Machine Learning Methods for Early Prediction of Sustained Ignition

We use machine learning methods to seek means of early prediction for sustained ignition in reactive flow simulations. Three configurations are considered: a geometrically complex simulation of a turbulent H₂-air jet in crossflow in addition to two simpler ersatz models for this implemented on computationally cheap 2D and 3D cubic meshes. The objective is to correctly classify the ignition attempts as either igniting or non-igniting, based on the an early stages of the simulation. Several machine learning implementations are evaluated, ranging from a simple two-layer perceptron to two- and three-dimensional convolutional neural networks (CNN) with multiple data channels. Logarithmic pre-conditioning of radical concentrations is found to significantly improve the machine learning performance. The performance of CNN's is evaluated both when trained and tested on the same type of data (full jet in crossflow simulation vs. ersatz case) and when a network trained on computationally cheap ersatz data is applied to the jet in crossflow case. In the latter case, we explore the fidelity required of the ersatz model in order to achieve good predictive accuracy when applied to the jet in crossflow.