FEM example in Python

M. M. Sussman
[email protected]
Office Hours: 11:10AM-12:10PM, Thack 622

May 12 June 19, 2014

1 / 45
Topics

Introduction

Code

Verify and run

2 / 45
Purpose

I Practice with Python

I Illustrate FEM in 1D in detail
I Coding strategies

3 / 45
Problem description

I ODE u 00 + 2u 0 + u = f = (x + 2)
I Neumann boundary conditions
I Why? Because all 3 terms, real solution with exponentials
I Exact u = (1 + x)e1x + x(1 ex )
I Done when get correct convergence rate to exact solution
I Other b.c., rhs are available

4 / 45
How to debug and test?

I Never write code without a test plan!

I Test as you go
I Choose exact solutions and test terms one at a time
I Have a test function as part of the code.
I When code is working
I Find problems similar to given, but with exact solutions
I Verify reasonable solution

5 / 45
Strategy

I Leave b.c. and r.h.s as general as possible

I quadratic elements
I Makes nontrivial exact solutions possible
I Model classes on FEniCS classes

6 / 45
Class structure

I Mesh
I Shapefns
I FiniteElement
I Including integration over this element
I FunctionSpace
I Including assembly of integrals

7 / 45
Pseudo code

mesh=Mesh( ... )
sfns=Shapefns( ... )
V=FunctionSpace(mesh,sfns)
R
b= f
A= 0i 0j + 2 0i j + i j
R R R

Modify A and b for boundary conditions

u=la.solve(A,b)

8 / 45
Topics

Introduction

Code

Verify and run

9 / 45
Coding strategy for classes

I Classes contain hidden data only

I Information accessed by accessor functions only

10 / 45
Details of classes Mesh and Shapefns

I Mesh
I Mesh(N,a,b)
N is number of elements
a is left endpoint
b is right endpoint
I coordinates() or coordinates(n)
returns all coordinates or nth coordinate
I cells() or cells(n)
returns all nodes numbers of nth cell or all node numbers
I size()
returns number of mesh cells
I Shapefns
I Shapefns()
I eval(n,xi) returns n ()
I ddx(n,xi) returns 0n ()
I size() returns the number of required nodes

11 / 45
Shapefns

class Shapefns(object):
"""
Define Quadratic Lagrange shape functions
These will be defined on the (local) interval [0,1], with
mid point 0.5
Shapefns()
eval(n,xi): phi[n](xi)
ddx(n,xi): dphi[n](xi)
size(): number of nodes for these shape functions
"""

12 / 45
Shapefns

class Shapefns(object):
"""
Define Quadratic Lagrange shape functions
These will be defined on the (local) interval [0,1], with
mid point 0.5
Shapefns()
eval(n,xi): phi[n](xi)
ddx(n,xi): dphi[n](xi)
size(): number of nodes for these shape functions
"""

def __init__(self):

12 / 45
Shapefns

class Shapefns(object):
"""
Define Quadratic Lagrange shape functions
These will be defined on the (local) interval [0,1], with
mid point 0.5
Shapefns()
eval(n,xi): phi[n](xi)
ddx(n,xi): dphi[n](xi)
size(): number of nodes for these shape functions
"""

def __init__(self):
"""
an array of functions for phi and deriv phi
"""
self.__phi=[lambda xi: 2.0 * (xi-0.5) * (xi-1.0) ,\
lambda xi: 4.0 * xi * (1.0-xi), \
lambda xi: 2.0 * xi * (xi-0.5)]

12 / 45
Shapefns

class Shapefns(object):
"""
Define Quadratic Lagrange shape functions
These will be defined on the (local) interval [0,1], with
mid point 0.5
Shapefns()
eval(n,xi): phi[n](xi)
ddx(n,xi): dphi[n](xi)
size(): number of nodes for these shape functions
"""

def __init__(self):
"""
an array of functions for phi and deriv phi
"""
self.__phi=[lambda xi: 2.0 * (xi-0.5) * (xi-1.0) ,\
lambda xi: 4.0 * xi * (1.0-xi), \
lambda xi: 2.0 * xi * (xi-0.5)]
# and dphi (derivative of phi w.r.t. xi)
# derivative of second factor * first + derivative of first factor * sec
self.__dphi=[lambda xi: 2.0 * (xi-0.5) + 2.0*(xi - 1.0) ,\
lambda xi: -4.0 * xi + 4.0*(1.0 - xi), \
lambda xi: 2.0 * xi + 2.0*(xi - 0.5)]

12 / 45
Shapefns

class Shapefns(object):
"""
Define Quadratic Lagrange shape functions
These will be defined on the (local) interval [0,1], with
mid point 0.5
Shapefns()
eval(n,xi): phi[n](xi)
ddx(n,xi): dphi[n](xi)
size(): number of nodes for these shape functions
"""

def __init__(self):
"""
an array of functions for phi and deriv phi
"""
self.__phi=[lambda xi: 2.0 * (xi-0.5) * (xi-1.0) ,\
lambda xi: 4.0 * xi * (1.0-xi), \
lambda xi: 2.0 * xi * (xi-0.5)]
# and dphi (derivative of phi w.r.t. xi)
# derivative of second factor * first + derivative of first factor * sec
self.__dphi=[lambda xi: 2.0 * (xi-0.5) + 2.0*(xi - 1.0) ,\
lambda xi: -4.0 * xi + 4.0*(1.0 - xi), \
lambda xi: 2.0 * xi + 2.0*(xi - 0.5)]
self.__N=3 #number of nodes in quadratic Lagrange polynomial

