Acceleration of the complex reacting flow simulation with a generalizable neural network based on meta-learning

Abstract

A chemical decoupled method is proposed in this paper. To accelerate the complex reacting flow simulation, the chemistry and flow computation is decoupled by a Neural Network (NN). A chemical kinetics database including 2,106,496 flow states and corresponding reacting source terms has been generated. Nevertheless, the data presents imbalance and spans a wide range of magnitudes, which challenges the NN training. The normalization method, NN’s construction, and performance evaluation indicators are optimized to address this challenge. Another problem is that different processes, such as transport or reaction, dominate different flow regions. This phenomenon requires the NN sufficient generalization to accurately map the complex reactions and flow states, even when the dominant process is different. The generalization of the NN is improved through the meta-learning strategy, the residual connection, and the inter-layer sine normalization. The effect of these enhancements is tested by noisy data and simulations with different turbulence model and boundary conditions. The method has been demonstrated in a supersonic combustion simulation with 16 species and 137 reactions. The maximum mean square relative error between the simulation with NN and the traditional simulation is 13.59 %. The calculation time is reduced from 6415.989 min (about 4.456 days) of the traditional simulation to 1983.576 min (about 1.377 days) of the simulation with NN. This method has application potential in accelerating the complex reacting flow simulation.