). The good match has thus been taken as an indicator that PSDs peaks can be used to determine characteristic tumbling times and the same methodology has been applied to a tethered polymer as well. However, for the case of a tethered chain it is not easy to identify a periodic event such as a specific rare fluctuation. Therefore, it is not surprising that we do not confirm the existence of a characteristic time that is an order of magnitude larger than the internal relaxation time. One must here distinguish between “cyclic” (repetitive) events and periodic events. A Poisson time process of rate has a welldefined time scale , however, the occurrence of such events is not periodic; the delay between successive events is exponentiallydistributed. In Ref. PolymerTumbling_PDFs () such an exponential distribution is proposed even for the delay between successive tumbling events for a free chain in unbounded flow. The PSD of such a process is expected to be that of white noise (i.e., flat) for frequencies small compared to , and typically a powerlaw decay for larger frequencies (gray noise). The occurrence and shape of any local maxima (peaks) or frequencies comparable to depends on the exact nature of the correlations at that time scale.
v.2 Model Parameters
As explained in IV, we have made several runs for different polymer lengths and also bead sizes. One set of runs used either or large beads each about 10 times larger than a solvent particle, using DSMC with or without hydrodynamics (see Section III.4.4) for the solvent. Another set of runs used either or small beads each identical to a solvent particle, using DSMC or pure MD for the solvent. The beads were rough in the sense that noslip conditions were applied for the solventsolute interface (see Section III.4.1). All of the runs used open boundary conditions (see Section III.3), and the typical halfwidth of the interior region was or cells around the polymer chain. The difference in the results (such as relaxation times) between these runs and runs using or runs using periodic BCs were negligible for the chain sizes we studied ^{15}^{15}15It is expected that using a small would truncate the (longranged) hydrodynamic interactions and thus increase the relaxation time. We observe such effects for the chains, however, the effect is too small compared to the statistical errors to be accurately quantified.. The solvent was a hardsphere MD or DSMC fluid with volume fraction , which corresponds to a moderately dense liquid (the melting point is ). The runs were run for with , and such a run takes about 6 days on a single 2.4GHz DualCore AMD Opteron processor. Even for such long runs the statistical errors due to the strong fluctuations in the polymer conformations are large, especially for correlation functions at long time lags .
v.3 Relaxation Times
The relaxation time of the polymer is welldefined only for linear models. It is often measured by fitting an exponential to the autocorrelation function of the endtoend vector , where denotes the position of the th bead PolymerDynamics_Review (). We will separately consider the different components of the endtoend vector and fit an exponential ^{16}^{16}16The initial relaxation of the various autocorrelation functions is faster than exponential, and the statistical error at longer times is large even for long runs. We therefore fit the exponentials to the portion of the curves at small times, when . The fits are not perfect and there are large statistical errors depending on the length of the run and the number of samples used to average , and the relaxation times (and thus Weissenberg numbers) we quote should be taken as approximate. to the , and autocorrelations functions to obtain the relaxation times , and as a function of Wi. We find that is always the largest, especially for large Wi (for , by symmetry), and is always smaller by at least a factor of two ^{17}^{17}17This is because of the constraint that the polymer chain must be above the plane at all times, which reduces the available configuration space., even for , as illustrated in the inset in Fig. 2. We take as the definition of the polymer relaxation time.
The relaxation times we observe for are consistent with what is predicted from theoretical considerations, , where is the viscosity and is the effective bead radius. Using the viscosity (based on Enskog theory) of the MD liquid ^{18}^{18}18Direct measurements of the viscosity of the DSMC liquid show that it has viscosity rather close to that of the corresponding MD liquid for the specific parameters we use. and the tether length as , we calculated for the case of with large beads, to be compared to the numerical results from DSMC . The MD runs for the case of large beads are not long enough to determine the relaxation time accurately. We expect that the difference between MD and DSMC will become more pronounced for smaller beads, and indeed, for we obtain . Turning hydrodynamics off in DSMC extends the relaxation times (and also the collapse times for an initially stretched polymer) by a factor of , as already observed using MPCD PolymerCollapse_Yeomans () and as predicted by Zimm theory ^{19}^{19}19It is difficult to directly compare DSMC with and without hydrodynamics since switching hydrodynamics off, in our model, affects the friction force between the beads and the solvent. This is unlike the models were the friction force is an added phenomenological term that has an adjustable coefficient. . Figure 2 illustrates the dependence of on Wi, and similarly for the and directions. Quantitatively similar (but not identical) results are observed independently of the details of the polymer model and even the existence of hydrodynamic relaxations.
