I will be attending the ASME Internal Combustion Engine Division Fall technical conference (ICEF2013) the upcoming week in Dearborn, Mi. I will be presenting the summary of my studies on chemical kinetics acceleration for multi-component fuel combustion simulations that I've completed throughout 2012. My presentation will be on October 14th at 4:00p within the Numerical Simulation session. The study, named Computationally Efficient Simulation of Multi-Component Fuel Combustion Using a Sparse Analytical Jacobian Chemistry Solver and High-Dimensional Clustering (details on the ASME website), was carried out with the collaboration of Dr. Anand Krishnasamy, Dr. Youngchul Ra and Prof. Rolf Reitz of the University of Wisconsin and focused on the development of algorithms for speeding up detailed chemistry calculations when in presence of multiple or multi-component fuels. The study tackled two major bottlenecks:
- Grouping cells with similar reactivity to reduce the number of chemical kinetics integrations cannot rely on simplified mixture state parameters such as the widely used equivalence ratio, as similar phi-T values can correspond to extremely different reactivities if different fuels are present in the cell. For this reason, I developed a high-dimensional clustering (HDC) algorithm that builds a high-dimensional representation of the chemical state space in multiple variables, and performs cell clustering using a new algorithm of the k-means class, called BBC-kmeans, that is tailored for clustering data in many dimensions, where the points are typically extremely sparse. The most important feature of this method is that it is rigorously tolerance-bounded, and the speed-up is sacrificed in case the requested accuracy cannot be achieved. Nevertheless, this allowed huge speed-ups of about three times even with very coarse grids, and showed logarithmic increase in the speed-up with increasing numbers of cells.
- The computational time required by the integration of chemical kinetics when multi-component fuel ignition mechanisms are used. This was challenging as on-the-fly reduction methods are not viable for skeletal multi-component mechanisms that can have as few as about one hundred species. The advanced numerics in SpeedCHEM allowed about one order of magnitude speedup in comparison with a reference chemistry code.
The 30-fold CPU time reduction achieved on coarse grids is now being tested on multi-dimensional engine grids for real-world engine geometries with up to one million cells.