The Treatment of Rotational Motion in Molecular Dynamics

Roland Sonnenschein, Aatto Laaksonen* and Enrico Clementi

IBM Corporation, Data Systems Division, Dept. 48B,M S 428, Kingston, New York 12401 and National Foundation for Cancer Research, Kingston Laboratory, Kingston, New York 12401
Received 4 November 1985; accepted 18 March 1986.

I. INTRODUCTION

In the past, several attempts have been to obtain a well-behaved equation of motion to describe the reorientations of rigid molecules in the context of molecular dynamics simulation of molecular liquids. Quite recently7we found a second-order differential equation which allows a more accurate numerical solution then either the original proposal by Evans4 or the SHAKE algorithm.8 We report the results of a comparison with a method due to Fincham.' As a measure, the deviation of numerical trajectories starting from the same initial conditions was chosen

11. THE EQUATION OF MOTION FOR THE QUATERNION PARAMETERS

Following Evans4z5we define the quaternion parameters (QP) in terms of Goldstein's Euler angle^.^

5 = sin($/2) - sin(($ - (p)/2)

77
=

5=

40)/2) sin($/2) - sin(($ - p)/2)

sin(O/Z) cos(($
-

sin($/2) - cos(($ They obey the constraint

(p)/2)

(1)

G.a'=
where

(2)

instead of the conservation of energy which has been rather overstressed in the past. In the next section both methods are introduced briefly, while the following section contains the results of test calculations. These calculations revealed also the effect of interpolation error of the noncoulombic forces due to the increased accuracy. An alternate tabulation scheme is developed which reduces imprecision leaving the memory requirements unchanged. In our opinion, however, the use in MD simulations of tabulation as a shortcut for the force calculation should be avoided whenever possible.
*Permanent address: Department of Physical Chemistry, Arrhenius Laboratory, University of Stockholm, S-106 91 Stockholm, Sweden

rotates the laboratory into the molecular fixed system. The time derivative is related to the angular velocity & by

1-59

-x,

77,

5\

From the Eulerian equationsg the following differential equation was obtained:

-2ij

ij

Journal of Computational Chemistry, Vol. 7, No. 5, 645-647 (1986) 0 1986 by John Wiley & Sons, Inc.

CCC 0192-8651l86050645-03$04.00

646

Sonnenschein, Laaksonen, and Clementi

where the components of G can be expressed in terms of a' and G, so that ( 5 )does not depend explicitly on 5; any more. Fincham's algorithm is based on the same quaternion parameters. It is given in terms of a leap frog scheme relating the +angular momentum in th_elaboratory frgme S, the in principal frame J,",the torque T" and the an" to gular velocity ; the quaternion parameters:

C (GA
L=l

GL~I2

(9)

where A t is the step size. Note that the introduction of a half step n + ?42 should take care of the fact that $ does explicitly depend on [see (511.

111. DETAILS OF TEST CALCULATIONS

Starting from the same initial conditions for N = 256 water molecules represented by the MCY potential" trajectories of 0.5 ps length were generated with the step lengths 1.0 x sec and 2.0 x sec (referred to as A and B). For one pair (6) was used to calculate the orientational motion while employing the Stormer formula of lowest order for the center of mass motion." The other pair was obtained integrating ( 5 ) as well as Newton's e q u a t i o n s w i t h a predictorcorrector scheme developed by Gear.12 The averaged kinetic energy corresponded t o a temperature of 300 K. While Sonnenschein7 had restricted the comparison of different algorithms to energy conservation, the deviation of the trajectories were studied in more detail at this time. The following quantities were calculated from the simulations as a function of time:

Fa denotes a center of mass vector, Cm the corresponding velocity where the subscript i X refers to the ith molecule in trajectory X . While CA,(t) and QAB(t) describe the evolution of displacement between the trajectories A and B, VABmeasures the correlation between the velocities. The correlation being greater than one at times indicates that the magnitudes of both vectors have also become quite different. The resulting curves can be seen in Figure 1.The superiority of the second order scheme is evident. As it was integrated with Gear's predictor corrector method making one loop only, the deviation can be taken as an estimate for the error made applying the bigger time step. Note that the accuracy for the leap frog scheme is increased very little when the step length At is reduced further (see Fig. 2). This is partially due to the interpolation error made in calculating the non-coulombic c o n t r i b u t i o n s t o force and energy as can be seen from the following example: Let the box side have a length of 20 A. Supposing kT/2 at 300 K as the kinetic energy for collision and the 0-0 contribution to the MCY potential energy one obtains the following values from the table:
~

(Roo1

Exact Energy
0.28492729 0.27986378 0.27244760

Interpolated
0.27993541

0.084 0.0842 0.0845

The exact value and the interpolated value differ considerably. If one tabulated the energy as a function of l/(Roo) instead using a table length of 2000 once more the accuracy of the linear interpolation is increased considerably:
(Roo)

Exact Energy
0.28492729 0.28238343 0.27986378

Interpolated
0.28239250

0.0840 0.0841 0.0842

Whether this improvement is sufficient is still questionable because the error propagation with the equations of motion encountered in molecular dynamics simulation may

Rotational Motion in Molecular Dynamics

COM Deviation 1./.5

647

0.6

0.4

0.3

40,01.
i

Quaternion Deviation 1./2.

I
ol

40,0135.0-

Quaternion Deviation 1J.5

35.030.0
-

30.0

, 25.0 E

25.0-

i 2c.o-

2 20.0-

5.3
1.2,-

l+ : .

,
0.100E-12

*/f
0.300E-12

0.500E-12

Figure 2. Top: center of mass deviation between trajectory A and a third one obtained with half the step size 0.5E-15 seconds. Bottom: quaternion deviation, as in top.
0.3 0.2
0,,

_COM Velocity Deviation 1.0/2.0fs

References
1. J. Barojas, D. Levesque, and B. Quentrec, Phys. Rev. A 7, 1092, (1973). 2. P.S. Y.Cheung, Chem. Phys. Lett., 13, 40 (1976). 3. F. H. Stillinger and A. Rahman, J . Chem. Phys., 60, 1545 (1974). 4. D.J. Evans, Mol. Phys., 34, 317 (1977). 5. D .J . Evans and S. Murad, Mol. Phys., 34, 327 (1977). 6. D Fincham, CCP5 Znformation Quarterly, No. 2, : 1981. 7. R. Sonnenschein, J. Comput. Phys., 59,347 (1985). 8. J.P. Ryckaert, G. Ciccotti, and H.J.C. Berendsen, J . Comput. Phys., 23, 327, (1977). 9. H. Goldstein, Classical Mechanics, Addison-Wesley, Reading, MA. 10. 0. Matsuoka, E. Clementi, and M. Yoshimine, J . Chem. Phys., 64, 1351, (1976). 11. L. Verlet, Phys. Rev. 98, 159, (1967). 12. C. W. Gear, AEC Research and Development Report NO. ANL-7126 (1966).

be serious. The investigation of this problem has to be postponed until a more accurate MD program becomes available.
IV. CONCLUSIONS

The results reported here strongly suggest that the simple leap-frog schemes used to integrate equations of motion for MD simulations although very popular should be abandoned in favor of more accurate methods.