To realize efficient computational fluid dynamics (CFD) prediction of two-phase flow, a multi-scale framework was proposed in this paper by applying a physics-guided data-driven approach. Instrumental to this framework, Feature Similarity Measurement (FSM) technique was developed for error estimation in two-phase flow simulation using coarse-mesh CFD, to achieve a comparable accuracy as fine-mesh simulations with fast-running feature. By defining physics-guided parameters and variable gradients as physical features, FSM has the capability to capture the underlying local patterns in the coarse-mesh CFD simulation. Massive low-fidelity data and respective high-fidelity data are used to explore the underlying information relevant to the main simulation errors and the effects of phenomenological scaling. By learning from previous simulation data, a surrogate model using deep feedforward neural network (DFNN) can be developed and trained to estimate the simulation error of coarse-mesh CFD. The research documented supports the feasibility of the physics-guided deep learning methods for coarse mesh CFD simulations which has a potential for the efficient industrial design.
Deep Dive into Computationally Efficient CFD Prediction of Bubbly Flow using Physics-Guided Deep Learning.
To realize efficient computational fluid dynamics (CFD) prediction of two-phase flow, a multi-scale framework was proposed in this paper by applying a physics-guided data-driven approach. Instrumental to this framework, Feature Similarity Measurement (FSM) technique was developed for error estimation in two-phase flow simulation using coarse-mesh CFD, to achieve a comparable accuracy as fine-mesh simulations with fast-running feature. By defining physics-guided parameters and variable gradients as physical features, FSM has the capability to capture the underlying local patterns in the coarse-mesh CFD simulation. Massive low-fidelity data and respective high-fidelity data are used to explore the underlying information relevant to the main simulation errors and the effects of phenomenological scaling. By learning from previous simulation data, a surrogate model using deep feedforward neural network (DFNN) can be developed and trained to estimate the simulation error of coarse-mesh CFD.
1
Computationally Efficient CFD Prediction of Bubbly Flow using Physics-Guided Deep
Learning
Han Bao1, Jinyong Feng2*, Nam Dinh3, Hongbin Zhang1
1 Idaho National Laboratory, P.O. Box 1625, MS 3860, Idaho Falls, 83415, ID, USA
2Massachusetts Institute of Technology, Cambridge, 02139, Massachusetts, USA
3North Carolina State University, Raleigh, 27695, North Carolina, USA
Abstract:
To realize efficient computational fluid dynamics (CFD) prediction of two-phase flow, a multi-
scale framework was proposed in this paper by applying a physics-guided data-driven approach.
Instrumental to this framework, Feature Similarity Measurement (FSM) technique was developed
for error estimation in two-phase flow simulation using coarse-mesh CFD, to achieve a comparable
accuracy as fine-mesh simulations with fast-running feature. By defining physics-guided
parameters and variable gradients as physical features, FSM has the capability to capture the
underlying local patterns in the coarse-mesh CFD simulation. Massive low-fidelity data and
respective high-fidelity data are used to explore the underlying information relevant to the main
simulation errors and the effects of phenomenological scaling. By learning from previous
simulation data, a surrogate model using deep feedforward neural network (DFNN) can be
developed and trained to estimate the simulation error of coarse-mesh CFD. In a demonstration
case of two-phase bubbly flow, the DFNN model well captured and corrected the unphysical
“peaks” in the velocity and void fraction profiles near the wall in the coarse-mesh configuration,
even for extrapolative predictions. The research documented supports the feasibility of the physics-
guided deep learning methods for coarse mesh CFD simulations which has a potential for the
efficient industrial design.
Keywords: deep learning, two-phase bubbly flow, coarse-mesh CFD, physical feature, data
similarity
2
- Introduction
Owing to the advancement of high-performance computing and computational methods,
modeling and numerical simulations have become instrumental in the design, analysis and
licensing of nuclear power plants. Compared to system codes using lumped-parameter models,
computational fluid dynamics (CFD) methods have been widely used for solving transport
equations of fluid mechanics by using local instantaneous formulations with finer mesh sizes,
where small-scale flow features could be captured. While CFD has the potential to accurately
predict the flow behavior and reduce the need for dedicated reactor-operational experiments, it
suffers from three key challenges for the system-level analysis of NPP behaviors, namely high
computational costs, user effects, and limited understanding on error sources of CFD simulation.
The main limitation of applying CFD methods to practical industrial applications is the
computational cost. Since discretizing the temporal and spatial space on a much smaller scale,
CFD simulations require many more cells than a system thermal-hydraulic simulation. One of the
most representative examples is direct numerical simulation (DNS) method. As a first principle
based method, DNS directly solves the Navier-Stokes equations without any closure models, thus
making it serve as high fidelity benchmark data, especially in the single-phase study. By coupling
with interface tracking method (Hirt and Nichols, 1981; Sussman et al., 1994; Unverdi and
Tryggvason, 1992), DNS extends its capability to simulate two-phase flow. In the state-of-the-art
two-phase DNS simulation (Fang et al., 2018), in order to resolve each individual bubble and
turbulent eddies down to the smallest turbulent length scale, i.e., Kolmogorov scale (Kolmogorov,
1941), it requires 1.10 billion cells and ~730,000 core-hours to simulate the reactor subchannel
with hydraulics Reynolds number of 80,000 whereas the hydraulics Reynolds number under
reactor operational conditions is ~500,000. Productive CFD simulations have to be performed on
large supercomputers rather than a multi-core computer. To bypass the computational cost of the
fully resolved high Reynolds number case, researchers either conducted separate effect studies
with well-controlled flow conditions (Bunner and Tryggvason, 2003; Feng and Bolotnov, 2017a,
2017b; Feng and Bolotnov, 2017) to develop individual closures, or adopt computational efficient
Reynolds-averaging Navier-Stokes method (Brewster et al., 2015; Feng et al., 2018).
3
Another aspect leading to the limitation of CFD simulation on system-level analysis is the user
effect, particularly on multiphase flow CFD. CFD codes are designed to be general flow solvers,
applicable to nearly every scale of flow problem encountered in engineering practice, spanning
from high Mach number compressible ai
…(Full text truncated)…
This content is AI-processed based on ArXiv data.