Coal Pvrolvsis Simulation By GPU-based Reactive Force Field Molecular Dynamics (ReaxFF MD)

Coal pyrolysis refers to the thermal decomposition of coal in a vacuum at a specific temperature, which plays an important role in the efficient and clean coal conversion and utilization. Coal undergoes a rapid loss of moisture and volatiles, followed by a myriad of coupled complex reaction pathways when pyrolysis occurs. The heterogeneous nature of coal and the complexity of the pyrolysis process have made it very difficult to access the comprehensive mechanisms of coal pyrolysis, even with the state-of-the-art experimental approaches.As a method for investigating chemical reaction with high accuracy, quantum mechanics (including the widely used DFT method) is computationally intensive, resulting in the fact that it has little applicability for complex coal pyrolysis. Reactive force field (ReaxFF), a recent and novel bond order potential, is able to describe the evolving of formation, transition and dissociation of chemical bonds with accuracy close to DFT but with very much reduced computational costs when combined with molecular dynamics (ReaxFF MD). With the capability to simulate larger and more complex molecular systems involving chemical reactions without pre-defining reaction pathways, ReaxFF MD provides a new and promising approach for molecular simulation of complex coal pyrolysis system and the underlying with chemical reactions.Combining high performance computing and cheminformatics analysis, the first GPU based ReaxFF MD program was created in the thesis to investigate the behaviors of coal pyrolysis process using the large-scale coal models constructed and with different ranks. The simulation results including product distributions and initial reaction mechanisms are reasonable, which is hardly accessible by experiments or other computational methods. The main results are summarized as the following.In order to reach spatio-temporal scales of nanometers and nanoseconds for complex coal pyrolysis, the first GPU-enabled ReaxFF MD program, GMD-Reax was created, both for the increase of simulation system size and acceleration of ReaxFF MD running on desktop workstations. Any possible care has been taken in the implementation to have the ReaxFF algorithms mapped to the proper organization of GPU threads and data structures, to minimize sources of overhead and to optimize kernel parameters to achieve excellent performance. The performance of GMD-Reax has been benchmarked for coal models with atoms ranging from 1378-27,283 against that of LAMMPS parallel running on 8 CPU cores. The overall performance of single precision GMD-Reax running on a computational node with a single C2050 GPU attached achieves speedups as high as 5.9-16.1 times over the FORTRAN code of ReaxFF in LAMMPS and 3.3-8.4 times over that of the C code. Accordingly, double precision GMD-Reax achieves speedups as high as 2.6-8.1 times against the FORTRAN code of ReaxFF in LAMMPS and 1.5-4.2 times against that of the C code. In addition, the paralleled implementation of ReaxFF on multi-GPUs (MPI-GMD-Reax) is developed by taking advantage of MPI. MPI-GMD-Reax has a capability for simulating a much more complex system with around 400,000 atoms when running on four processors with 8 K20C GPUs.In order to explore the coal pyrolysis behaviors for different coal ranks, large-scale coal models were constructed based on a combination of experiments (proximate and ultimate analysis, I3C NMR analysis) and classical coal models with atoms from-5000 to-28,000, including a hypothetical concept-proof bituminous coal model, Liulin bituminous coal models and Hailaer brown coal models. The largest Liulin coal molecular model consists of 28,351 atoms and is the second largest coal model ever simulated with ReaxFF MD simulation.The GMD-Reax simulations were performed at different conditions to investigate the effects of temperature (Hailaer brown coal:800-2600 K; Liulin bituminous coal: 1000-2600 K), heating rate (2,8,10,20,40 K/ps) and moisture on the coal pyrolysis with the large-scale coal molecular models constructed. The analysis of ReaxFF MD trajectories shows that the evolution tendencies of product profile (char, tar and gas) and main pyrolyzates (CO2, CO, naphthalenes, phenols and benzenes) as a function of time and temperature are in broad agreement with experimental results in literatures and Py-GC/MS experiments. The results obtained from product yield, element composition (C/H/O) and mass distribution with time and temperature show that 2000 K is the transition simulation temperature at which the dominant primary pyrolysis