Gradient boosted decision trees for combustion chemistry integration

Abstract

This study introduces the gradient boosted decision tree (GBDT) as a machine learning approach to circumvent the need for a direct integration of the typically stiff system of ordinary differential equations that govern the temporal evolution of chemically reacting species. Stiffness primarily relates to the chemistry integration and here, hydrogen/air systems are taken to train and test the ensemble learning approach. We use the LightGBM (Light Gradient Boosting Machine) algorithm to train GBDTs on the time series of various self-igniting mixtures from the time of ignition to equilibrium composition. The GBDT model provides reasonable predictions of the species compositions and thermodynamic states at the next time step in an a priori study. A much more challenging a posteriori study shows that the model can reproduce a full time–history profile of the igniting H2/air mixtures, as the results agree very well with those obtained from a direct integration of the ODEs. The GBDT model can be deployed as standalone C++ codes and a speed-up by one order of magnitude has been demonstrated. The GBDT approach can thus be considered as an efficient method to represent the chemical kinetics in the simulation of reactive flows. It provides an alternative to deep artificial neural networks (ANNs) that is comparable in accuracy but easier to couple with existing CFD codes.

Keywords

• Ensemble learning
• Gradient boosting
• Chemical kinetics
• Hydrogen combustion