This paper reports on a fast predictive combustion tool employing detailed chemistry. The model is a stochastic reactor based, discretised probability density function model, without spatial resolution. Employing detailed chemistry has the potential of predicting emissions, but generally results in very high CPU costs. Here it is shown that CPU times of a couple of minutes per cycle can be reached when applying detailed chemistry, and CPU times below 10 seconds per cycle can be reached when using reduced chemistry while still catching in-cylinder in-homogeneities. This makes the tool usable for efficient engine performance mapping and optimisation.
To meet CPU time requirements, automatically load balancing parallelisation was included in the model. This allowed for an almost linear CPU speed-up with number of cores available. As the number of cores increased, temporarily idle CPU's and computer cluster overhead cost was found to start affecting the overall CPU cost, but speed-up was observed up to 200 cores.
A clustering algorithm allowing for any number of controlling parameters was further utilised. The algorithm clusters the different particles based on user provided parameters and dispersion thresholds. After finishing the chemistry step, the clustered solutions are mapped back to the individual particles while preserving each individual particle's distance from its cluster mean. The clustering algorithm was found to give a larger CPU speed-up the more particles were used and also to be effective both for detailed and reduced chemical mechanisms.