Neural Network Prediction of Boundary Layer Flashback

Abstract

Flashback is a major concern for engine operation and safety, particularly with progress toward renewably producible and cleaner-burning fuels, such as hydrogen fuel blends. This work extends prior progress in developing models for predicting the onset of boundary layer flashback. While prior attempts have developed models based on analytical theory or through phenomenological considerations, problem complexity has inhibited flashback understanding and, hence, model performance. The goal of this work is to address current model performance limitations by leveraging the representational flexibility offered by neural networks (NNs) in predicting boundary layer flashback. This is demonstrated through two applications. The first demonstrates the utility of training an NN on only a subproblem, thereby preserving model intuition. The second presents a predictive boundary layer flashback model using only a NN. Focus is placed on developing NN models which are practical; the input and output variables are easily measurable and controllable prior to experimentation.