v.4 Cyclic Dynamics
We now turn our attention to crosscorrelations between polymer extensions in the and directions. We have found that the crosscorrelations lags are most visible in the and positions of the last bead, . Our results for are shown in Fig. 3, along with as an inset. The results for compare well with those in Ref. TetheredPolymer_FullMD (), although we see the secondary peak developing at somewhat lower Wi. We do not see any evidence for the existence of a critical Wi: There are peaks at both positive and negative time in for all Wi. Some crosscorrelations, such as , have a large positive or negative cusp at the origin at and it is this cusp that masks the peaks at nonzero lags for small Wi.
In Fig. 4 we compare at for several different models ^{20}^{20}20The Weissenberg numbers were calculated after the runs were completed and therefore the different runs are not at the exact same Wi number. and see a good match, even for the DSMC runs ignoring hydrodynamics (momentum conservation). This indicates that the dynamics of the chains is primarily driven by the competition between the internal stochastic motion (entropy) and the external forcing due to the shear, and not hydrodynamic interactions between the beads or the effect of the motion of the chain on the flow.
We do not discuss the origin and locations of the peaks in the crosscorrelation functions in detail in this work. These peaks are indicative of the existence of a correlated motion in the plane, but do not uniquely identify that motion. An important question to address is the existence of a time scale other than the internal relaxation time ). In Fig. 5 we show a renormalized crosscorrelation function
in an unsuccessful attempt to collapse the data for different Wi. While the match is not perfect the picture does not point to the existence of a time scale shorter than . We also do not see any convincing evidence for coherent and reproducible correlations on time scales significantly larger than , even in various power spectral densities. Our results do not rule out the possibility of a repetitive motion of the chain with widely varying cyclic times (e.g., exponential tail) but we have not observed any direct evidence for such cycling either. We will report more detailed results on the dynamics of tethered polymer chains along with comparisons with Brownian dynamics and LatticeBoltzmann in future work.
Vi Conclusions
We presented a stochastic eventdriven molecular dynamics (SEDMD) algorithm that combines hardsphere eventdriven molecular dynamics (EDMD) with direct simulation Monte Carlo (DSMC), aimed at simulating flow in suspensions at the microscale. The overall algorithm is still eventdriven, however, the DSMC portion of the algorithm can be made timedriven for increased efficiency. The fundamental idea is to replace the deterministic (MDlike) interactions between particles of certain species with a stochastic (MClike) collision process, thus preserving the phase space dynamics and conservation laws but ignoring the liquid structure. The SEDMD methodology correctly reproduces hydrodynamic behavior at the macroscale but also correctly represents fluctuations at the microscale. A similar algorithm has been proposed using timedriven (softparticle) MD and a multiparticle collision variant of DSMC DSMC_MPCD_MD_Kapral ().
As an application of such a methodology we have considered the simulation of polymer chains in a flowing solution, and in particular, a polymer tethered to a hard wall and subject to shear flow. We have implemented open boundary conditions that adaptively adjust the simulation domain to only focus on the region close to the polymer chain(s). The algorithm is found to be efficient even though it is not parallelized, and it is found to reproduce results obtained via molecular dynamics and other algorithms in the literature, after adjusting for the correction to transport coefficients and compressibility of the DSMC fluid relative to the MD fluid.
