MODELING AND SIMULATING REACTIVE MOLECULAR DYNAMICS USING REAXFF

R EAX FF REACTIVE MD SIMULATION OF A COMB USTING CHAR STRUCTUR E

MARKUS OHLENFORST

C ONTENTS
0. Introduction ............................................................................................................................ 3 1. ReaxFF Modeling and Implementation ............................................................................... 4 1.1. Reactive interactions ....................................................................................................... 4 1.1.1. Modeling bond orders .............................................................................................. 4 1.1.2. Equilibrating charges ................................................................................................ 5 1.2. Numerical Aspects ........................................................................................................... 6 1.2.1. The QEq system ........................................................................................................ 6 1.2.2. Solving QEq ............................................................................................................... 7 1.2.3. ILUT-based preconditioning ..................................................................................... 7 1.3. Algorithmic Techniques ................................................................................................... 8 1.3.1 Generating neighbors ................................................................................................ 8 1.3.2. Bonded and non-bonded force computations ......................................................... 9 2. ReaxFF - Simulation based analysis of a new fuel ................................................................ 10 2.1 Simulation details and results ........................................................................................ 10 2.1.1 Mechanisms of initial pyrolysis ............................................................................... 10 2.1.2 Mechanisms of initial combustion .......................................................................... 13 2.2 Output analysis extracting kinetic properties ............................................................. 16 3. Closing remarks .................................................................................................................... 17

Images on cover sheet from: F. Castro-Marcano, A. M. Kamat, M. F. Russo Jr., A. C. T. van Duin, J. P. Matthews, Combustion of an Illinois No. 6 coal char simulated using an atomistic char representation and the ReaxFF reactive force field Combustion and Flame, 159 (3), 1272-1285 (2012).

0. I NTRODUCTION

his work was written in the context of the CES-Seminar-class for master students in Computational Engineering Science at the RWTH Aachen University (WS 12/13). Task was to write a paper about a current topic of research represented by one or more recent publications. Prof. Dr. Ismail from the chair for Molecular Simulations and Transformations gave Reactive Molecular Dynamics as the topic for this work and supervised it as well. Aforementioned publications for this work were: [1]: A.C.T. van Duin, S. Dasgupta, F. Lorant, and W.A. Goddard III, ReaxFF: A reactive force field for hydrocarbons, J. Phys. Chem. A, 105 (2001), pp. 93969409. [2]: H. M. Aktulga, S. A. Pandit, A. C. T. van Duin, A. Y. Grama, Reactive Molecular Dynamics: Numerical Methods and Algorithmic Techniques, SIAM J. Sci. Comput., 34(1), C1C23. [3]: Liu, Lianchi and Bai, Chen and Sun, Huai and W.A. Goddard III, Mechanism and Kinetics for the Initial Steps of Pyrolysis and Combustion of 1,6-Dicyclopropane-2,4hexyne from ReaxFF Reactive Dynamics, J. Phys. Chem. A, 115 (19). pp. 4941-4950.

In reactive molecular dynamics, ReaxFF (reactive force field) is a force field that can be used for simulations of the (reactive) behavior of various chemical systems. It was developed by Adri van Duin, William A. Goddard, III and co-workers at the California Institute of Technology and published in 2001. Its advantages in comparison with common techniques in the field of molecular dynamics are presented and the underlying model, as well as a way to implement it is explained in the first chapter ReaxFF - Modeling and Implementation. The second chapter, Simulation based analysis of a new fuel, is about the results of a recent study, in which researchers applied ReaxFF for the analysis of a new fuel (additive), 1,6-Dicyclopropane-2,4-hexyne. Aspiration is to help the reader to gain access to the understanding of how to distinguish ReaxFF from other techniques, what model it relies on, how it can be implemented and finally, how it is used in practice.

