Data-driven modeling of Dual-Dilution Combustion in Advanced Spark Ignition engines using Machine Learning

Nowadays, it is fundamental to have fast and reliable virtual tools to accelerate the development of robust and efficient engines, in order to meet the increasingly stringent regulations introduced by the European Commission on both pollutant and greenhouse gas (GHG) emissions. In particular, when an ultra-lean dual-dilution approach is implemented, as in this project, conventional 0D/1D CFD simulation software can be difficult and highly time-consuming to calibrate for achieving a reliable combustion model. Therefore, Machine Learning (ML) and Artificial Neural Network (ANN) can represent a valid alternatives to predict the combustion profile under these challenging conditions. The aim of this thesis is to establish three different ML and ANN models for a ultra lean dual-dilution engine, based on experimental data obtained during the development of the PHOENICE (PHev towards zerO EmissioNs & ultimate ICE efficiency) H2020 project. More in detail, a fully connected neural network was designed to predict the Wiebe parameters, while a Gaussian Process Regression (GPR) model and an additional fully connected neural network were both developed to directly predict the burn rate curves. These three models were able to capture the non-linear relationship between the core control variables and combustion with good accuracy, predicting the main combustion characteristics such as Mass Fraction Burned (MFB-10, MFB-50) and the combustion duration (MFB10–75). The best performing model, the fully connected neural network designed for burn rate prediction, was able to reach regression values between 0.9299 and 0.9953, while the root mean square errors between the ANN predicted and the experimental measurements were within the range of 0.57–0.90 °CA