stage will turn into secondary pyrolysis stage where recombination of intermediates into coal char will dominate the pyrolysis process both for bituminous and brown coal model. The behaviors of coal pyrolysis between Hailaer brown coal and Liulin bituminous coal are similar in the main pyrolyzate (char, tar, gas) evolution and gas generation sequence (H2O, CO2, CO, CH4, H2). However, the simulation results indicates that differences exist between the pyrolysis of brown coal and bituminous coal pyrolysis in kinetics, the pyrolysis of Hailaer brown coal occurs at relatively low temperature, about 100-300 K lower than that of Liulin bituminous coal, the weight loss kinetic rate of Hailaer brown coal is higher than that of Liulin bituminous coal with several folds, less than a factor of 10.With the aid of VARxMD, the complex and detailed chemical reactions has been revealed from the ReaxFF MD simulations. The generation and consumption of HO·, H3O· and CHO2 radicals, aromatic ring-opening pathways and phenol generation mechanisms in coal pyrolysis are obtained and reasonable. In addition, the detailed process of bridge bond breaking during coal pyrolysis simulations was obtained for the first time, including the breaking sequence (-CH2-O->-COOH>-CH2-CH2-> Car-O-> Car-CH2-> Car-Car) and amounts of major bridge bonds, and their breaking percentage and evolution tendencies with time and temperature. These reaction mechanisms are hardly accessible by experiments or other computational methods.In particular, the model scale effects on the ReaxFF MD simulation trajectories for coal pyrolysis were discussed for the first time in this thesis. The results show that when the model scale is larger than 2000 atoms, ReaxFF MD simulation has good scalability for coal model size, yet the proper molecular model scale depends largely on the targets to be investigated by simulations. The coal model with-2000 atoms is large enough to investigate the generation trends of small gas molecules, while the observation of fragment evolution with more than six carbon atoms (benzene, phenol and naphthalene) requires larger-scale molecular models with atoms at least around 30,000 atoms. The large-scale coal models can describe more closely the diversity and complex properties of real coal, and their pyrolysis simulation results by ReaxFF MD could be more statistically based and thus more reasonable of the comprehensive understanding obtained for coal pyrolysis.Furthermore, the paper extended the coal pyrolysis systems to coal-like system, cellulose pyrolysis system, to investigate its initial reaction mechanisms for effective utilization of biomass. It is readily observable for the simulated evolution tendency of the major pyrolysis products (glycolaldehyde, levoglucosan and water) with time and temperature that agrees well with the Py-GC/MS experimental observations of cellulose pyrolysis at 673-1073 K and obtained from in literature. And further analysis of the simulation trajectories by VARxMD reveals some of the detailed initial chemical reaction schemes that might take place in cellulose pyrolysis.By comparing the main product evolution trends of cellulose pyrolysis between the short and long duration time in ReaxFF MD simulations, it is found that short time simulations at higher temperatures could resemble the evolution tendencies obtained from long time simulations at lower temperatures. These observations provide validation further the ground for the simulation strategy of the artificially increased high temperatures to avoid long simulations. Although it is very difficult to predict how much the increased simulation temperature will affect the compound generation, the reasonable evolving trends of main products obtained both in coal and cellulose pyrolysis (cellulose pyrolysis: appearing sequence of C6H10O5, C2H4O2and H2O maximum amount; coal pyrolysis:gas sequence and evolution tendency of naphthalenes) and their agreement with the Py-GC/MS experiments and literature suggest that direct simulation of large-scale model by ReaxFF MD is a promising method to predict pyrolysis behaviors.This work demonstrates a new methodology for investigating the comprehensive coal pyrolysis mechanism by direct simulation of large-scale coal models with GPU-enabled high performance computing code. With the unique capability of VARxMD, the underlying complex chemical reactions and important pathways in coal pyrolysis can be revealed. The methodology is applicable in other complicated molecular systems for a profound understanding of the complex chemical reactions in pyrolysis of very complicated coal molecular systems.