1. R EAX FF M ODELING AND I MPLEMENTATION

eaxFF fills the gap between quantum mechanical (QM) and empirical force field (EFF) based computational chemical methods. Although QM methods can generally be used for all chemical systems, regardless of their connectivity, they are not practicable for large systems with sizes in the order of thousands of atoms. Because of the computational expense QM methods are primarily in use for single point or local energy minimization. Hightemperature molecular dynamics simulations, like the ones presented in this work, are in general too time-consuming. ReaxFF was developed to simulate the molecular dynamics of large scale chemical systems (sizes in the order of thousands of atoms) efficiently. Contrary to traditional force fields and like QM methods, it is able to model chemical reactions. Changing bonds are not a problem, whereas for traditional force fields, the functional form depends on having all bonds defined explicitly. ReaxFF replaces explicit bonds by bond orders, which allows continuous bond forming/breaking. It is developed to be as general as possible and has been parameterized and tested for hydrocarbon reactions, transition-metal-catalyzed nanotube formation, and high- energy materials. How the ReaxFF model considers reactive interactions exactly and how it can be implemented is shown in the following.

1.1. R EACTIVE

INTERACTIONS

y the reactive force field (ReaxFF) atoms are modeled as separate entities with different bond structures. These have to be updated at every time-step. In combination with the therefore necessary charge redistribution (charges on atoms change due to bonding changes), the dynamic bonding scheme represents the essential part of ReaxFF that distinguishes the model from classical MD or ab-initio methods. Before going into detail regarding the numerical and algorithmic aspects when implementing the force field, the incorporated reactive potentials in ReaxFF are briefly described in this section.

1.1.1. M ODELIN G

B OND ORDERS

The strength of the bond between a pair of atoms i and j is described by the bond order through the number of chemical bonds. ReaxFF models this quantity by a closed form [1]:

(1.1)

Here, the bond order depends on the types of atoms i and j and the distance in between. denotes a - , - or - bond. a and b are parameters belonging to the bond type. is the length which is optimal for this bond type. Since the actual bond type is unknown during simulation, the total bond order equals the following sum [1]: (1.2) Additionally, the total coordination number of each atom and 1 3 bond corrections in valence angles must be considered. This is suggested by the exemplary fact that the bond length and strength between O and H atoms in OH are different than those in H 2O. Concurrent corrections are employed in ReaxFF by [1]:

(1.3) stands for the deviation of atom i from its optimal coordination number, the overcoordination correction, and for

for 1 3 bond order

corrections. After bond orders are computed, according charge changes have to be incorporated. Then, a ReaxFF simulation continues much like a classical MD simulation.

1.1.2. E QUILIBRATI NG

CHARGES

Due to dynamic bonding in ReaxFF atoms underlie different charges for the duration of simulations. To redistribute charges periodically it would be most accurate to employ ab-initio methods. Unfavorable is that this would make the ReaxFF method unscalable. Therefore developers stuck to the QEq method [2]. According to that, the actual problem is approximated by looking for a set of charges constituting a minimal electrostatic energy of the system with the same net charge: (1.4)

i and j represent atom indices and is referred to as the partial charge on atom i. Specifying the addends in equation 1.4 through physical quantities gives: (1.5)

denotes the energy of an isolated neutral atom. is the electronegativity. stands for the idem potential or self-Coulomb of i, while corresponds to the Coulomb interaction between atoms i and j.

1.2. N UMERICAL A SPECTS

hile typical classical MD formulations are based on static charges on atoms, (re)assigning partial charges to atoms at each time-step constitutes one of the most computation-intensive parts of a ReaxFF simulation. As seen above, the charge reassignment problem can be approximated as charge equilibration with the objective of minimizing the electrostatic energy. One efficient way to solve problem (1.5) of the ReaxFF model numerically was presented by Aktulga et al. in 2012 [2] and is outlined in the following.

