A Fast Algorithm for the Discrete Element Method by Contact Force Prediction†

C. Tokoro1 Department of Resources and Environmental Engineering, Waseda University* K. Okaya and J. Sadaki Department of Geosystem Engineering, Graduate School of Engineering, The University of Tokyo

Abstract The discrete element method (DEM) takes enormous calculation time because it requires a very small time step, one small enough to represent the large frequency in the contact dynamic model. In general, the equations of motion of particles are solved using the second-order Adams-Bashforth method, which estimates the values of contact force in the following calculation time by linear extrapolation, or by multi-step methods such as the predictor-corrector method. Inspired by these two conventional methods, we propose a Contact Force Prediction Method that makes a larger time step possible. Our method uses the predicted values of contact force at every contact point, which are exact solutions or numerical solutions of differential equations that represent two particle contacts. It has been confirmed experimentally that the proposed method gives reasonable results of packing and discharge simulations, and accelerates DEM calculation 3 8 times. Keywords: Discrete Element Method, Calculation Time, Time Step, Numerical Integration, Fast Algorithm

1. Introduction Discrete Element Method (DEM) simulations on a number of granular systems have reported positive results since this method can estimate many effects at particle level1). However, it requires enormous computation time because it needs a sufficiently small time step to follow high fluctuations in the contact dynamic model between particles. To overcome the deficit of conventional DEM, speeding up DEM is proposed in this study. There are several approaches to speeding up DEM, such as promoting efficiency in the algorithm to detect contact between particles, increase in the time step and limitation of the number

* 3-4-1, Okubo, Shinjuku-ku, Tokyo, 169-8555, Japan 1 Corresponding author TEL: +81-3-5286-3327 FAX: +81-3-5286-3491 E-mail: tokoro@aoni.waseda.jp † This report was originally printed in J. Soc. Powder Technology, Japan, 40, 236-245 (2003) in Japanese, before being translated into English by KONA Editorial Committee with the permission of the editorial committee of the Soc. Powder Technology, Japan.

of particles to calculate. In this study, we report a method to increase the time step. In DEM calculation, the time step must be set smaller than the particle size and the density becomes smaller or the spring constant becomes larger, in order to obtain a stable numerical solution. Several methods to increase the time step have been reported, such as methods to set the spring constant smaller2)-4) and the particle size larger5) than the real value. This method is an effective method of increasing the time step, but it must be used after careful examination of the influence on the phenomenon to which DEM calculation is applied, because delays in contact detection cannot be avoided. The purpose of this study is to develop an original DEM in which a large time step is possible by improving the algorithm needed to obtain the numerical solution to the contact dynamic equation between particles. We propose a Contact Force Prediction Method in which predicted values of contact force at every contact point are calculated by a more stable method than the conventional DEM and used to solve particle velocity and location. The packing (Fig. 1) and discharge (Fig. 2) system is chosen here as an example

182

KONA No.23 (2005)

5 and a2 0.1 Outline of DEM Contact force f cij in DEM is calculated by projecting on the normal direction (f nij) and the tangential direction (f sij) as shown below: f cij (t) f nij (t) f sij (t). n is the normal unit vector. (1) where (t) means a function of time t. where k is the spring constant.2 m 3 4 ksij dnij (t)dsij (t)sij (t) hsij dnij (t)vsij (t) a a (6) (7) (8) dsij (t)  vsij(t)dt  sij (t) vsij(t ) vsij(t ) ri i(t) vsij (t) vij (t) vnij (t) Fig. and it is shown that the proposed method is effective for calculating the packing ratio or the discharge rate. The normal direction force is calculated from the following equations: f nij (t) 1 2 knij dnij (t)n ij (t) hnij dnij (t)vnij (t) a a (2) (3) (4) (5) dnij (t) (ri rj ) x i(t) x j(t) vij (t) vi(t) vj(t) vnij (t) vij (t)·nij (t) nij (t). Subscript i denotes the physical values of particle i. d is the overlap between particles.of a phenomenon in which contact between many particles is important. Delays in contact search and stability of calculation When the time step is increased. (9) KONA No. In the following. r is the radius of the particle. and r denotes the relative value.23 (2005) 183 . 2 An example of discharge simulation rj j (t) nij (t). The tangential direction force is calculated from the following equations: f sij (t) 0. the effects of these factors with the conventional method are examined. DEM calculation may become incapable of reproducing a real phenomenon for the following reasons: one is delays in the contact detection and another is the error of the differential method. 2. 1 An example of packing simulation 2.3 m Hole 0. h is the damping constant. ij denotes between particles i and j. and x is the location vector. a1 and a2 are constants that depend on the type of contact dynamic model: a1 1 and a2 0 for the linear spring model and a1 1. j denotes particle j. Fig. n denotes the normal direction value.25 for the Hertz model6). v is the velocity vector.