We studied the dynamics of a tethered polymer subject to pure shear and found consistent results between MD and DSMC and also previous TDMD studies. We find that neither the size of the polymer beads relative to the solvent particles, nor the correct representation of the hydrodynamic interactions in the fluid, qualitatively alter the results. This suggests that fluctuations dominate the dynamic behavior of tethered polymers, consistent with previous studies. Our results do not find periodic motion of the polymer and show that the crosscorrelation between the polymer extensions along the flow and shear directions shows a doublepeak structure with characteristic time that is comparable to the relaxation time of the polymer. This is in contrast to other works that claim the existence of a new timescale associated with the cyclic motion of the polymer. We will investigate these issues further and compare with Brownian dynamics and LatticeBoltzmann simulations in future work.
We expect that this and related algorithms will find many applications in micro and nanofluidics. In particular, the use of DSMC instead of expensive MD is suitable for problems where the detailed structure and chemical specificity of the solvent do not matter, and more general hydrodynamic forces and internal fluctuations dominate. Using a continuum approach such as NavierStokes (NS) equations for the solvent is questionable at very small length scales. Furthermore, the handling of singularities and fluctuations is not natural in such PDE methods and various approximations need to be evaluated using particlebased methods. Since the meshes required by continuum solvers for microflows are very fine, it is expected that the efficiency of particle methods will be comparable to PDE solvers. Nevertheless, algorithms based on fluctuating hydrodynamics descriptions will be more efficient when fluctuations matter. Comparisons and coupling of DSMC to fluctuating NS solvers is the subject of current investigations FluctuatingHydro_AMAR ().
Vii Acknowledgments
This work was performed under the auspices of the U.S. Department of Energy by the University of California Lawrence Livermore National Laboratory under Contract No. W7405Eng48 (UCRLJRNL233235).
Appendix A AED variants of DSMC
In this Appendix we discuss a fully asynchronous eventdriven (AED) implementation of DSMC. The advantage of asynchronous algorithms is that they do not introduce any artificial time scales (such as a time step) into the problem AED_Review (). We have validated that the AED algorithm produces the same results as the timedriven one by comparing against published DSMC results for plane Poiseuille flow of a rare gas DSMC_PlanePoiseuille (). We have also implemented traditional timedriven (TD) DSMC and find identical results when the time step is sufficiently small. We find that the eventdriven algorithm is almost an order of magnitude slower than the time driven one at higher densities, and only becomes competitive at very low densities (which is the traditional domain of interest for DSMC). The overhead of the AED algorithm comes from the need to repredict the next event and update the event queue whenever a particle suffers a DSMC collision. This cost is in addition to the equivalent cost in the timedriven algorithm, namely, moving the particles forward in time and colliding them.
The AED algorithm introduces a new type of event, a stochastic (trial) collision between two DSMC particles that are in the same cell (see Section III.2). These trial collisions occur in a given cell as a Poisson process (i.e., exponentially distributed waiting time) with a rate given by Eq. (1). There are several approaches to scheduling and processing DSMC collisions directly borrowed from algorithms for performing Kinetic Monte Carlo simulations (which are synchronous eventdriven algorithms AdvancedKMC_Tutorial ()). The simplest, and in our experience, most efficient, approach to AED DSMC is to use cell rejection to select a host cell for the stochastic collisions. The rate of DSMC collisions is chosen according to the cell with maximal occupancy , . The randomly chosen cell of occupancy is accepted with probability and a random pair of particles and are chosen from . Since the DSMC fluid is perfectly compressible, the maximal cell occupancy can be quite high for very large systems, and this leads to decreasing cell acceptance probability as the size of the system increases.
One can avoid cell rejections altogether. The first option is to associate stochastic collisions with cells and schedule one such event per cell. The event time is easily predicted at any point in time to occur at time , where is a uniform random deviate in . These event times are put in the event queue (which may be separate from the particle one and then the two queues may be merged only at the top). The event times (and thus the cell queue) need to updated whenever a cell occupancy changes, that is, whenever a cell transfer is processed. This makes this algorithm inefficient. Another alternative is to recognize that the sum of a set of independent Poisson processes is a Poisson process with a rate that is the sum of the individual rates, . That is, DSMC collisions occur in the system as a Poisson process with rate . When processing such an event one has to first choose the cell with probability , which requires some additional data structures to implement efficiently AdvancedKMC_Tutorial (). For example, the cells could be grouped in lists based on their occupancy and then an occupancy chosen first (with the appropriate weight), followed by selection of a cell with that particular occupancy.
