Finite volume method network for the acceleration of unsteady computational fluid dynamics: Non‐reacting and reacting flows

Abstract

Summary Despite rapid improvements in the performance of the central processing unit (CPU), the calculation cost of simulating chemically reacting flow using CFD remains infeasible in many cases. The application of the convolutional neural networks (CNNs) specialized in image processing in flow field prediction has been studied, but the need to develop a neural network design fitted for CFD has recently emerged. In this study, a neural network model introducing the finite volume method (FVM) with unique network architecture and physics‐informed loss function was developed to accelerate CFD simulations. The developed network model, considering the nature of the CFD flow field where the identical governing equations are applied to all grids, can predict the future fields with only two previous fields unlike the CNNs requiring many field images (>10 000). The performance of this baseline model was evaluated using CFD time series data from non‐reacting flow and reacting flow simulation; counterflow and hydrogen flame with 20 detailed chemistries. Consequently, we demonstrated that (a) the FVM‐based network architecture provided significantly improved accuracy of multistep time series prediction compared to the previous MLP model (b) the physic‐informed loss function prevented non‐physical overfitting problem and ultimately reduced the error in time series prediction (c) observing the calculated residuals in an unsupervised manner could monitor the network accuracy. Additionally, under the reacting flow dataset, the computational speed of this network model was measured to be about 10 times faster than that of the CFD solver.

Predicting Spatiotemporal Distributions in a Bubbling Fluidized Bed for Biomass Fast Pyrolysis Using Convolutional Neural Networks

“…Although past research has demonstrated the potential of DL in accelerating predictions of flow fields, most of them are short-term forecasts within two hundred steps, which is far from industrial applications. For instance, the FVMN model proposed by Jeon et al can predict ten future oil fields based on two. The model developed by Bazai can predict 170 steps but requires 18,000 frames for training.…”

A review of the application of artificial intelligence to nuclear reactors: Where we are and what's next

“…Ayodeji and Amidu [ 52 ] developed a surrogate model based on a deep feedforward neural network to predict the turbulent eddy viscosity in RANS simulations; the results closely matched those of an actual turbulent model. Jeon and Lee [ 53 ] constructed a neuron-network-based model to simulate the principles of the finite volume method (FVM) in fluid dynamics. The performance was evaluated using unsteady reacting flow datasets and showed good agreement with reference data at one-tenth of the computational cost.…”

A machine learning model for reconstructing skin-friction drag over ocean surface waves

“…2019; Kochkov et al. 2021; Jeon, Lee & Kim 2022), develop improved turbulence closures (e.g. Ling, Kurzawski & Templeton 2016; Wang, Wu & Xiao 2017; Beck, Flad & Munz 2019; Duraisamy, Iaccarino & Xiao 2019), reconstruct turbulent flow fields from spatially limited data (e.g.…”