nVidia Modulus Tutorial – Lid Driven Cavity Flow讀後心得

學習目標

• 使用Modulus中Geometry module產生2D Geometry
• 設定BC-Boundary Conditions
• 設定待解之PDF( Flow Equation)
• Loss function和network調教
• 使用Modulus作後處理

問題描述

上邊界以1m/s速度朝x方向移動，其餘邊界皆為靜止。雷諾數(Reynold number)設定為10。

使用案例設定

以PINNs為基礎求解PDE的步驟：設定幾何形狀設定物理現象所對應之PDE產生neural network solver。如果對應到Modulus，所使用的module與class分別為：CSG/STL(產生point cloud)、TrainDomain class與Solver class。問題與Modulus的對應如上圖所示。

使用Modulus解決LDC包含以下幾個步驟：
Step 1：Importing the required packages

from sympy import Symbol, Eq, Abs

from modulus.solver import Solver
from modulus.dataset import TrainDomain, ValidationDomain
from modulus.data import Validation
from modulus.sympy_utils.geometry_2d import Rectangle
from modulus.csv_utils.csv_rw import csv_to_dict
from modulus.PDES import NaviewStokes
from modulus.controller impotr ModulusController

Step 2：產生幾何形狀

# make geometry
height = 0.1
width = 0.1
vel = 1.0

# define geometry
rec = Rectangle((-width / 2, -height / 2), (width / 2, height / 2))
geo = rec

可使用sample_boundary產生sample point，並使用var_to_vtk產生Paraview可顯示之.vtu檔。

samples = geo.sample_boundary(1000)
var_to_vtk(samples, './geo')

Step 3：定義邊界條件與設定LDC所對應之PDE

# define sympy varaibles to parametize domain curves
x, y = Symbol("x"), Symbol("y")

class LDCTrain(TRainDomain):
  define __init__(self, **config)
  super(LDCTrain, self)._init_()
  
  #top wall
  topwall = geo.boundary_bc(outvar_sympy={'u': vel, 'v': 0},
                        batch_size_per_area=10000,
                        lambda_sympy={'lambda_u': 1.0 -20 * Abs(x), # weight edges to be zero
                                      'lambda_v': 1.0},
                        criteria=Eq(y, height / 2))
   self.add(topwall, name="Topwall")
   
   # no slip
   bottomwall = geo.boundary_bc(outvar_sympy={'u': 0, 'v': 0},
                                batch_size_per_area=10000,
                                criteria-y < height / 2)
   self. add(bottomwall, name="NoSlip")

   # interior
   interior = geo.interior_bc(outvar_sympy={'continuity': 0, 'momentum_x': 0, 'momentum_y': 0},
                              bounds={x: (-width / 2, width / 2),
                                      y: (-height i 2, height 1'2)},
                              lambda_sympy={'lambda_continuity': geo.sdf,
                                            'lambda_momentum_x': geo.sdf,
                                            'lambda_momentum_y': geo.sdf},
                              batch_size_per_area=400000)
    self.add(interior, name="Interior")

Modulus solver的training data的來源為系統之BC與PDE，可由TrainDomain定義。另外，使用solver求解必須定義相關的loss function。
Boundary Condition：使用boundary_bc函數設定。
Equations to solve：考慮2D LDC，我們必須求解continuity equation與x,y方向的momentum equation。初始值為零表示不存在forcing/source項目。
Lambda_sympy表示loss function weighting。下圖可知在左上角與右上角有最大的不連續性，利用SDF(Signed Distance Field)設定可獲得較快的收斂與避免在邊界的不連續性。

Step 4：產生驗證資料

# validation data
mapping = {'Points:0': 'x', 'Points:1': 'y', 'U:0': 'u', 'U:1': 'v', 'p': 'p'}
openfoam_var = csv_to_dict('openfoam/cavity_uniformVe10.csv, mapping)
openfoam_var['x'] t= -width / 2 # center OpenFoam data
openfoam_var['y'] t= -height / 2 # center OpenFoam data
openfoam_invar_numpy = {key: value for key, value in openfoam_var.items() if key in ['x', 'y']}
openfoam_outvar_numpy = {key: value for key, value in openfoam_var.items() if key in ['u', 'v']}
class LDCVal (ValidationDomain):
  def_init_(self, **config):
  super(LDCVal, self).__init_
  val = Validation. from_numpy(openfoam_invar_numpy, openfoam_outvar_numpy)
  self.add(val, name='Val')

Modulus solver唯一基於物理條件之neural network solver，因此並不需要training data，但我們可以將CFD或其他solver的結果匯入為validating data。我們將OpenFOAM的計算結果以.csv方式匯入，視為solver的validation domain。

Step 5：建立Neural Network Solver

class LDCSolver (Solver):
  train_domain = LDCTrain
  val_domain = LDCVal

  def _init_(self, **config):
    super (LDCSolver, self).__init__(**config)
    self.equations = NavierStokes(nu=0.01, rho=1.0, dim=2,
                                  time=False).make_node()
    flow_net = self.arch.make_node(name='flow_net',
                                   inputs=['x', 'y'],
                                   outputs=['u', 'v', 'p'])
    self.nets = [flow_net]

    @classmethod
    def update_defaults(cls, defaults):
      defaults.update ({
          'network_dir': './network_checkpoint_ldc_2d',
          'decay_steps': 4000,
          max_steps':400000
      })
      
if __name__ == '__main__':
  ctr = ModulusController(LDCSolver)
  ctr.run()

solver繼承Modulus內建的Solver Class，需要設定train_domain與val_domain。
LDC所遵循的PDE為Navier-Stokes方程式，Modulus已有內建，需要給定的參數為：dim=2(2D)、ν(kinematic viscosity)=0.01 m2/sec與ρ(密度)=1.0kg/m31.

$$Re=\frac{\rho VD}{u}=\frac{\nu D}{\gamma }=\frac{1.0\times 0.1}{0.01}=10$$

Modulus預設的network包含6個hidden layer，每一個layer包含512個node。指定network的輸入為(x,y)，輸出為(u,v,p)

Step 6：執行Neural Network Solver

python ldc_2d.py