Combining machine learning with 3D-CFD modeling for optimizing a DISI engine performance during cold-start

Abstract

This work presents a methodology for using machine learning (ML) techniques in combination with 3D computational fluid dynamics (CFD) modeling to optimize the cold-start fast-idle phase of a gasoline direct injection spark ignition (DISI) engine. The optimization process implies the identification of the range of operating parameters, that will ensure the following criteria under cold-start conditions: (1) a fixed IMEP of 2 bar (BMEP of 0 bar), (2) a stoichiometric exhaust equivalence ratio (based on carbon-to-oxygen atoms) to ensure the efficient operation of the after-treatment system, (3) enough exhaust heat flux to ensure a rapid light-off of the after-treatment system, and (4) reduced NOx and HC emissions. A total of six operating parameters will be identified as having a significant influence on cold-start engine performance. These parameters are associated with the fuel injection strategy (end of the second injection, injection pressure, and fuel mass); combustion strategy (spark timing, spark energy); and intake airflow (intake manifold pressure). Performing an optimization study exclusively using multi-cycle (at least 3 cycles) 3D CFD simulations would be an arduous task. For example, to achieve an exhaust equivalence ratio of 1 and an IMEP of 2 bar, multiple iterations would be required for fuel mass (to account for film formation), intake manifold pressure (to ensure enough air in-cylinder), and spark timing (to ensure the fixed load). This process would be more convoluted and expensive with the addition of constraints for exhaust heat flux and emissions. A promising approach to tackling such a complicated optimization process is to employ the concept of machine learning, which demands a database formed by the six operating parameters mentioned above. The current work will demonstrate a strategy of combining CFD modeling with advanced Gaussian Process Regression (GPR)-based ML models to make predictions about DISI cold-start behavior with acceptable accuracy and a substantially reduced computational time.