A generative adversarial network (GAN) approach to creating synthetic flame images from experimental data

Abstract

Modern diagnostic tools in turbulent combustion allow for highly-resolved measurements of reacting flows; however, they tend to generate massive data-sets, rendering conventional analysis intractable and inefficient. To alleviate this problem, machine learning tools may be used to, for example, discover features from the data for downstream modeling and prediction tasks. To this end, this work applies generative adversarial networks (GANs) to generate realistic flame images based on a time-resolved data set of hydroxide concentration snapshots obtained from planar laser induced fluorescence measurements of a model combustor. The generative model is able to generate flames in attached, lifted, and intermediate configurations dictated by the user. Using k-means clustering and proper orthogonal decomposition, the synthetic image set produced by the GAN is shown to be visually similar to the real image set, with recirculation zones and burned/unburned regions clearly present, indicating good GAN performance in capturing the experimental data statistical structure. Combined with techniques for controlling the configuration of generated flames, this work opens new avenues towards tractable statistical analysis and modeling of flame behavior, as well as rapid and inexpensive flame data generation.