1.2.1. T HE QE Q

S YSTEM

To get linear systems out of equation 1.5, the method of Lagrange multipliers can be used. After some computation you acquire: (1.6) (1.7) Here, represents a vector of size N, containing parameters determined based on the types of atoms in the system. N is the number of incorporated atoms. H represents the QEq N x N sparse coefficient matrix where the diagonal of H consists of the polarization energies of atoms, and the off-diagonal elements hold the electrostatic interaction coefficients between atom pairs. s and t denote fictitious charge vectors of size N. Finally the partial charges are computed with the help of the fictitious charges: (1.8) In fact, a direct solver could give the solution to the linear systems in (1.6) and (1.7). Nevertheless Krylov subspace methods are much cheaper for moderate to large sized systems. A closer look at how the system above can be solved in a highly efficient way is given in the next section.

1.2.2. S OLVING QE Q
H is a sparse matrix because the number of non-zeros in H is only on the order of a few hundred entries per row, and thats independent of the system size. Reason is, non-bonded interactions in ReaxFF are computed within the cut-off radius only. However, Krylov subspace solvers are iterative methods depending upon a costly sparse matrix-vector multiplication at each iteration. Therefore, a preconditioning technique can be employed to convert the linear system into an equivalent one that has improved spectral properties. Furthermore, computing linear, quadratic, or higher order extrapolations on the solutions from previous steps makes sense since individual time-steps in reactive MD simulations have to be much shorter than those in typical classical MD methods. Consequently, good initial guesses can be made and are even more favorable.

1.2.3. ILUT- BASED

PRECONDITIONIN G

Incomplete LU factorization (ILU) proved to be a useful basis for preconditioning. This technique is effective and broadly used in the context of solving sparse linear systems since it helps in reducing the iteration counts. Even though, calculating the ILU factors and putting them in use as preconditioners frequently is computationally expensive, especially when solving the QEq problem at each step of a ReaxFF simulation. Indeed, the simulation environment evolves slowly in a ReaxFF simulation. That indicates that the QEq coefficient matrix H, as well as the ILU factors L and U evolves slowly, too. For that reason, analogue to the reuse of the solution from the previous step as initial guess, the same factors L and U can be assumed effectively as preconditioners over several steps. It is possible to use the same preconditioner over tens to thousands of time-steps with only a slight increase in the iteration count. Of course, that depends on the displacement rate of atoms in the system and the specified accuracy of the solution. In the sPuReMD-implementation of ReaxFF by Aktulga et al. [2], factors from such an ILU factorization of the H matrix are employed with a threshold (ILUT). That means, that all entries in the L and U factors with values less than the specified threshold are set to zero for optimization reasons. By reducing the threshold, the factors L and U can be turned into higher quality preconditioners. Disadvantageous is that factorization and application of the preconditioner takes considerably longer. An optimal threshold value has to be found empirically depending on the type of the experiment to be simulated. Concerning the PGMRES (Generalized Minimal Residual Method with Preconditioning) and PCG (Conjugate Gradients with Preconditioning) solvers (both Krylov subspace methods) for solving the system (1.6)-(1.8), the following can be said: PGMRES doesnt only perform better, but shows the higher longevity of preconditioners as well [2]. sPuReMD employs PGMRES as the default solver.

1.3. A LGORITHMIC T ECHNIQUES

s shown in Algorithm 1, the high-level structure of the sPuReMD-implementation of ReaxFF looks like one of a classical MD code. Still, due to dynamic bonding and charge equilibration requirements, each of the listed components entails a significantly higher complexity compared to a nonreactive classical MD implementation. ALGORITHM 1 GENERAL STRUCTURE OF AN ATOMISTIC MODELING CODE. Read geometry, force field parameters, user control file Initialize data structure for t=0 to nsteps do Generate neighbors Compute energy and forces Evolve the system Output system info end for In the following, the algorithmic techniques and optimizations applied in sPuReMD are dealt with. The developers objective was to deliver excellent per time-step performance and linear time scaling in system size.