as shown in Table 1. The particle diameter is 10 mm. has already been reported5). ∆t *( ) Fig. but the following two problems occur when the time step is large.0 0. 8). but it is usually used because of its convenience in the calculation9). because two particles that are in contact with each other continue to come close during a delay in contact detection without receiving contact force from each other.8 1 Normalized Time Step. A method whereby the calculation time is returned to the starting point when two particles make contact with each other and where particle locations are recalculated. n/n0( ) 0. (10) n0 n0 A B C where m is the friction constant.1 0. As an example. In this study. this problem is solved using the contact force prediction method described below.2 0.where s is the tangential unit vector and is the angular vector velocity. 10).3 be substituted for the physical properties of ice pellets13).0 109 0.3 106 5.3 0. whose initial velocity is zero. The contact time between particles in this simulation is 3. which overcomes this problem. we consider a difference between a real phenomenon and calculation result using the three particles below. 4 is the normalized 184 KONA No. it is difficult to theoretically decide the spring constant value except when the particle shape is a disk. As shown in Fig. 4 Normalized velocity of the 3-particle problem Table 1 Parameter settings for DEM simulation of ice pellets Properties Density Linear Spring Constant Young’s Modulus Poisson’s Ratio g Friction Coefficient Unit [kg/m ] [N/m] [N/m2] [ ] [ ] [ ] 3 Value 800 2. parameter values used in the simulation can 0. there are three particles. 12) have been chosen in this simulation.4 0 Fig. Two contacts with a small time lag may be misjudged as occurring at almost the same time because of the large time step. and the time step is set up in 1/100 of the contact time. a3 and a4 are constants that depend on the type of the contact dynamic model: a3 a4 0 for the linear spring model and a3 0.6 0.32 1. One is the large contact force between particles due to the large strain.7 10 5 seconds. In the linear spring model. Although natural gas hydrate pellets11). Fig. without rotation in a onedimensional line. B and C.2 0.5 and a4 0.4 Normalized Velocity.25 for the Mindlin model7).4 0. ∆t* in Fig.2 0. A.23 (2005) . which is not practical. B. 3. f sij (t) mij f nij (t)s ij (t). 3 3-particle contact problem C B A 0. The total tangential force is limited by the Coulomb frictional limit. This phenomenon changes the behavior of particles after they make contact because the consumption of the collision energy is different. and a linear model is used. and the particles of both ends.3 0. in which the calculation error becomes sufficiently small.2 An error from delays in contact detection Contact detection between particles may produce a maximum delay that equals the time step. A and C. This delay can be ignored when the time step is small. 2.1 0 0. collide at the same velocity with the central particle. 4 shows the calculated velocity of each particle after contact when the beginning of the contact between particle C and B is after ∆t* from the beginning of contact between particle A and B. Another problem is that several particles make contact with the same particle at same time.

