Pre/Post deep learning

Some deep learning workflows applied to physics contexts require the projection of fields defined on an unstructured mesh onto a rectilinear grid, and inversely.

Includes

[1]:
import pyvista
pyvista.global_theme._jupyter_backend = 'panel' # remove this line to get interactive 3D plots
import numpy as np
from Muscat.Bridges.PyVistaBridge import PlotMesh, MeshToPyVista
from Muscat.Helpers.IO.TemporaryDirectory import TemporaryDirectory

PREPROCESSING

Read the solution generated by a physical code

[2]:
from Muscat.IO import XdmfReader as XR
reader = XR.XdmfReader(filename = "PrePostDeepLearning_Input.xmf")
reader.Read()
grid = reader.xdmf.GetDomain(0).GetGrid(0)

# Read the mesh
uMesh = grid.GetSupport()

# Read the solution field 'U'
indexU = grid.GetPointFieldsNames().index("U")
U = grid.GetPointFields()[indexU][:,0:2]
uMesh.nodeFields["U"] = U
PlotMesh(uMesh,scalars="U", show_edges=True, cpos="xy", show_scalar_bar=False)
../_images/notebooks_PrePostDeepLearning_6_0.png

Create a rectilinear mesh of size 48*48 excluding the left part of the mesh (x<0)

[3]:
Nx = 48
Ny = 48
boundingMin, boundingMax = uMesh.ComputeBoundingBox()
Lx = boundingMax[0] - 0.
Ly = boundingMax[1] - boundingMin[1]

from Muscat.Containers.ConstantRectilinearMeshTools import CreateConstantRectilinearMesh
unstructuredRectMesh = CreateConstantRectilinearMesh(dimensions=[Nx,Ny], spacing=[Lx/(Nx-1), Ly/(Ny-1)], origin=[0., boundingMin[1]])
PlotMesh(unstructuredRectMesh, show_edges=True, cpos="xy")
../_images/notebooks_PrePostDeepLearning_8_0.png

Compute the projection operator from unstructured mesh to structured mesh

[4]:
from Muscat.FE.FETools import PrepareFEComputation
from Muscat.Containers.MeshFieldOperations import GetFieldTransferOp
from Muscat.FE.Fields.FEField import FEField

inputFEField = FEField(name="U",mesh=uMesh)

operator, status, entities  = GetFieldTransferOp(inputFEField, unstructuredRectMesh.nodes, method = "Interp/Clamp", verbose=True)

# Compute the projected field on the structured mesh
projectedU = operator.dot(U)

unstructuredRectMesh.nodeFields["projectedU"] = projectedU
PlotMesh(unstructuredRectMesh,scalars="projectedU", show_edges=True, cpos="xy", show_scalar_bar=False)
Starting C++ FieldTransfer
Done c++ FieldTransfer
../_images/notebooks_PrePostDeepLearning_10_1.png

Export the structured mesh and projected field in xdmf format (in ASCII)

[5]:
# To visualize this xdmf file you can use ParaView (downloadable from https://www.paraview.org/)
from Muscat.IO import XdmfWriter as XW
writer = XW.XdmfWriter(TemporaryDirectory.GetTempPath() + 'PrePostDeepLearning_OutputI.xdmf')
writer.SetHdf5(False)
writer.Open()
writer.Write(unstructuredRectMesh,PointFields=[projectedU], PointFieldsNames=["U"])
writer.Close()

POSTPROCESSING

[6]:
# Compute the projection operator from the structured mesh to the unstructured mesh (inverse projection)
inputFEField = FEField(name="U",mesh=unstructuredRectMesh)
operator, status, entities = GetFieldTransferOp(inputFEField, uMesh.nodes, method =  "Interp/Clamp", verbose=True)

# Compute the inverse-projected projected field on the unstructured mesh
inverseProjected_ProjectedU = operator.dot(projectedU)

uMesh.nodeFields["inverseProjected_ProjectedU"] = inverseProjected_ProjectedU
PlotMesh(uMesh,scalars="inverseProjected_ProjectedU", show_edges=False, cpos="xy", show_scalar_bar=False)
Starting C++ FieldTransfer
Done c++ FieldTransfer
../_images/notebooks_PrePostDeepLearning_14_1.png

CHECK ACCURACY

[7]:
# Create a clippedMesh from the unstructured mesh by removing the left part (x<0)
# (with field, inverse-projected projected field and difference between them)
from Muscat.Containers.Filters.FilterObjects import ElementFilter
from Muscat.FE.FETools import ComputePhiAtIntegPoint, ComputeL2ScalarProductMatrix
from Muscat.Containers.MeshInspectionTools import ExtractElementsByElementFilter
from Muscat.Containers.MeshModificationTools import CleanLonelyNodes
from Muscat.Containers.MeshFieldOperations import CopyFieldsFromOriginalMeshToTargetMesh

uMesh.nodeFields['delta'] = U - inverseProjected_ProjectedU

elFilter = ElementFilter(zone = lambda p: (-p[:,0]))
meshClipped = ExtractElementsByElementFilter(uMesh, elFilter)
CleanLonelyNodes(meshClipped)

