Online in situ prediction of 3-D flame evolution from its history 2-D projections via deep learning

Abstract

Online in situ prediction of 3-D flame evolution has been long desired and is considered to be the Holy Grail for the combustion community. Recent advances in computational power have facilitated the development of computational fluid dynamics (CFD), which can be used to predict flame behaviours. However, the most advanced CFD techniques are still incapable of realizing online in situ prediction of practical flames due to the enormous computational costs involved. In this work, we aim to combine the state-of-the-art experimental technique (that is, time-resolved volumetric tomography) with deep learning algorithms for rapid prediction of 3-D flame evolution. Proof-of-concept experiments conducted suggest that the evolution of both a laminar diffusion flame and a typical non-premixed turbulent swirl-stabilized flame can be predicted faithfully in a time scale on the order of milliseconds, which can be further reduced by simply using a few more GPUs. We believe this is the first time that online in situ prediction of 3-D flame evolution has become feasible, and we expect this method to be extremely useful, as for most application scenarios the online in situ prediction of even the large-scale flame features are already useful for an effective flame control.

Using machine learning to detect the turbulent region in flow past a circular cylinder

“…2019), flow field reconstruction and prediction (Maulik & San 2017; Maulik et al. 2018; Fukami, Fukagata & Taira 2019; Huang, Liu & Cai 2019; Lee & You 2019) and, more relevant to the present study, flow field identification (Colvert, Alsalman & Kanso 2018; Alsalman, Colvert & Kanso 2019; Wu et al. 2019 b ).…”

“…To avoid the above two limitations in the conventional methods, the machine learning method is useful. In recent years, machine learning has been used to study many problems in fluid mechanics, including turbulence modelling (Ling & Templeton 2015;Ma, Lu & Tryggvason 2015;Ling, Kurzawski & Templeton 2016b;Parish & Duraisamy 2016;Xiao et al 2016;Gamahara & Hattori 2017;Vollant, Balarac & Corre 2017;Wang, Wu & Xiao 2017;Wang et al 2018;Wu, Xiao & Paterson 2018;Duraisamy, Iaccarino & Xiao 2019;Wu et al 2019a;Zhou et al 2019), flow field reconstruction and prediction (Maulik & San 2017;Maulik et al 2018;Fukami, Fukagata & Taira 2019;Huang, Liu & Cai 2019;Lee & You 2019) and, more relevant to the present study, flow field identification (Colvert, Alsalman & Kanso 2018;Alsalman, Colvert & Kanso 2019;Wu et al 2019b). To solve the identification problems, there are two main machine learning approaches, namely, classification and clustering.…”

Artificial neural network based chemical mechanisms for computationally efficient modeling of hydrogen/carbon monoxide/kerosene combustion

“…Furthermore, U-net can not only convert the feature map into a vector, but also reconstruct the feature from this vector while greatly reducing sample requirements. Therefore, when considering the complex multiscale problem of combustion, a natural choice would be to us U-net as the basis and some successful attempts have been made [23,32]. However, the original U-net is meant for a segmentation task, the output layer is designed to represent a categorical distribution.…”

A deep learning framework for hydrogen-fueled turbulent combustion simulation

“…Furthermore, U-net can not only convert the feature map into a vector, but also reconstruct the feature from this vector while greatly reducing sample requirements. Therefore, when considering the complex multiscale problem of combustion, a natural choice would be to us U-net as the basis and some successful attempts have been made [26,35]. However, the original U-net is meant for a segmentation task, the output layer is designed to represent a categorical distribution.…”