5 compared with DEM with a small time step so that the effect of the contact between the other particles can be ignored. when the contact detection between particle A and B is shifted by the time step in the system in which the contact between A and B and the contact between C and B take place at the same time.1 0.15 0. location x i(t ∆ t) and angular location i(t ∆ t) after time step ∆ t are calculated using contact force fi(t) and torque Ti(t). but its calculation is sufficiently fast and stable for a large time step. On the other hand. Fig. Moreover. Neither is second-order accuracy guaranteed with the Leapfrog method. 5 is the simulation result of the normalized velocity error of particle A calculated from Fig.8 1 Normalized Time Step. The explicit method is used in DEM and particle i’s velocity vi(t ∆ t). is the general calculation procedure for DEM.6 0. Therefore. although the correction process is usually repeated until an error KONA No. 5 Error of velocity of particle A 2. ∆t *( ) Fig. several solutions with the conventional DEM are compared using the linear spring model in the system in which a particle without rotation makes contact with a wall in one dimension. angular velocity i(t ∆ t). divided by tc.2 0. the contact time between two particles: ∆t * Contact Detection YES NO ∆t . Although the predicted value of the contact force in the second-order Adams-Bashforth method is calculated by linear extrapolation. tc (11) Calculation of Contact Force The velocity is also normalized by the velocity of particle A and C before they make contact. the difference between the beginning of the contact between particles A and B and the beginning of the contact between particles C and B.05 0 0 0. the appropriate numerical solution for DEM with a large time step is not a highly precise method.2 0.23 (2005) 185 . except its shadowed area.4 0. because the particle velocity is defined by only the middle point of a time step. 4.value of ∆t. 6 Error of Velocity of Particle A . the easiest with first-order accuracy is the explicit Euler method14). the second-order Adams-Bashforth15).  (n n0)/n0 ( ) Algorithm of calculation of contact force and movement in the conventional method and the new method 0. It can be concluded that DEM with a large time step is the method that allows the error shown in Fig. there is the problem of second-order accuracy not being strictly guaranteed unless appropriate processing is performed at the discontinuous points. 6. In the following. such as the beginning and end points of the contact. unless the value of the velocity during the contact is recalculated using an appropriate method.3 Stability of calculation Fig. Calculation of Estimated Value NO All Contact Points are finished? YES NO All Particles are finished? YES Calculation of Particle Movement Fig. Although several numerical methods of solving the differential equation that represents the contact between two particles can be considered for the conventional DEM. Leapfrog16) and predictor-corrector methods10) are used with the conventional DEM.

From Equation (15). the explicit Euler method is expressed in the matrix description as follows: nk dk 1 1 where tc is the contact time when h 0.46 h nk Leap Frog dk nk 1/2 1 No Good nk h 1 fk m ∆t 1/2 * h2 k 2 p 4 m 2 xk h2 k 1 4 m 2 xk 1. time step.01 dk h ˆ nk ˆ dk PredictorCorrector 1 nk Good nk 1 3 h 1 m 2 dk 3 2 nk 1 2m dk 1 2 nk fk 1 2 1 1 2 nk fk 1 h2 k m (mnknk·k k m k m 1 kdk dk ·k k 2 1) h nk 1 1 ∆t * fk) 2 BN M2 1 p h2 2 2 (mnk kdk ) h (nk dk 1 nk 1 dk ) 2. it is common to limit the correction process to 1 or 2 times to shorten the computational time in conventional DEM. Table 2 compares the 4 above-mentioned methods. tc (16) The analysis contents are explained as follows. It is necessary to store all contact states until the calculation of predicted values for the velocity and location are finished because predicted values are calculated after all contact forces are calculated with this method. One-dimensional collision of one particle is assumed. is simplified with h.08 1) h dk 1 ( fˆk 4h2 2 (mnk·k 1 2 kdk ·k h ˆ (nk 1 n k) No Good 186 KONA No. ∆t * in Table 2 is defined as follows: ∆t * ∆t . In addition.23 (2005) . it is found that the eigenvalue of matrix A may be larger than 1. tc p m k (17) A nk dk (13) The conservation of energy in the explicit Euler method is examined when h 0 as follows: Table 2 Comparison of 4 conventional methods (k t/∆t. (14) The determinant of A is as follows: A  1 k 2 h h h. m m (15) where m is the mass of the particle. This processing increases computational time.decreases enough in the predictor-corrector method. This shows the possibility of deviating from a stability domain of a calculation with the explicit Euler method. The contact force is calculated assuming a linear spring model as follows: fk kdk hnk. and subscript i and n are omitted. subscript k is defined with k t/∆ t and ∆ t. h ∆t) Method Equations nk Explicit Euler dk 1 Calculation Stability at h 0 Energy Conser vation at h 0 Calculation Time 1 nk fk m Unstable h2 h dk k m 2 2 (mnk kdk ) 1 h nk 2-Order AdamsBashforth 1 nk 1 2 nk No Good nk 1 3 h dk 1 m 2 dk 3 2 1/2 fk 1 2 fk 1 hk(nk xk Unstable 1 nk 1 xk ) h2 k m 2 (mnk·k 1 2 kdk ·k 1) 1. (12) A 1 h h h m h 1 k m . The calculation stability in Table 2 is calculated in the same way as when h 0. In the case of the explicit Euler method.