Finally, it is also possible to use a mixture of the asynchronous and timedriven variants of DSMC. The asynchronous algorithm (e.g., based on cell rejection) can be used for DSMC particles in eventdriven cells, and the timedriven one elsewhere. This may be useful in situations where the timescale of the eventdriven component (i.e., the solute and nearby solvent particles) is significantly smaller than the time step and thus time stepping would lead to discretization artifacts.
In the AED variant of DSMC constant pressure BCs (see Section III.4.2) can be implemented by adding a new type of acceleration event. When such an event is processed, all of the particles are brought to the same point in time (synchronized), the velocities of each DSMC particle is incremented by (here is the elapsed time since the last acceleration event), and the event handling is restarted. The acceleration events occur as a Poisson process with a suitably chosen rate, for example, ensuring that the average or maximal change in velocity is a fraction of the average particle velocity. Note that the choice of this acceleration rate introduces an artificial time constant in the algorithm similar to the time step in timedriven DSMC.
Appendix B MultiParticle Collisions in DSMC
Under dense liquid conditions, DSMC binary collisions are so numerous (see Section III.4.3) that the velocities of the particles are effectively thermalized to the local Maxwell distribution. We have implemented a variant DSMC algorithm in which at every time step the velocities of all of the particles are redrawn from a local Maxwellian, preserving the total linear momentum and energy in each cell DSMC_Pullin (). We found that this variant of DSMC is less efficient than and behaves similarly to the usual binarycollision DSMC. Reference DSMC_TRMC () describes a more general algorithm (TRMC) that combines binary collisions for a subset of the particles with drawing from a local Maxwellian for the remainder of the particles, and under dense liquid conditions this typically degenerates to complete randomization of all of the velocities at every time step. Until a theoretical framework is established for the behavior of DSMClike algorithms at high densities the classical DSMC algorithm seems to be the best alternative in terms of simplicity, efficiency, and theoretical foundation. The effect of collision rules and cell size on multiparticle collision dynamics has been studied and it was found that increased collisional viscosity is desirable for achieving realistic convection to diffusion ratios DSMC_MPCD_Gompper (); MultiparticleDSMC_Polymers ().
We mention that, strictly speaking, we should use as in Eq. (1) not the volume of the cell, but the unoccupied cell volume (that is, the portion of the cell not covered by nonDSMC particles) ^{21}^{21}21Our implementation makes the additional approximation that nonDSMC particles are also counted in in Eq. (1), instead of keeping a separate count of just the DSMC particles inside each cell. If the polymer beads are larger than a cell than this approximation does not matter since no cell can contain the centroid of both a DSMC and nonDSMC particle.. It is however difficult to dynamically maintain an accurate estimate of the cell coverage, and the complication does not appear to be worth the implementation complexity. In particular, an approximation is already made in neglecting the structure of the fluid near a polymer bead (i.e., the solvation layer), and furthermore, the majority of the cells that are partially covered by a polymer bead will be entirely or almost entirely covered so that they would at most contain a single DSMC particle, in which case the probability of a DSMC collision would be very low anyway. Finally, as explained in Section III.4.3, the exact collision frequency does not really matter. In the context of multiparticle collision dynamics, Ref. DSMC_MPCD_CylinderFlow () proposes the use of virtual particles filling the partiallyfilled cells as a way to achieve more accurate stick boundary conditions.
References
 [1] W. E. Alley, P. Covello, and B. J. Alder. Complex flows by nanohydrodynamics. Mol. Phys., 102(19):0026–8976, 2004.
