Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Acceleration of RANS Solver Convergence via Initialization with Wake Extension Models

Kazuko W. Fuchi, Eric M. Wolf, Christopher R. Schrock, Philip S. Beran

TL;DR

This work tackles the high computational cost of steady-state RANS simulations by warm-starting runs with wake-aware initial fields. It introduces a three-region domain and a CNN-based wake extender, combined through a partition-of-unity (POFU) scheme to produce full-domain initial fields from region-specific predictions. The CNN wake extender, trained from a single CFD simulation, recursively predicts downstream wake development and yields large convergence gains (up to 16.4× faster wall-clock time and 26.3× fewer iterations for a NACA0012 case) compared with freestream initialization, outperforming simpler wake models that require wake extension far downstream. The results suggest that representing the entire wake is key to maximizing acceleration, with significant implications for accelerating design cycles and guiding generalization to broader flow conditions.

Abstract

Use of appropriate initialization to warm-start Reynolds-averaged Navier-Stokes (RANS) simulations of turbulent flow can facilitate convergence and lead to efficient use of computational resources. In this work, a method to model downstream wake development in external turbulent flow is proposed and used for RANS solver convergence acceleration. To balance the model accuracy and cost, the proposed method divides the analysis domain into three regions: near-body, wake and off-body. An approach based on a convolutional neural network is introduced as an efficient method to predict the downstream wake development. The model training only requires data from a single simulation, and its use is demonstrated to be effective in accelerating the RANS simulation when combined with an accurate flow prediction in the near-body region. The simulation using the proposed method took 26.3x fewer iterations, achieving 16.4x speedup in wall-clock time, compared to a baseline run using freestream initialization.

Acceleration of RANS Solver Convergence via Initialization with Wake Extension Models

TL;DR

This work tackles the high computational cost of steady-state RANS simulations by warm-starting runs with wake-aware initial fields. It introduces a three-region domain and a CNN-based wake extender, combined through a partition-of-unity (POFU) scheme to produce full-domain initial fields from region-specific predictions. The CNN wake extender, trained from a single CFD simulation, recursively predicts downstream wake development and yields large convergence gains (up to 16.4× faster wall-clock time and 26.3× fewer iterations for a NACA0012 case) compared with freestream initialization, outperforming simpler wake models that require wake extension far downstream. The results suggest that representing the entire wake is key to maximizing acceleration, with significant implications for accelerating design cycles and guiding generalization to broader flow conditions.

Abstract

Use of appropriate initialization to warm-start Reynolds-averaged Navier-Stokes (RANS) simulations of turbulent flow can facilitate convergence and lead to efficient use of computational resources. In this work, a method to model downstream wake development in external turbulent flow is proposed and used for RANS solver convergence acceleration. To balance the model accuracy and cost, the proposed method divides the analysis domain into three regions: near-body, wake and off-body. An approach based on a convolutional neural network is introduced as an efficient method to predict the downstream wake development. The model training only requires data from a single simulation, and its use is demonstrated to be effective in accelerating the RANS simulation when combined with an accurate flow prediction in the near-body region. The simulation using the proposed method took 26.3x fewer iterations, achieving 16.4x speedup in wall-clock time, compared to a baseline run using freestream initialization.
Paper Structure (14 sections, 9 equations, 16 figures, 4 tables, 1 algorithm)

This paper contains 14 sections, 9 equations, 16 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Method
  3. Model Ensemble through POFU
  4. Wake Extension Models
  5. Freestream in the wake
  6. Uniform wake extension
  7. CNN wake extender model
  8. ML Initialization of CFD Simulations
  9. Results
  10. Data Preparation and Wake Model Training
  11. Data preparation - turbulent flow around NACA0012 airfoil
  12. CNN wake extension model training
  13. CFD Initialization Study
  14. Conclusion

Figures (16)

  • Figure 1: Three modeling regions of the analysis domain.
  • Figure 2: POFU window functions for (a) near-body, and (b) wake regions, with $W=1$ in white and $W=0$ in gray areas.
  • Figure 3: Examples of flow field using different wake models: (a) freestream conditions, (b) uniform wake extension, and (c) CNN wake extension model.
  • Figure 4: CNN model is used to predict the field difference downstream.
  • Figure 5: Wake extension through the recursive use of CNN model.
  • ...and 11 more figures