1.3.1 G ENERATING

NEIGHBORS

Both bonded and non-bonded interactions between atoms are truncated after a certain cut-off distance in ReaxFF. They account for 45 for bonded interactions and 1012 for non-bonded ones. To gain the relevant neighbors of each atom sPuReMD applies a procedure called binning or link-cell method [2]. First, a three-dimensional grid structure is created by partitioning the specified geometry of the simulation into small cubic cells. Hereafter, atoms are binned into these cells depending on their spatial coordinates. Obviously, potential neighbors of an atom are either in the same cell or in neighboring cells within the cut-off distance measured from its own cell. One can achieve O(k) neighbor generation complexity for each atom with k being the number of neighbors, averaged over all atoms. However, depending on the applied method and the specified cell size, k can be considerably high. Its possible to generate neighbor lists in linear time. Though, this part belongs to the most computationally expensive components of an MD simulation, which suggests lowering the large constant associated. Therefore, different optimizations are implemented in sPuReMD. Two important ones deal with the cell dimensions and the atom list respectively. On the one hand reducing cell dimensions comes along with a reduced search space per atom, which helps to lower the neighbor search time to a certain extent. On the other hand ongoing reduction (shape of the search space approximating a sphere) involves an overhead 8

associated with managing the increased number of cells. For normally sparse and crowded cells 1/2 should work as a near-optimal value. Atoms that fall into the same cell are regrouped in the atom list. This allows a better neighbor search performance because of the better use of the cache.

1.3.2. B ONDED

AND NON - BONDED FORCE COMPUTA TIONS

There are four major steps involved in the computation of forces in ReaxFF: system wide calculation of the bond orders between atom pairs, determination of the forces due to the given bonds, assignment of the partial charges on each atom and computation of the nonbonded forces in the system. ELIMINATING BOND ORDER DERIVATIVES LIST For the most part, bonded potentials are determined by the strength of the bonds between the atoms involved. A closer look at (1.3) reveals that relies on all the uncorrected bond orders of both atoms i and j. This indicates that when regarding the force due to the i-j bond, the term is non-zero for all atoms k that share a bond with atom i or j. Since a single bond might contribute by this means to various bonded interactions, the evaluation of could be necessary multiple times over a single time-step. Instead of storing the bond order derivatives for frequent lookups to the physical memory, sPuReMD postpones their computation until the end of a time-step. Then, when the bonded interactions are calculated, the needed s are computed and added with the preevaluated according coefficients to the net force on atom k. This way, simulations of much larger systems are affordable on a single processor due to the saving of abundant memory. Moreover, it reduces the time for computing forces.

LOOKUP-TABLES FOR NON-BONDED INTERACTIONS Also the modeling of non-bonded interactions is more complex in ReaxFF than in the classical counterparts. The application of a lookup table and interpolation to approximate complex expressions is common for optimizing MD simulations. sPuReMD employs cubic spline interpolation in combination with a compact lookup table for accurate approximations of non-bonded energies and forces [2]. Considerable performance improvements justify an acceptable loss in terms of accuracy.

2. R EAX FF - S IMULATION BASED ANALYSIS OF A NEW FUEL

o show exemplary, how useful results are generated and extracted from simulation outputs in the field of modern reactive molecular dynamics, the essence of an early research work is outlined in this section. In 2010, Liu et al. used the ReaxFF force field successfully for the analysis of a new fuel material (1,6-dicyclopropane-2,4-hexyne) [3]. Since it has immense heats of pyrolysis and combustion, this fuel material promised to be a potential high-energy fuel or fuel additive suitable for rockets and marine vessels. It was assumed that by igniting the main fuel more easily, 1,6-dicyclopropane-2,4-hexyne would provide more energy in the ignition process. For the time being, the researchers directed their attention to the initial mechanism and kinetic analysis for both pyrolysis and combustion. That should help to understand whether the new material was stable enough in the preheating/injection process and how it would help the ignition. Since the objective was to follow the important initial chemical processes in a relatively large reactive system over a range of temperatures for periods of time of at least 10 ns, the choice fell onto ReaxFF instead of a quantum mechanics calculation program. Initial pyrolysis and combustion were analyzed by simulating unimolecular and multimolecular systems for both processes. The resulting important reaction steps of all unimolecular simulations were compared with results from quantum mechanics (QM) for validation purposes.