 [2] M. Ripoll, K. Mussawisade, R. G. Winkler, and G. Gompper. LowReynoldsnumber hydrodynamics of complex fluids by multiparticlecollision dynamics. Europhys. Lett., 68(1):106, 2004.
 [3] T. M. Squires and S. R. Quake. Microfluidics: Fluid physics at the nanoliter scale. Rev. Mod. Phys., 77(3):977, 2005.
 [4] E. S. G. Shaqfeh. The dynamics of singlemolecule DNA in flow. Journal of NonNewtonian Fluid Mechanics, 130:1–28, 2005.
 [5] G. De Fabritiis, M. Serrano, R. DelgadoBuscalioni, and P. V. Coveney. Fluctuating hydrodynamic modeling of fluids at the nanoscale. Phys. Rev. E, 75(2):026307, 2007.
 [6] G. W. Slater, Y. Gratton, M. Kenward, L. McCormick, and F. Tessier. Deformation, Stretching, and Relaxation of SinglePolymer Chains: Fundamentals and Examples. Soft Materials, 2:155–182, 2004.
 [7] P. S. Doyle, B. Ladoux, and J.L. Viovy. Dynamics of a Tethered Polymer in Shear Flow. Phys. Rev. Lett., 84(20):4769–4772, 2000.
 [8] S. Barsky, R. DelgadoBuscalioni, and P. V. Coveney. Comparison of molecular dynamics with hybrid continuum–molecular dynamics for a single tethered polymer in a solvent. J. Chem. Phys., 121(5):2403–2411, 2004.
 [9] Y. Gratton and G. W. Slater. Molecular dynamics study of tethered polymers in shear flow. Eur. Phys. J. E, 17:455–465, 2005.
 [10] R. DelgadoBuscalioni. Cyclic Motion of a Grafted Polymer under Shear Flow. Phys. Rev. Lett., 96(8):088303, 2006.
 [11] G. M Wang and W. C Sandberg. Nonequilibrium allatom molecular dynamics simulations of free and tethered DNA molecules in nanochannel shear flows. Nanotechnology, 18(13):135702, 2007.
 [12] C. M. Schroeder, R. E. Teixeira, E. S. G. Shaqfeh, and S. Chu. Characteristic Periodic Motion of Polymers in Shear Flow. Phys. Rev. Lett., 95(1):018301, 2005.
 [13] R. DelgadoBuscalioni. Dynamics of a Single Tethered Polymer under Shear Flow. AIP Conference Proceedings, 913(1):114–120, 2007.
 [14] R. M. Jendrejack, J. J. de Pablo, and M. D. Graham. Stochastic simulations of DNA in flow: Dynamics and the effects of hydrodynamic interactions. J. Chem. Phys., 116(17):7752–7759, 2002.
 [15] J. P. HernandezOrtiz, J. J. de Pablo, and M. D. Graham. Fast Computation of ManyParticle Hydrodynamic and Electrostatic Interactions in a Confined Geometry. Phys. Rev. Lett., 98(14):140602, 2007.
 [16] O. B. Usta, A. J. C. Ladd, and J. E. Butler. LatticeBoltzmann simulations of the dynamics of polymer solutions in periodic and confined geometries. J. Chem. Phys., 122(9):094902, 2005.
 [17] N. Sharma and N. A. Patankar. Direct numerical simulation of the Brownian motion of particles by using fluctuating hydrodynamic equations. J. Comput. Phys., 201:466–486, 2004.
 [18] D. Trebotich, G. H. Miller, P. Colella, D. T. Graves, D. F. Martin, and P. O. Schwartz. A Tightly Coupled ParticleFluid Model for DNALaden Flows in Complex Microscale Geometries. Comp. Fluid Solid Mech., pages 1018–1022, 2005.
 [19] G. Giupponi, G. De Fabritiis, and P. V. Coveney. Hybrid method coupling fluctuating hydrodynamics and molecular dynamics for the simulation of macromolecules. J. Chem. Phys., 126(15):154903, 2007.