It is necessary for the time step to set up a smaller value in the case of collision of many particles. The dynamic equations of particle i and j in a one-dimensional collision system are as follows: mi dni(t) f ij (t) ∑ f i (t) dt dnj (t) dt f ij (t) ∑ f j (t). Adscript in the second-order Adams-Bashforth method shows the calculation time increases when appropriate processing is performed at discontinuous points. Contact force prediction method 3. ˆ ˆ ˆ f i (t ∆ t) ∑ f cij (t ∆ t) f oi (t ∆ t). the following formulae for the beginning and end points are used to guarantee second-order accuracy: nk 1/2 nk h/2 nk nk h/2 1/2 fk . we propose a method whereby a stable numerical solution can be found for a large time step in a system of multiple particle collisions. Fig. Adscript with the Leapfrog method shows the calculation time increases for the reasons given above. a onedimensional collision of one particle with a wall is assumed. because the frequency of the contact dynamic model becomes larger when the number of contact points increases. 6 shows the calculation procedure for this method. “Good” shows the case where the energy is not conserved but the eigenvalue is 1. For example. which is defined as follows: ∆ e (mn 2 k nk·k 1 1 Although the Leapfrog method has the above-mentioned problem. In the following section. 6. More concretely. 3. 1 1 (20) where “ ˆ ” denotes a predicted value. so more appropriate processing such as the multi-step method is needed to guarantee second-order accuracy. mj (25) where the second term of the right-hand side in each equation is a summation of forces acting on particles i and j except for the contact force between particles i and j. j kd2 1) (mn 2 kd2 ). However. but there is a conservative quality for a conversion when the eigenvalue of matrix A is 1. this means that the calculated energy value shows the minute fluctuations up and down around the true energy value and does not greatly decrease or increase. Table 2 shows that it is the superior of the three methods with second-order accuracy in calculation stability and energy conservation. With the above-mentioned consideration. As compared to the conventional calculation procedure.mn 2 1 k kd2 1 k k 1 h (mn 2 kd2 )Œmn 2 kd2 . is added. As an example of a multi-particle collision. m fk . The meaning of using a predicted value of a twoparticle collision for a multi-particle collision is examined here from the viewpoint of numerical analysis. Thus. k k k k m (18) 2 Equation (18) shows that energy is not strictly conserved. m (21) (22) However. it is necessary to note that the contact detection holds most of the calculation time rather than the calculation of contact force in a state of crowded particles. such as the beginning and end points of the contact to guarantee second-order accuracy. With the Leapfrog method. k k k 1 (19) and dk·k 1 are defined as follows: 1 nk 2 1 dk 2 1 nk·k dk·k 3 nk 2 3 dk 2 (23) . We refer to this method in which a predicted value of the contact force is used in this way as the “contact force prediction method” in this study. the shadowed area in Fig. (24) The calculation time in Table 2 shows the time needed to calculate a one-dimensional collision of a particle when the calculation time in the explicit Euler method is 1.23 (2005) 187 . Numerical equations in terms of the energy conservation in Table 2 show the error of energy conservation.1 Outline of contact force prediction method A predicted value of the contact force at the stage where two particles collide is solved by this method. The energy conservation in Table 2 is examined in the same way as when h 0. we assumed a one-dimensional collision of many particles without rotation. it is necessary for the calculation time needed to obtain the predicted value to be small. achieving second-order accuracy for the velocity of particle i depends on the following KONA No. Equation (22) cannot be used for the explicit method when h 0. the procedure for predicting the contact force at every contact point.

