Bagged ensemble of ANN for predicting laminar flame speed of toluene reference fuels and syngas blends in SI engines

Abstract

Laminar Flame Speed (LFS) is a critical parameter that determines the burn rate in Spark Ignited (SI) engines, as it influences the laminar burnup process of air/fuel mixture entrained by turbulent flame. For various SI engine modeling approaches, accurate prediction of LFS under engine-relevant conditions is crucial. Syngas, often created from various renewable feedstocks, offers significant potential for mitigating engine knock and engine-out emissions. This study introduces a novel approach using ensembles of neural networks as a non-linear regression method to predict LFS of gasoline surrogates and syngas blends under engine-relevant conditions. To address the scarcity of experimental data, we performed 1-D flame simulations to generate a sufficiently large LFS dataset (225,200 cases in total) which was supplemented by available experimental data in literature. To mitigate the potential bias associated with chemical mechanism selection, multiple kinetic mechanisms were utilized for the flame simulations. Considering the thermodynamic conditions during flame propagation in SI engines, the dataset covers conditions from 300 K to 1000 K, 1 bar to 50 bar, and equivalence ratios from 0.8 to 1.2. Moreover, a large number of Toluene Reference Fuel (mixtures of iso-octane, n-heptane, toluene)/syngas blends up to 11,140 mixtures were explored. After evaluating multiple ML models, a two-hidden-layer neural network was selected for optimal performance. To improve robustness of predicted LFS, the neural network was further refined by employing ensembles of five such neural networks, each separately trained on 80 % of the dataset. The model developed in this study represents a substantial advancement in accurate and computationally efficient LFS prediction. It also introduces an effective approach to account for mechanism-associated bias by utilizing multiple chemical mechanisms.