Adaptive global pathway selection using artificial neural networks: A-priori study

Abstract

A novel chemistry reduction scheme incorporating neural networks and a pre-existing reduction algorithm called Global Pathway Selection (GPS) is proposed and validated. GPS is a state-of-the-art chemistry reduction tool that can provide accurate predictions for any given detailed mechanism. However, this accuracy becomes limited by the overfitting of some reactions for the entire combustion process. We propose a new adaptive GPS approach that conducts GPS reductions and creates multiple reduced mechanisms. The adaptive GPS method is then used to train a supervised neural network to enable the identification of important species for the given state vector. This network is then used to dynamically update the simulator's reduced mechanism according to any given condition. This method is referred to as supervised learning-aided GPS (SL-GPS) in this study. Because the mechanism is continuously modified to better represent the simulation's state, and each such modification requires only a call of a pre-trained network, this method significantly improves the accuracy of predictions using the reduced mechanisms with reduced computational cost. The adaptive GPS method is tested in 0D simulations of methane combustion using the GRI 3.0 detailed mechanism. The results obtained using the adaptive GPS and SL-GPS are compared to the predictions of the detailed mechanism and classical GPS. The SL-GPS method is shown to perform at greater accuracies than the classic GPS while using fewer species and reactions, and at a similar accuracy to the adaptive GPS with which the neural network was trained. The capabilities of adaptive and SL-GPS methods are even more profound for mechanisms with a greater number of species and reactions, i.e., PCRL-Mech1 mechanism for ethanol combustion which has been used for validation in this study.