Neural network approach to response surface development for reaction model optimization and uncertainty minimization
Yue Zhang, Wendi Dong, Laurien A. Vandewalle, Rui Xu, Gregory P. Smith, Hai Wang
Abstract
We examine the state-of-the-art neural network (NN) approach and its flexible implementations in combustion reaction model uncertainty quantification (UQ), optimization, and uncertainty minimization (UM). The work is motivated by addressing the problem of limited scalability of the traditional polynomial response surface methodology in handling large size of rate parameters and target data sets. Features of the NN training, accuracy, and trade-offs in several key aspects of the NN application are discussed. We show that for high-dimensional reaction model optimization and UM, a shallow NN with only one hidden layer is more robust and accurate than the polynomial response methodology. Further, we demonstrate that NN allows for adaptive training. New neural networks that augment new input parameters or updates in a trial reaction model can be adapted from the existing networks with much smaller training efforts. In addition, deep neural networks are capable of covering functional dependencies of initial thermodynamic conditions and boundary conditions, thus yielding generalized response surfaces with rate parameters and thermodynamic/mixture conditions as the input for a given combustion property. The NN approach can be readily integrated into the framework of the Method of Uncertainty Minimization using Polynomial Chaos Expansions (MUM-PCE) developed earlier. We present a test case that uses the trial Foundational Fuel Chemistry Model Version 2.0 (FFCM-2, a model consisting of 96 species and 1054 reactions for combustion of relevant C0–4 species), optimizing it against FFCM-1 targets, to illustrate the efficiency and accuracy of the NN method.