Aiding the Development of Molecular-Level Kinetic Mega Models through Parallelization

Many processes in the energy and petrochemical industries involve large numbers of reactions and/or species. In order to optimize the molecular output of these processes a map of the molecular landscape must be understood. Through the use molecular-level models this information can be deciphered and used to probe/optimize these processes at molecular resolution. As expected, molecular-level models of these systems can quickly reach mega levels sizes in terms of reactions and/or species. Model development at these scales becomes time consuming and cumbersome in each of the model building, solving, parameter estimation, and editing phases. Advances in computer science have allowed for the development of new modeling frameworks that can be tailored to decrease the time spent and dynamically scale to variations in reaction network sizes. The Dynamic Model Builder (DMB) is a C++ based modeling framework that accounts for varying model sizes while functioning independently of program compilation by the virtue that the model can be created and simulated directly from system memory. The DMB framework was further enhanced to deal with large systems through parallelization strategies of the reactions, mole balances, and ODE solver routines on a high performance cluster (HPC). DMB’s reactions and the mass balances were parallelized on the CPU using OpenMP/MPI while the ODE implicit solver routine was parallelized on a CPU-GPU hybrid system using CUDA based MAGMA GPU libraries for the LU decomposition routine. These parallelization approaches decreased simulation and kinetic parameter estimation time of large molecular systems. To illustrate the robustness of this modeling framework a lignin structure was developed, reacted based on pyrolysis chemistry, and modeled between the temperatures of 300-500 °C. Optimization of kinetic parameters was achieved from literature experimental data and produced a model output that shows good agreement with experimental values.