analysis of a data driven closure model trained from a reacting DNS of a Low-Swirl Burner

Machine learning (ML) advances coupled with the exponential increase in supercomputing capabilities presents tremendous promise in developing complex models for reacting flows from direct numerical simulations (DNS). In this study, we describe a model development effort via application of machine learning techniques to a reacting flow DNS. Specifically, we describe our work in applying various ML models to data from reacting DNS of a Low-Swirl Burner (LSB) [1].

In the current study, supervised learning techniques are utilized within the class of deep learning algorithms to investigate reacting flow sub-grid models. DNS data, in the form of moments and dissipation rates of mixture fraction and progress variable, are used as the input parameters. A-priori analysis is used to demonstrate the efficacy of the ML techniques to generate a sub-grid representation of the source term for progress variable. Models generated from both Random forest and a Deep Neural Network with 10 hidden layers and 20 nodes are compared. Both methods show tremendous promise and are found to produce peak conditional means within 15% of the DNS data across multiple filter widths.

References

[1] Day M, Tachibana S, Bell J, Lijewski M, Beckner V, Cheng RK. Combust Flame. 2012;159(1):275-290