 [20] N. Kikuchi, J. F. Ryder, C. M. Pooley, and J. M. Yeomans. Kinetics of the polymer collapse transition: The role of hydrodynamics. Phys. Rev. E, 71(6):061804, 2005.
 [21] D. Trebotich, G. H. Miller, and M. D. Bybee. A Hard Constraint Algorithm to Model Particle Interactions in DNAladen Flows. Nanoscale and Microscale Thermophysical Engineering, 11(1):121–128, 2007.
 [22] K. Mussawisade, M. Ripoll, R. G. Winkler, and G. Gompper. Dynamics of polymers in a particlebased mesoscopic solvent. J. Chem. Phys., 123(14):144905, 2005.
 [23] S. H. Lee and R. Kapral. Mesoscopic description of solvent effects on polymer dynamics. J. Chem. Phys., 124(21):214901, 2006.
 [24] D. Trebotich, G. H. Miller, and M. D. Bybee. A Penalty Method to Model Particle Interactions in DNAladen Flows. To appear in J. Nanosci. Nanotech., 2007.
 [25] C. Aust, M. Kroger, and S. Hess. Structure and dynamics of dilute polymer solutions under shear flow via nonequilibrium molecular dynamics. Macromolecules, 32(17):5660–5672, 1999.
 [26] J. B. Bell, A. Garcia, and S. A. Williams. Numerical Methods for the Stochastic LandauLifshitz NavierStokes Equations. Phys. Rev. E, 76:016708, 2007.
 [27] S. A. Williams, J. B. Bell, and A. L. Garcia. Algorithm Refinement for Fluctuating Hydrodynamics. Submitted, 2007.
 [28] K. Kadau, C. Rosenblatt, J. L. Barber, T. C. Germann, Z. Huang, P. Carles, and B. J. Alder. The importance of fluctuations in fluid mixing. PNAS, 104(19):7741–7745, 2007.
 [29] F. Xijunand N. PhanThien, S. Chen, X. Wu, and T. Y. Ng. Simulating flow of DNA suspension using dissipative particle dynamics. Physics of Fluids, 18(6):063102, 2006.
 [30] F. J. Alexander and A. L. Garcia. The Direct Simulation Monte Carlo Method. Computers in Physics, 11(6):588–593, 1997.
 [31] F. Baras, M. Malek Mansour, and A. L. Garcia. Microscopic simulation of dilute gases with adjustable transport coefficients. Phys. Rev. E, 49(4):3512–3515, 1994.
 [32] F. J. Alexander, A. L. Garcia, and B. J. Alder. A Consistent Boltzmann Algorithm. Phys. Rev. Lett., 74(26):5212–5215, 1995.
 [33] A. L. Garcia and W. Wagner. The limiting kinetic equation of the Consistent Boltzmann Algorithm for dense gases. J. Stat. Phys., 101:1065–86, 2000.
 [34] J. F. Ryder and J. M. Yeomans. Shear thinning in dilute polymer solutions. The Journal of Chemical Physics, 125(19):194906, 2006.
 [35] Jose Maria Montanero and Andres Santos. Simulation of the Enskog equation [agrave] la Bird. Phys. Fluids, 9(7):2057–2060, 1997.
 [36] A. Frezzotti. A particle scheme for the numerical solution of the Enskog equation. Phys. Fluids, 9(5):1329–1335, 1997.
 [37] T. Ihle, E. Tüzel, and D. M. Kroll. Consistent particlebased algorithm with a nonideal equation of state. Europhys. Lett., 73:664–670, 2006.
 [38] B. J. Alder and T. E. Wainwright. Studies in molecular dynamics. I. General method. J. Chem. Phys., 31:459, 1959.
 [39] A. Donev, S. Torquato, and F. H. Stillinger. Neighbor List CollisionDriven Molecular Dynamics Simulation for Nonspherical Particles: I. Algorithmic Details II. Applications to Ellipses and Ellipsoids. J. Comp. Phys., 202(2):737–764, 765–793, 2005.