2.1 S IMULATION

DETAILS AN D RESULTS

otential functions and parameters from a prior ReaxFF analysis of the initiation mechanisms and kinetics of pyrolysis and combustion of the JP-10 hydrocarbon Jet Fuel were adopted for the analysis of 1,6-dicyclopropane-2,4-hexyne. These parameters were trained and validated against results from quantum mechanics. The RD simulations ran subject to the conditions of a constant number of atoms, constant volume and constant temperature (NVT ensemble). Additionally, periodic boundary conditions (PBC) were applied. The temperature was controlled with the help of a Berendsen thermostat. QM calculations were performed by the TURBOMOLE 5.8 package.

2.1.1 M ECHANISMS

OF INITIAL PYROLYSIS

For the unimolecular pyrolysis simulations, a single fuel molecule was examined in a cubic cell with 16.29 sides. In the first place, the system was equilibrated at 300 K for 10 ps with 0.1 fs time steps. After that it was heated up to 2500 K steadily over a time of 10 ps. Then simulation was accomplished for another 1 ns at 2500 K. The resulting pressure accounted for 84 MPa which is realistic for the preheat/injection phase in engines. For the sake of 10

proper statistical sampling and to gain a variety of likely mechanisms ten independent simulations of the unimolecular pyrolysis were performed. The researchers have noticed two different initial decomposition reactions when examining the results, see table 1.

T ABLE 1: I NITIAL REACTIONS FROM R EAX FF UNIMOLECULAR PYROLYSIS SIMULATION [3]

INITA: The ring of the cyclopropyl structures breaks to create the two stable molecules ethylene and 1,3,5-hexatriyne. INITB: The fuel molecule is isomerized, which causes a biradical intermediate to form, that then transfers a hydrogen atom to form more stable intermediates.

T ABLE 2: S PECIFIC REACTIONS FR OM UNIMOLECULAR PYRO LYSIS SIMULATIONS [3]

11

Following these two initial reactions, there are many other reactions in the unimolecular pyrolysis, see table 2. You can derive the lead initial reaction from the name of each elementary reaction. All in all, it can be seen that most ReaxFF energies are close to the QM calculations. Only the energies of some radical decomposition reactions and radical-radical reactions are undervalued. It is assumed that the reason for this was ReaxFF underestimating the energy of radicals [3]. Otherwise, the trends in the total energy are in good agreement with the QM calculations which confirms the accuracy of the ReaxFF force calculation. The researchers categorized the observed reactions and put them into two groups including different types, see markings in table 2 and corresponding legend underneath. NORAD: Reactions with no radicals involved in reactants. YESRAD: Reactions that involve radicals in the reactants. While dissociation reaction energies for NoRad are computed positive (endothermic), what induces less active species, but with entropy release getting crucial in pyrolysis, reaction energies for the radical-radical and molecule-radical reactions of the YesRad group are negative (exothermic), what induces reactive radicals leading to a variety of pyrolysis. Figure 1 visualizes the energetics of the unimolecular pyrolysis for three of the ten pyrolysis simulations. It can be seen that the unimolecular pyrolysis reaction behavior is endothermic. Clearly entropy has a certain weight in this process.

F IGURE 1: R ELATIVE ENERGIES ( KCAL / MOL ) OF THE THREE PATHWAYS

OBSERVED IN

R EAX FF UNIMOLECULAR PYROLYSIS [3]

The thermal conditions for the multimolecular pyrolysis simulations were the same as for the unimolecular ones. To get the same density as in the unimolecular structure, a periodic cubic box with 60 sides was taken and filled with 50 fuel molecules. Again, the system was equilibrated in the first place (300 K, 50 ps, 0.1 fs time step) and then heated to 2500 K steadily (10 ps time step). Finally RD simulation was performed for another 100 ps (2500 K, 0.25 fs time step). Figure 2 shows the fragment distributions and potential energy profiles with time for pyrolysis of the multimolecular system at 2500 K. 12

F IGURE 2: F RAGMENT DISTRIBUTIONS AND POTENTIAL

ENERGY PROFILE OF MULTIMOLECULAR PYROLYSIS

[3]

