Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

  • Wall-bounded turbulence is an unresolved problem.
  • Interactions among coherent structures in the flow are explored.
  • A deep-learning method is used to predict the velocity field in time.
  • A game-theoretic algorithm is used to assess the importance of each structure.
  • Results are in agreement with previous observations in the literature.
  • The process has the potential to shed light on numerous fundamental phenomena.

Paper Content

Introduction

  • Osborne Reynolds published the first scientific article on turbulent flows 140 years ago
  • Navier-Stokes equations can only be solved analytically for elementary flow configurations
  • Turbulent flows are relevant to climate emergency, transportation, combustion, aerodynamics, energy generation, and urban pollution
  • Turbulence is a multi-scale phenomenon in both time and space
  • Integrating Navier-Stokes equations requires extremely fine computational meshes
  • Supercomputers are powerful enough to integrate these equations numerically
  • Direct numerical simulation (DNS) provides a complete flow characterization
  • DNS can provide a large amount of high-quality data
  • Coherent structures are flow regions associated with momentum transfer and turbulent-kinetic-energy (TKE) production
  • Explainable artificial intelligence (XAI) is used to gain a more profound knowledge of the flow physics
  • Deep convolutional neural networks (CNNs) and Shapley additive explanation (SHAP) values are used
  • Ejections and sweeps are the main contributors to the exchange of streamwise momentum
  • CNNs can predict the evolution of turbulent channel flow
  • SHAP ranks the importance of each structure for predicting the following flow field

Results

  • Flow is driven by a pressure gradient
  • Spatial coordinates are x, y, and z
  • Length and width of channel are Lx = 2πh and Lz = πh
  • Flow is statistically symmetric
  • Flow is driven by friction Reynolds number Reτ = 125
  • Q events are regions of high Reynolds stress
  • Q events can be divided into 4 types: outward interactions, ejections, inward interactions, and sweeps

Prediction of the velocity field

  • A CNN is trained to predict the evolution of the velocity field
  • Results show the high accuracy of the neural-network prediction
  • Simulated and predicted flow fields are nearly identical
  • Mean error between the simulation and the prediction is 0.47u τ
  • Maximum error is 2.33u τ , which is 3.5% and 15% of the bulk velocity

Explainability of the cnn predictions

  • CNN used to calculate contribution of turbulent structures to prediction
  • SHAP values used to assess importance of each Q event
  • Total importance of each typology of Q event calculated by summing up value of every structure belonging to class
  • Percentual importance of each class calculated by dividing total SHAP value of class by total SHAP value
  • SHAP density calculated to evaluate contribution of each structure to turbulent flow
  • Ejections most important in absolute terms, but sweeps most important per unit of volume
  • Two families of structures observed: wall attached and wall detached
  • Wall attached sweeps most influential structures per unit of volume
  • Ejections and sweeps form coherent pairs which transport shear stress
  • Most important structures elongated in streamwise direction
  • Total contribution per unit of volume of both ejections and sweeps compensates

Discussion and conclusions

  • XAI is used to predict the importance of intense Reynolds-stress structures
  • Sweeps and ejections are the most relevant structures for fully-developed turbulence in a channel
  • Computational tools are used to quantify the importance of each structure
  • Sweeps and ejections have a high influence on the flow
  • Larger structures correspond to wall-attached ejections and have the most negative SHAP values

Methods

Numerical simulations and flow case under study

  • Trained CNN using 4,900 velocity fields from DNS
  • Used LISO code to run simulations of wall-bounded turbulence

Deep-neural-network architecture and prediction

  • CNNs have been used to predict velocity fields
  • Equation 4 describes the convolution operation
  • The network used in this work consists of 4 layers of 3D CNN blocks

Explainability of the neural network

  • Deep learning can be used to predict complex relationships between inputs and outputs.
  • SHAP framework is used to evaluate the contribution of each feature to the prediction of the next flow field.
  • Mask function is used to convert from the original space into the space of the structures.
  • SHAP-kernel method is based on two components: LIME and Shapley values.
  • SHAP values explain the CNN prediction as a sum of the marginal contributions of the input features.
  • Reference matrix contains all zeros.
  • SHAP values are negative and including a feature reduces the error.
  • Most important structures are wall-attached ejections.
  • SHAP values per unit of volume are higher for wall-attached sweeps.
  • Correlation between SHAP values and Reynolds stress disappears when scaled by volume.
  • High-SHAP-density sweeps are the most important structures.
  • CNN uses ReLU and sigmoid activation functions.
  • SHAP-kernel algorithm is divided into different stages.
  • Results are predicted using the CNN for each structure.
  • Error between prediction and simulated solution is used to determine the SHAP value of each feature.