61 188 KONA No. In the following.22 2 p 4 m 2 nk h2 k 1 4 m 2 nk OK 1. ∑ f i (t) and ∑ f j(t). a method whereby the inf luence of other contact forces is ignored for the predicted value is considered. However. as mentioned in Section 2. “Exactly Good” in terms of energy conser vation means that the particle energy is conserved strictly according to the examination similar to Equation (18).22 m h2 2 h3 6 h4 24 2 2 2 (mnk kdk ) h OK 2.23 (2005) . Equation (24) is substituted for Equation (26) as follows: ni(t ∆t) ni(t) where df (t) fˆij (t ∆ t) f ij (t) ∆ t ij dt df (t) fˆi (t ∆ t) f i (t) ∆ t i . ∑ f i (t) and ∑ f j (t). four methods whereby stable solutions can be achieved in a large time step are selected and shown in Table 3.equation: ni(t ∆ t) ni(t) ∆ t dni(t) dt (∆ t)2 d 2ni(t) . 2 dt 2 (26) where the terms above third-order accuracy for ∆ t are omitted. Although various methods can be used to obtain the predicted value.(30) mj mi permits an error of a certain order for the calculation of the contact force. is ignored. (34) h/2 1 From Equation (28) and Equation (30). it is not necessary to achieve second-order accuracy with this method because DEM with a large time step h dk 1 m 1 (nk 2 1 h This differs from the conventional Runge-Kutta method with second-order accuracy in that Equation Table 3 Comparison of 4 methods of calculating the estimated value of force (k t/∆t. Equation (27) shows that first-order accuracy is achieved even if the inf luence of the other contact forces. h ∆t) Method Calculation Stability at h 0 Energy Conser vation at h 0 0 Implicit Trapezoidal Stable Exactly Good h2 k Modified 2-Order Runge-Kutta ∆t * Application to the Hertz and Mindlin Model Calculation Time NO 1. and the stability of calculation and energy conservation are examined when h 0. it is found that second-order accuracy is not achieved unless the inf luence of the other contact forces. dt (28) ∆t f ij (t) fˆij (t ∆t) ∑ f i (t) ∑ fˆi (t ∆t) . The manner of consideration is similar to that used in Table 2. The modified Runge-Kutta method with second-order accuracy is as follows: nk dk nk 1/2 nk dk nk dk h/2 1/2 fk m nk fk 1/2 (31) (32) (33) nk). is considered in the predicted value. dfij (t) dn (t) dni(t) kij nj (t) ni(t) hij j dt dt dt kij nj (t) ni(t) hij fij (t) ∑ f j (t) fij (t) ∑ f i (t) .77 Good k 4-Order Runge-Kutta ∆t * 2 M2 p No Good 0 Exact Solution Stable Exactly Good NO 1. 2mi (27) (29) On the other hand. the following equation relates to Equation (28).

(36) KONA No.. the predicted value of the contact force in the tangential direction is limited by the Coulomb frictional limit:  fˆsij (t ∆ t) mij  fˆnij (t ∆ t). as mentioned above. With the conventional DEM. is obtained as follows: * ˆ a1 ˆ a2 ˆ fˆnij knij dnij (t ∆ t) hnij dnij nnij (t ∆ t). (40) (35) where a1 and a2 are similar to the constants in Equation (2). The time step has to be set to less than the contact time tc. Particles fall freely from a height of 1 m. The linear spring model is used for the contact force model. 7 shows the time change of the energy summation of all particles in the case of the contact force fˆnij (t ∆ t) * fˆnij fˆnij (t ∆ t) 0.e. (38) where a3 and a4 are similar to the constants in ˆ Equation (6). ∆ tcij is the time when the two-particle * collision finishes. respectively. fˆnij (t ∆ t).23 (2005) 189 . nnij (t ∆ t) ˆ and nsij (t ∆ t) define the predicted particle velocity in the normal and tangential direction. for example: ˆ ∆ tcij fnij (t) ∆ t.2. In this study. These values can be obtained by either method. “Exact Solution” in Table 3 means a method whereby an exact solution to the differential equation of the contact dynamic model is used for the predicted value. respectively. In all calculations. The properties of the particles have already been shown in Table 1. as mentioned above.2.0 10 4 seconds. A predicted value is calculated by projecting on the normal direction and the tangential direction similar ˆ ˆ to the contact force. ˆ ˆ a3 fˆsij (t ∆ t) ksij dnij (t ∆ t)dsij (t ∆ t) ˆ 4 ˆ hsij dnij (t ∆ t)nsij (t ∆ t) a (37) ˆ when ∆ t Œ∆ tcij.(34) is the trapezoid method. fˆnij 0. so the predicted value of the contact force in the normal direction is also not permitted as follows: * when fˆnij Œ0. The periodic boundary condition is used in all side walls in order to moderate the wall effect. However. as mentioned in Section 2. it is interesting that this method is stable even with such little modification. * when fˆnij 0.3 m. fˆsij (t ∆ t). the number of particles is 200. 3. From Table 3. The method that can be applied to the non-linear spring model and permits a large time step is the fourth-order RungeKutta method. Fig. the effect of the time step that can be set larger is bigger for the total speedup of DEM calculation than the calculation time of the predicted value is small. The contact time between two particles in this calculation is 2. and the calculation space is a quadratic column whose bottom is 0. and is obtained by the following linear approximation equation. but these methods cannot be applied to the non-linear spring model. i. ˆ a3 ˆ ˆ ˆ fˆsij (t ∆ t) ksij dnij (t ∆ tcij)dsij (t ∆ tcij) ˆ a4 ˆ ˆ ˆ hsij dnij (t ∆ tcij)nsij (t ∆ tcij). and the spring constant is determined from the contact time by the Hertz spring model.3 m 0. This is because the calculation time of the contact detection is larger than that of the calculation of the contact force. the exact solution method is used to obtain the predicted value of the contact force with the contact force prediction method. Although the calculation when h 0 is unstable with the conventional RungeKutta method with second-order accuracy..1 Packing simulation The packing simulation results produced by the free fall of particles are shown here. “Stable” in terms of the stability of calculation in Table 3 means that a stable solution for the predicted value of the contact force can be achieved with a large time step. ∆t * 1. there is another limit for the time step in DEM calculation. is obtained as follows: when ∆ t ˆ ∆ tcij. The calculation time is smallest with the implicit trapezoid method and the exact solution method. 3. i. tension in the normal direction is not permitted by the no-tension joint. * fnij (t) f nij (39) Moreover.2 Application examples of the contact force prediction method Some examples of DEM calculation using the abovementioned original method are introduced here. The particle size is 50 mm. However. dnij (t ∆ t) and dsij (t ∆ t) define the predicted value of the particle location in the norˆ mal and tangential direction. the implicit trapezoid method and the exact solution method are superior to the other two methods in terms of the large time step and energy conservation.e. the packing and discharge system are used as an example of multi-particle collision. and the predicted value of the contact force in the normal direction. The predicted value in the tangential direction.

