An automated machine learning framework for piston engine optimization

Abstract

This study introduces an automated machine learning-based optimization framework for internal combustion engine applications. This framework includes a SuperLearner model, an ensemble of several base learners combined with optimization algorithms, and an active learning approach. The SuperLearner model's performance was maximized by optimizing the hyperparameters using an elitist-based genetic algorithm. This work aims to demonstrate the framework on a smaller dataset of size 64 sourced randomly from the literature dataset for a heavy-duty gasoline compression ignition engine. The database consisted of nine input features representing engine control parameters and five target features representing engine performance and emissions. The multi-objective problem was converted into a single objective problem to maximize a merit value. A merit value of 103.9 was achieved using this framework, which was higher than the best merit values found within the reduced (101.97) and source (103.2) datasets. This high merit value was obtained with about 96% lesser data than the source dataset. This study demonstrated that the developed framework could lead to optimum solutions, starting with smaller datasets and fewer additional experimental or simulation runs.