CopyFieldsFromOriginalMeshToTargetMesh(uMesh, meshClipped)
nbeOfNodes = meshClipped.GetNumberOfNodes()

deltaClippedMesh = meshClipped.nodeFields['delta']

PlotMesh(meshClipped,scalars="delta", show_edges=False, cpos="xy", show_scalar_bar=True)
../_images/notebooks_PrePostDeepLearning_16_0.png

Computetion of the integral of the delta

[8]:
from Muscat.Helpers.TextFormatHelper import TFormat
from Muscat.Helpers.Timer import Timer

print("Compute the L2(Omega) norm of delta by applying numerical quadrature from Lagrange P1")
print("Finite element integration using three different methods")

#### Method 1 #######################################################################################################################
print(TFormat.Center(TFormat.InRed("Method 1: ")+TFormat.InBlue("'by hand'")))

timer = Timer("Duration of method 1")
#compute method 1 three times
for i in range(2):
    timer.Start()
    integrationWeights, phiAtIntegPoint = ComputePhiAtIntegPoint(meshClipped)

    vectDeltaAtIntegPoints = np.empty((2,phiAtIntegPoint.shape[0]))
    for i in range(2):
        vectDeltaAtIntegPoints[i] = phiAtIntegPoint.dot(deltaClippedMesh[:,i])

    normDeltaMethod1 = np.sqrt(np.einsum('ij,ij,j', vectDeltaAtIntegPoints, vectDeltaAtIntegPoints, integrationWeights, optimize = True))
    timer.Stop()

print("norm(Delta) =", normDeltaMethod1)
############
#### Method 2 #######################################################################################################################
print(TFormat.Center(TFormat.InRed("Method 2: ")+TFormat.InBlue("using the mass matrix")))

timer = Timer("Duration of method 2").Start()

massMatrix = ComputeL2ScalarProductMatrix(meshClipped, 2)

normDeltaMethod2 = np.sqrt(np.dot(deltaClippedMesh.ravel(order='F'), massMatrix.dot(deltaClippedMesh.ravel(order='F'))))

timer.Stop()
print("norm(Delta) =", normDeltaMethod2)

#### Method 3 #######################################################################################################################
print(TFormat.Center(TFormat.InRed("Method 3: ")+TFormat.InBlue("using Muscat integration tool")))

from Muscat.FE.IntegrationTools import IntegrateField
timer = Timer("Duration of method 3").Start()
deltaSquare_U0 = FEField("delta_U0",meshClipped,data= deltaClippedMesh[:,0].flatten() )**2 + FEField("delta_U1",meshClipped,data= deltaClippedMesh[:,1].flatten() )**2
F = IntegrateField(deltaSquare_U0, ElementFilter(dimensionality=2))
timer.Stop()
print("norm(Delta) =", np.sqrt(F))

#### Method 4 #######################################################################################################################
print(TFormat.Center(TFormat.InRed("Method 3: ")+TFormat.InBlue("using the weak form engine")))

from Muscat.FE.Integration import IntegrateGeneral
from Muscat.FE.SymWeakForm import GetField, GetTestField
from Muscat.FE.Spaces.FESpaces import LagrangeSpaceP1, ConstantSpaceGlobal
from Muscat.FE.DofNumbering import ComputeDofNumbering

timer = Timer("Duration of method 4").Start()

U = GetField("U",2)
Tt = GetTestField("T",1)

wf = U.T*U*Tt

#meshClipped

field1 = FEField("U_0",mesh=meshClipped, data = deltaClippedMesh[:,0].flatten())
field2 = FEField("U_1",mesh=meshClipped, data = deltaClippedMesh[:,1].flatten())

numbering = ComputeDofNumbering(meshClipped,ConstantSpaceGlobal)
unknownField = FEField("T",mesh=meshClipped,space=ConstantSpaceGlobal,numbering=numbering)

elFilter = ElementFilter()

K, F = IntegrateGeneral(mesh=meshClipped,
                    wform=wf,
                    constants={},
                    fields=[field1,field2],
                    unknownFields=[unknownField],
                    integrationRuleName="LagrangeP1Quadrature",
                    elementFilter=elFilter)

normDeltaMethod3 = np.sqrt(F[0])

timer.Stop()
print("norm(Delta) =", normDeltaMethod2)
############

print(Timer.PrintTimes())
Compute the L2(Omega) norm of delta by applying numerical quadrature from Lagrange P1
Finite element integration using three different methods
******************** Method 1: 'by hand' *******************
norm(Delta) = 0.0001466583268059577
************** Method 2: using the mass matrix *************
norm(Delta) = 0.00014665832680595773
********** Method 3: using Muscat integration tool *********
norm(Delta) = 0.0001555718367085269
*********** Method 3: using the weak form engine ***********
norm(Delta) = 0.00014665832680595773

Duration of method 1:   (2) : 6.293e+00 s (mean 3.147e+00 s/call)
Duration of method 2:   (1) : 3.520e-02 s (mean 3.520e-02 s/call)
Duration of method 3:   (1) : 3.619e-01 s (mean 3.619e-01 s/call)
Duration of method 4:   (1) : 1.984e-02 s (mean 1.984e-02 s/call)