Therefore. In Fig. Fig. a big difference between the conventional method and the proposed method is not found. and the particles are allowed to fall freely from the outlet to the f loor. 9 shows the time change of the energy summation of all particles. Similar to Fig. it is not dispersed in nt 3 with the proposed method.63. 9. Moreover.6 0. 3. a distance of under 1 m. It is believed that this is because the calculation error is canceled out due to the movement of all particles. When the calculation time of 1 step is considered. New Fig. The calculated packing ratio becomes larger as the time step becomes 0. As for the discharge behavior. 7 Kinetic energy of total particles in packing simulation prediction method or the conventional method.553 and 0.62 0. so the proposed method can set the time step 4 times larger than that of the conventional method. In other words.) 15 nt 6 (Conv. In Fig. How fine the time step must be set with the proposed method depends on how much calculation error is permitted.2. Although the calculation is dispersed in nt3 with the conventional method. 0.5 (New) nt 3 (New) nt 3 (Conv. (s) Fig.25 nt 1.74. The random packing ratio should be between the closest packing ratio. 8 shows the comparison between the conventional method and the proposed method for the calculation results of the packing ratio.4 0.2 times larger than with the conventional method.59 0. 8 Result of packing simulation 190 KONA No.58 0. Although the calculation is dispersed in nt 3 with the conventional method. Incidentally.61 Packing Ratio. a 0.60 0.5 with the proposed method.) nt 300 (Conv. it is not dispersed in nt 3 with the proposed method. the proposed method accelerates DEM calculation 3 times.55 1 10 100 nt ( ) 1000 10000 Conv. ( ) 0.2-m diameter circular outlet is opened at the bottom of the packed particles bed.23 (2005) . the kinetic energy is not dispersed in nt 6 with the conventional method and nt 1. there have been various considerations regarding the random packing of mono-dispersed particles in the long term17).2 0. the proposed method accelerates DEM calculation 8 times when the time step is set 10 times larger than conventionally. experimental results of mono-dispersed spherical particles by Westman and White17) show that their packing ratio is between 0. 18).2 Discharge simulation In order to simulate a system in which particle movement is more dynamic than in the packing system. ntŒ30 with the proposed method. where tt is the contact time needed for one particle to contact a wall.57 obtained in this study is the appropriate value. nt is the number of partitions in the contact time as follows: nt tc . Fig. Fig. the proposed method can set the time step 10 times larger than that of the conventional method. The calculation time of 1 step with the proposed method is 1. Thus. theoretically 0.52. 7. and the packing ratio in a simple cube lattice. although the packing ratio does not depend on the time step in ntŒ300 with the conventional method. 8.57 0.8 1 Time. ∆t (41) larger. It is assumed that this is because the overlap at the beginning of contact between two particles is maintained.56 0. the packing ratio 0. respectively.) 10 20 Kinetic Energy. ( J ) 5 0 0 0. even with the large time step. 10 shows a comparison between the conventional method and the proposed method for the calculation results of the number of discharged particles from the outlet.

