Predicting combustion variability using machine learning from the flow field data at spark timing for a gasoline direct injection engine

Machine learning was used to predict three combustion metrics based on the flow field data at spark angle ( SA), obtained from large eddy simulations (LES). These metrics were peak cylinder pressure (PCP), crank angle of PCP (CA of PCP), and indicated mean effective pressure (IMEP). The ML model was evaluated for use as a surrogate for detailed combustion model in studies of cycle-to-cycle variation (CCV). The computational time for combustion predictions (from SA to exhaust valve opening (EVO)) was 15 hours on 24 cores while the trained ML model was able to make predictions in 3 seconds ( a speed up factor of 7200). The ML model was trained on a single core in 45 minutes using 88 LES simulation cycles. As the LES simulation data at SA has high fidelity (in this case approximately 260,000 computational cells) a workflow to filter the data to a reasonably sized feature set for use in ML was developed. Subvolumes were defined and the Pearson Correlation Coefficient (PCC) was used to evaluate the features. An ensemble ML approach was employed, which combines the predictions of multiple base learners to improve the final prediction. The ML model predictions for the three combustion metrics showed good correlation with the CFD simulations.