A large number of different radicals can be observed in the middle. Here, C4H, C2H, C2H3, C2 and C4H2 represent the major species. On the one hand the initial steps of multimolecular are mainly the same as the ones of the unimolecular pyrolysis. The fuel molecules mostly decompose to ethylene and 1,3,5-hexatriyne because of the initial reaction InitA. On the other hand the following reactions after this initial decomposition were considerably different for the multimolecular system. Much more complex species and the existence of many radicals lead to the conclusion that intermolecular radical reactions are crucial in the pyrolysis. Furthermore, the multimolecular pyrolysis proceeds faster than the unimolecular one. This is a result of the additional radical reactions in the multimolecular pyrolysis. Analysis of major product distributions at the end of 100 ps RD of pyrolysis as a function of temperature (not presented here) demonstrate that the H2 concentration apparently grows with temperature, pointing at the increased importance of H abstractions and activity of H radicals in the high temperature pyrolysis. Rsum: The pyrolysis of the fuel seems to begin with unimolecular pyrolysis involving radicals from secondary decompositions that accelerate the process. Additionally it was found that the endothermic, entropy-driven abstraction of ethylene from the fuel molecule is the most important initial step. In parallel, isomerization of the fuel molecule occurs as an occasional initial reaction (20%). Despite not being the main reaction, the isomerization creates radicals that influence the multimolecular pyrolysis significantly.

2.1.2 M ECHANISMS

OF INITIAL COMBUSTION

For the analysis of the initial mechanism of combustion, first, a single fuel molecule (C10H10) was situated in a cubic periodic box with 20 sides. 13 additional oxygen molecules caused an equivalent ratio of ca. 1.0. After equilibrating the system at 300 K (10 ps, 0.1 fs time step) and heating it to 1500 K steadily (10 ps, 0.25 fs time step) with bonding being disabled to prevent combustion reactions from occurring beforehand, again, 10 parallel independent

13

simulations were performed. Three different initial combustion reactions were observed for the unimolecular model: INITR1: O2 attack on the cyclopropyl structure, which causes the formation of a fivemembered peroxide ring. INITR2: O2 attack on the middle C-C bond of the diyne. That cracks to form two C5H5O radicals. INITR3: O2 attack on the cyclopropyl structure causes ring-opening and formation of a 7membered peroxide ring. Following these two initial reactions, there are many other reactions in the unimolecular combustion, see table 3.

T ABLE 3: S PECIFIC REACTIONS FR OM UNIMOLECULAR COMBUSTION SIMULATIONS [3]

14

Obviously, oxygen attack reactions and radical reactions are the most important subsequent reactions with the oxygen molecule as a radical acceptor and generator, thus inducing other radical reactions to accelerate the oxidation. In Figure 3, relative energies for the three different pathways of unimolecular combustion are shown.

F IGURE 3: R ELATIVE ENERGIES ( KCAL / MOL ) OF THE THREE PATHWAYS

OBSERVED IN

R EAX FF UNIMOLECULAR COMBUSTION [3]

The combustion processes are all exothermic, which signifies that combustion proceeds more easily than pyrolysis. Again, the comparison between ReaxFF and QM reaction energy values verifies the accuracy of the ReaxFF force field. Though, once more the ReaxFF simulations seem to have underestimated the energies of radicals. To examine the initial mechanisms of multimolecular combustion, 30 fuel molecules and 390 oxygen molecules were placed in a cubic box with 62.0 sides. Result was a multimolecular system with an equivalent ratio of ca. 1.0 and the same density as the unimolecular one. After equilibrating the box at 300 K (10 ps, 0.1 fs time step) and heating it to 1500 K (10 ps, 0.25 fs time step) with bond interactions being turned off, RD simulation was accomplished at 1500 K for 1 ns using 0.25 fs time steps.