 [40] S. W. Smith, C. K. Hall, and B. D. Freeman. Molecular Dynamics for Polymeric Fluids Using Discontinuous Potentials. J. Comp. Phys., 134(1):16–30, 1997.
 [41] S. B. Opps, J. M. Polson, and N. A. Risk. Discontinuous molecular dynamics simulation study of polymer collapse. J. Chem. Phys., 125(19):194904, 2006.
 [42] S. Peng, F. Ding, B. Urbanc, S. V. Buldyrev, L. Cruz, H. E. Stanley, and N. V. Dokholyan. Discrete molecular dynamics simulations of peptide aggregation. Phys. Rev. E, 69(4):041908, 2004.
 [43] H. D. Nguyen and C. K. Hall. Molecular Dynamics Simulations of Fibril Formation by Random Coil Peptides. Proc. Natl. Acad. Sci., USA 101:16180, 2004.
 [44] A. Donev. Asynchronous EventDriven Particle Algorithms. In K. S. Perumalla, editor, Proceedings of the 21st International Workshop on Principles of Advanced and Distributed Simulation. IEEE Computer Society, June 2007.
 [45] A. L. Garcia and W. Wagner. Time step truncation error in direct simulation Monte Carlo. Phys. Fluids, 12:2621–2633, 2000.
 [46] A. L. Garcia and W. Wagner. Generation of the Maxwellian Inflow Distribution. J. Comput. Phys., 217:693–708, 2006.
 [47] A. L. Garcia, J. Bell, Wm. Y. Crutchfield, and B. J. Alder. Adaptive Mesh and Algorithm Refinement using Direct Simulation Monte Carlo. J. Comp. Phys., 154:134–155, 1999.
 [48] S. Wijesinghe, R. Hornung, A. L. Garcia, and N. Hadjiconstantinou. Threedimensional Hybrid ContinuumAtomistic Simulations for Multiscale Hydrodynamics. Journal of Fluids Engineering, 126:768–777, 2004.
 [49] A. Lamura, G. Gompper, T. Ihle, and D. M. Kroll. Multiparticle collision dynamics: Flow around a circular and a square cylinder. Europhys. Lett., 56(3):319–325, 2001.
 [50] Y. Inoue, Y. Chen, and H. Ohashi. Development of a Simulation Model for Solid Objects Suspended in a Fluctuating Fluid. J. Stat. Phys., 107(112):85–100, 2002.
 [51] F. J. Uribe and A. L. Garcia. Burnett description for plane poiseuille flow. Phys. Rev. E, 60(4):4063–4078, 1999.
 [52] F. Alexander, A. L. Garcia, and B. J. Alder. Cell Size Dependence of Transport Coefficients in Stochastic Particle Algorithms. Phys. Fluids, 10:1540, 1998. Erratum: Phys. Fluids, 12 731 (2000).
 [53] T. Bartel, T. Sterk, J. Payne, and B. Preppernau. DSMC simulation of nozzle expansion flow fields. AIAA Paper No. 942047, 1994.
 [54] S. Gerashchenko and V. Steinberg. Statistics of Tumbling of a Single Polymer Molecule in Shear Flow. Phys. Rev. Lett., 96(3):038304, 2006.
 [55] Aleksandar Donev. Asynchronous eventdriven particle algorithms. In PADS ’07: Proceedings of the 21st International Workshop on Principles of Advanced and Distributed Simulation, pages 83–92, Washington, DC, USA, 2007. IEEE Computer Society.
 [56] M. A. Novotny. A Tutorial on Advanced Dynamic Monte Carlo Methods for Systems with Discrete State Spaces, volume IX of Annual Reviews of Computational Physics, pages 153–210. World Scientific, Singapore, 2001.
 [57] D. I. Pullin. Direct simulation methods for compressible inviscid idealgas flow. J. Comp. Phys., 34:231–244, 1980.
 [58] L. Pareschi and S.Trazzi. Numerical solution of the Boltzmann equation by Time Relaxed Monte Carlo (TRMC) methods. Int. J. Numer. Meth. Fluids, 48:947–983, 2005.