the packing and discharge system is chosen as an example of the multi-particle collision. four methods in which a large time step can be set are examined. KONA No. and their characteristics are shown. 10 Result of discharge simulation 4.) 3 (Conv. we propose an original method called the contact force prediction method. or by multi-step methods such as the predictor-corrector method.) 50 0 0 1 Time. However. In this paper.5 (New) nt 3 (New) nt 3 (Conv. As for the method of obtaining the predicted value of the contact force with the proposed method. Thus.) nt 6 (Conv. Under the present parameters. and the proposed method accelerates DEM calculation 3 8 times. (s) 2 3 Fig.6 5 Kinetic Energy. some calculation results of mono-dispersed particles are introduced as an imminent example. such as the method proposed here. Inspired by these two conventional methods.5) k : Time Index ( ) m : Mass of Particles (kg) n : Unit Vector from the Center of Particle to Contact Point ( ) nt : Number of Steps in Contact Time ( ) r : Radius of Particle (m) s : Tangential Unit Vector ( ) t : Time (s) tc : Contact Time (s) T : Torque Vector of Particle (N · m) v : Velocity Vector of Particle (m/s) n : Velocity of Particle (m/s) f ˆ f fc ˆ fc fo 100 nt 1. Conclusion DEM in which a large time step is possible is developed by improving the algorithm to solve the contact force between two particles. an additional acceleration. In general. the equations of motion of particles are solved by the second-order Adams-Bashforth method. even if several techniques to accelerate DEM calculation. which estimates the values of the contact force in the following calculation time by linear extrapolation. (s) 2 3 Fig. ( J ) 4 3 2 1 0 nt nt nt nt nt 1. such as parallel computation or combination with a continuous method. are applied. it is confirmed that the contact force prediction method proposed in this study is useful for DEM acceleration.5 (New) 3 (Conv. it is expected that the number of particles calculated in DEM will be limited to 100 million.) nt 300 (Conv.23 (2005) 191 . Nomenclature : Force Vector of Particle (N) : Estimated Value of f (N) : Contact Force Vector of Particle (N) : Estimated Value of fc (N) : Force Vector of Particle except for Contact Force (N) ˆ f o : Estimated Value of f o (N) g : Acceleration of Gravity Vector (m2/s) h : Time Step in Table 2 and 3 (s) i : Particle Index ( ) I : Moment of Inertia of Particle (kg · m2) j : Particle Index ( ) k : Spring Coefficient (N/m or N/m1. ( ) 150 which uses the predicted value of the contact force at each contact point of two particles. Moreover. 9 Kinetic energy of total particles in discharge simulation 200 Number of Discharged Particles.) 0 1 Time. but we have already confirmed that the contact force prediction method is useful for two sizes of particle at a number of approximately 10 million. is needed for application to the general scale system.) 6 (Conv.) 300 (Conv.

