Applying machine learning techniques to predict laminar burning velocity for ammonia/hydrogen/air mixtures

Abstract

Ammonia utilisation in internal combustion engines has attracted wide interest due to the current trend toward decarbonisation, as ammonia is a zero-carbon fuel with different combustion properties to hydrocarbons. The laminar burning velocity (LBV) is a fundamental property of fuels with a significant effect on the combustion processes and accurate calculations and measurements of the LBV over a wide range of fuel blends, pressures and flow conditions is a time-consuming, complicated procedure. The main goal of the current study is to predict the LBV of NH3/H2/air mixtures using a hybrid machine learning (ML) approach based on a training dataset consisting of both the experimental LBV values and additional data obtained from numerical simulations with a detailed kinetic model. Initial ML model training data is collected from existing experimental LBV in the literature for NH3/H2/air mixtures. Then, synthetic data is generated using one-dimensional (1D) simulations to reduce data inhomogeneity and increase accuracy of the ML model. In total, 24 different ML algorithms are tested to find the best model both for the experimental and the hybrid dataset. The results suggest that both Gaussian Process Regression (GPR) and Neural Networks (NNs) can be utilised to predict LBV of NH3/H2/air mixtures with reasonable accuracy. The hybrid ML model achieved a coefficient of determination of R2 = 0.998. Finally, hybrid ML model hyperparameters are optimised to achieve a coefficient of determination of R2 = 0.999. It was also found that ML can speed up LBV computation from 9500 to 27000 times compared to 1D simulations with a reduced mechanism.