F IGURE 4: F RAGMENT DISTRIBUTIONS AND POTENTIAL ENERGY

OF MULTIMOLECULAR COMBUSTION

[3]

15

During 1 ns of multimolecular combustion simulations, the initial intermediates were C10H10O2 and C5H5O and the major products were CO2, CO, CH2O and C2H4. Their distribution, as well as the potential energy profile is shown in figure 4. Also multimolecular combustion seems to be exothermic. Furthermore, the resulting initial reactions with most initial intermediates being C10H10O2 and C5H5O are similar to the unimolecular system. Still, multimolecular combustion generates many more radical intermediates. Rsum: Combustion of 1,6-dicyclopropane-2,4-hexyne is initiated by the unimolecular oxidation. In this context, oxygen attacks any one of three different positions on the fuel molecule. Numerous radicals are created, causing the combustion to be unstable towards explosion. Resulting distributions at the end of the simulation at different temperatures (not outlines here) show that H2O, H2, and CO concentrations grow with increasing temperature. Obviously, more radical reactions appear at high temperature. The rate of the exothermic combustion process is much higher and its temperature is much lower than the according quantity in pyrolysis.

2.2 O UTPUT

ANALYSIS

EXTRACTING KINETIC P ROPERTIES

D simulations on the multimolecular models of pyrolysis and combustion were performed for several temperatures (1600-2500 K range, 100 K steps) to determine the kinetic properties for pyrolysis and combustion. The according simulations were accomplished for 100 ps at each temperature with a 0.25 fs time step for the pyrolysis model and for 1 ns with a 0.25 fs time step for the combustion model.

F IGURE 5: K INETIC ANALYSIS OF PYROLYSIS AND COMBUSTION [3]

In Figure 5, the log of the initial rate of the loss of fuel molecules is plotted versus (1/T) for pyrolysis and combustion. First-order kinetics exhibit the behavior of a single Arrhenius function (see upper right corner in figure 5). That suggests the extraction of an effective 16

activation energy and a pre-exponential factor A. An activation energy of and a pre-exponential factor of can be gathered from the pyrolysis results. On the assumption of unimolecular decomposition, transition state theory gives with , a negative activation of entropy which is consistent with the TST for multimolecular reactions. That indicates the involvement of a multimolecular transition state. Regarding combustion of 1.6-dicyclopropane-2.4-hexyne, figure 5 gives in combination with a pre-exponential factor of . More likely is a lower for combustion because O2 can stabilize the initial steps of bond breaking and the initial reaction steps tend to be more exothermic [3]. Comparing to the initial reaction step at 0 K of InitR1, InitR2 and InitR3 for the three different oxygen-attack positions leads to 2.19 kcal/mol at 1500 K after correction of temperature. On the assumption of TST a of -10.29 eu can be calculated. Oxygen attack causes the decrease of entropy at the transition state.

3. C LOSING REMARKS

oday, molecular modeling and simulation techniques are routinely used to investigate the structure, dynamics, reactivity, electronic charge distributions, dipoles and higher multipole moments, surface properties and thermodynamics of inorganic, biological and polymeric systems. Only the simplest calculations can be done by hand. Inevitably, computers are required to carry out molecular modeling of any reasonably sized system. The computer time and other quantities (e.g. memory or disk space) grow rapidly with the size of the system being studied. As already stated, highly accurate methods, such as abinitio methods that are based entirely on the theory from the first principles of the Schrdinger equation, are typically practical only for very small systems. Whereas other, faster methods, mostly based on empirical and semi-empirical force fields, are less accurate since they employ experimental results and relatively simple potential functions to approximate some elements of the underlying theory. In addition, because these force fields describe the system empirically rather than in a fundamental fashion, they are only applicable to systems similar to the ones from the training set. ReaxFF was developed to resolve this contradiction to a certain extent and seems to satisfy that aspiration in the fields of hydrocarbon reactions, transition-metal-catalyzed nanotube formation and high- energy materials. The chosen example pointed at its good accuracy in comparison with QM methods. But ReaxFF provides a much faster method, especially when implemented in an efficient way as by sPuReMD, see chapter 1. Thus, it opened up new possibilities for computational chemistry and is worth it to be dealt with.

17