Combustion machine learning of superheated butanol atomization in optical GDI engine

Bio-butanol is recognized as a carbon-neutral alternative to gasoline in internal combustion engines due to its favorable physicochemical properties, though the diversity of butanol isomers and their blending ratios with gasoline complicates the standardized characterization of their combustion behavior. This study proposes a machine learning (ML) approach for combustion analysis by extracting 28 features from high-speed flame images captured during the combustion of butanol isomer/gasoline blends in an optically accessible single-cylinder GDI engine under two fuel injection conditions: regular and superheated. Gasoline surrogates—primary reference fuel (PRF) and toluene primary reference fuel (TPRF), both with a RON of 92, were blended with 30 vol% of four butanol isomers (n-butanol, isobutanol, sec-butanol, and tert-butanol), generating 10 fuel blends evaluated across 100 combustion cycles per blend under each injection condition, resulting in 2000 total cycles with 200 flame images per cycle. Four ML models, decision tree (DT), adaptive boosting (AdaBoost), random forest (RF), and support vector machine (SVM), were employed to classify combustion characteristics, revealing that flame area, apparent flame speed, premixed/diffused flame intensities, and flame grayscale surface fractal dimension are critical features for predicting flame conditions. By integrating flame image classification with feature importance analysis, this work establishes an ML approach to advance combustion research and optimize sustainable fuel application.