109. Petroleum Technol. 8) Mindlin. 259-268. 3) Asakura. T.” Trans. R. 7) Mindlin. 17) Cumberland. and Muzzio. Y. 14) Hitotsumatsu. (1992).: “Compliance of Elastic Bodies in Contact. T. 95-99.” J.. M.: “Numerical Simulation of Fluidized Bed Using the Distinct Element Method (the Case of Spouting Bed). 16) Moakher. and Goodier. 327-344. H.” J. P. (1996). (2000). Japanese Assoc. and Fukuda.: “Expectation of Gas Hydrate Transport Technology. Soc. J. N. US Army Corps of Engineers. A.23 (2005) .” J. Mechanics. Appl.” Chem..” Powder Technol.” 122. 9.” Kagaku Souchi. S. and Crawford.): “Doshitu Rikigaku. Sangyo Tosho. Mechanics. R.29. Mixing and Segregation of Non-Cohesive Grains in 3D Tumbling Benders. D.: “Suuchi Kaiseki. (1953). Eng. Kawasaki.” Technical Report MRD-2-74 Missouri River Division. (1995). Mech. 58-551. R. 11) Okui. 12) Nakajima. K. (1997)... T.” Sigen to Sozai. Gihodo Shuppan. and Nakajima. J.: “Seppyou no Kouzou to Bussei. D..: “ Transporting System of Natural Gas Hydrate in Pellet Form. Kokon Shoin. T. 2) Kawaguchi. 66.” J.” p. N. S. Japan: “Funtai Simulation Nyumon.: “Study of Dynamical Analysis for Scale Up of Fluidized-bed of Solid Dosage Forms by DEM Simulation. (2001). (1999). Kawagoe..: “Dynamics of Granular Matter from the Physical Point of View (I). I. (1998). Sci. 104. 15) The Society of Powder Technology.” Proceedings of JSME Transportation and Logistics 2001. 219-224. Japan Soc.. 9) Taguchi. H. 192 KONA No. Funayama. J. Powder Technol. Y. Harada. (1970). 168174. Y. (1986).” 380.893-933. Japan.: “Elastic Spheres in Contact Under Varying Oblique Forces. Elsevier. 315-316. 20. Shinbrot. 5) Natsuyama. A.: “Rational Design of Tunnel Supports: A Computer Model for Rock Mass Behavior Using Interactive Graphics for the Input and Output of Geometrical Data. S. S.” 103. B. (1987). and Takaoki. 10) Xu. 52-16. T.: “Critical Rotation Speed for Ball-Milling. (ed. 3rd Edition. Eng. McGraw-Hill Book Company.: “Numerical Simulation of the Gas-Solid Flow in a Fluidized Bed by Combining Discrete Particle Method with Computational Fluid Dynamics. 2119-2125. and Yu. 240-246.” p. (2001). 112. 4) Watanabe. B. (1982). 32.25) : Angular Position Vector of Particle (rad) m : Friction Coefficient ( ) : Angular Velocity Vector of Particle (rad/s) Subscripts i : Particle i ij : Between Particle i and j j : Particle j n : Normal Direction s : Tangential Direction References 1) Cundall. (1969).: “ Theory of Elasticity. T. Asakura Shoten. Appl.14-39. Y. and Tsuji. 16. 13) Maeno. (Series B).. T.: “Experimentally Validated Computations of Flow.: Estimated Value of v (m/s) : Position Vector of Particle (m) : Displacement (m) : Estimated Value of d (m) ∆ t : Time Step (s) ∆ t * : Normalized Time Step ( ) ˆ ∆ tc : Estimated Value of Contact Time (s) ∆t : Time Difference of Contact Start Points (s) ∆t* : Normalized Time Difference of Contact Start Points ( ) g : Constant Related to Restitution Coefficient ( ) h : Viscosity Coefficient (kg/s or kg/s/m0. 6) Timoshenko. M. P. D.” ˆ n x d ˆ d Powder Technol.” p. 2785-2809. (1949). J. Tanaka.. and Horio. M..: “ The Packing of Particles.: “Simulation of Descending Particles in Water by DEM. (2001). 58-71. 18) Mogami.

the University of Tokyo.E. control and optimization of processes. She received her Doctor of Engineering in 2003 from the University of Tokyo. from the Dept. simulation. His research subjects in NSC were modeling. KONA No. etc. He moved to the Department of Geosystem Engineering. segregation. and her M. of Science and Engineering in 1998 at Waseda University. as a lecturer in 2000. the University of Tokyo. classification. degrees from the University of Tokyo in 1974 and 1976. Her research interests are resources processing and environmental technology. Katsunori Okaya Katsunori Okaya is Research Associate at the Department of Geosystem Engineering. and now is a co-worker.23 (2005) 193 . Graduate School of Engineering. Chiharu Tokoro was his doctor course student from 2000 2003.’ He joined Nippon Steel Corporation (NSC) in 1990 and worked as a researcher and a senior researcher until 2000. The theme of his thesis was ‘A Study of Control in Mineral Processing Plants. He is a member of the Mining and Materials Processing Institute of Japan. Graduate School of Engineering. She received her B.Sc. of Geosystem Engineering in 2000 at the University of Tokyo. random packing. Ms. respectively. and M.E.Sc. Okaya’s current research interests are in the handling of particles and grains (grinding. a member of the board of Japan Federation of Ocean Engineering Societies. dispersion. He obtained his B. simulation and the control of industrial plants. transportation and/or storage of NGH. the University of Tokyo. Graduate School of Engineering. His research interests include natural gas hydrate (NGH). both from the Department of Geosystem Engineering.Author’s short biography Chiharu Tokoro Chiharu Tokoro is Research Associate at the Faculty of Science and Engineering of Waseda University. especially of steel making plants and energy plants.) and simulations using DEM (Discrete Element Method). Jun Sadaki Jun Sadaki received a master degree in geosystem engineering in 1987 and a Doctor of Engineering degree in geosystem engineering in 1990. application of NGH to industrial processes as well as the modeling. the Society of Instrument and Control Engineers. from the Dept.

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.