12 / 45
Shapefns, contd

def eval(self,n,xi):
"""
the function phi[n](xi), for any xi
"""
return self.__phi[n](xi)

def ddx(self,n,xi):
"""
the function dphi[n](xi), for any xi
"""
return self.__dphi[n](xi)

def size(self):
"""
the number of points
"""
return self.__N

13 / 45
Shapefns, contd

def eval(self,n,xi):
"""
the function phi[n](xi), for any xi
"""
return self.__phi[n](xi)

def ddx(self,n,xi):
"""
the function dphi[n](xi), for any xi
"""
return self.__dphi[n](xi)

def size(self):
"""
the number of points
"""
return self.__N

13 / 45
What about the Mesh Class?

Homework, Exercise 11, 10 points

14 / 45
Test it!

I Place code at the end of the file to test it!

I Can preceed it with if __name__ ==
__main__:
I Test early, test often, dont throw away the
testing code!

15 / 45
Test code
N=5
rightpt = 5.0
print "\n\nFEM1D.py Test case, dx=",rightpt/N

mesh = Mesh(N,0.0,rightpt)
coords = mesh.coordinates()
print "mesh.coordinates()=",coords

sfns = Shapefns()
print "sfns.size()-3=", sfns.size()-3
xi = np.linspace(0,1,100)
import matplotlib.pyplot as plt
if True:
for n in range(3):
plt.plot(xi,sfns.eval(n,xi))
plt.show()
plt.plot(xi,sfns.ddx(n,xi))
plt.show()

16 / 45
Shape function plots

17 / 45
FiniteElement class
I Describes a single finite element
I A FunctionSpace will contain many of these
I Constructed from
1. Mesh
2. shape functions
3. this elements number (agrees with one of the Mesh intervals
4. 3 numbers to use for DOFs.
I Methods:
endpts(): element end points
dofpts(): array of DOF locations
dofnos(): DOF numbers
numDofs(): length of dofnos()
eval(n,x): evaluate n (x) (not )
ddxl(n,x): evaluate 0n (x) (not ) R
integral(f1,f2,derivative): e f1 (x)f2 (x)n (x) dx
I f1 and f2 are optional
I put 0 in integral if derivative=True
I returns a vector, one value for each n

18 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))
self.__numDofs = sfns.size()

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))
self.__numDofs = sfns.size()
assert(sfns.size() == len(dofnos))
self.__dofnos=dofnos

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))
self.__numDofs = sfns.size()
assert(sfns.size() == len(dofnos))
self.__dofnos=dofnos
self.__dofpts=np.linspace(self.__endpts[0],self.__endpts[1],self.__numDofs)

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))
self.__numDofs = sfns.size()
assert(sfns.size() == len(dofnos))
self.__dofnos=dofnos
self.__dofpts=np.linspace(self.__endpts[0],self.__endpts[1],self.__numDofs)
self.__sfns=sfns

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))
self.__numDofs = sfns.size()
assert(sfns.size() == len(dofnos))
self.__dofnos=dofnos
self.__dofpts=np.linspace(self.__endpts[0],self.__endpts[1],self.__numDofs)
self.__sfns=sfns
# Gauss points and weights: 3-pts are high enough for this
self.__gausspts = np.array(\
(.112701665379258311482073460022,.5,.887298334620741688517926539978))
self.__gausswts = np.array((5.0/18.0,8.0/18.0,5.0/18.0))

19 / 45
FiniteElement constructor
def __init__(self,mesh,sfns,eltno,dofnos):
"""
mesh is the mesh it is built on
sfns is the Shapefuns member
eltno is this elements number
endnos is a pair of ints giving the numbers of the endpoints
in the mesh
dofnos is an array of ints giving the numbers of the dofs
"""
# this element no. is same as mesh element no.
assert(0 <= eltno < mesh.size())
self.__eltno = eltno
endnos = mesh.cells(eltno)
assert(len(endnos) == 2)
self.__endpts = np.array(mesh.coordinates(endnos))
self.__numDofs = sfns.size()
assert(sfns.size() == len(dofnos))
self.__dofnos=dofnos
self.__dofpts=np.linspace(self.__endpts[0],self.__endpts[1],self.__numDofs)
self.__sfns=sfns
# Gauss points and weights: 3-pts are high enough for this
self.__gausspts = np.array(\
(.112701665379258311482073460022,.5,.887298334620741688517926539978))
self.__gausswts = np.array((5.0/18.0,8.0/18.0,5.0/18.0))
# for efficiency, generate an array of shape functions evaluated
# at the Gauss points
self.__gaussvals = np.empty([self.__numDofs,self.__gausspts.size])
for n in range(self.__numDofs):
self.__gaussvals[n,:]=sfns.eval(n,self.__gausspts[:])
19 / 45
FiniteElement accessor methods

def endpts(self):
""" access endpoints """
return self.__endpts

def dofpts(self):
""" access dofpoints """
return self.__dofpts

def dofnos(self):
""" access dof point numbers """
return self.__dofnos

def numDofs(self):
""" access numDofs """
return self.__numDofs

20 / 45
FiniteElement evaluation methods

def eval(self,n,x):
"""
evaluate the n-th shape function on this element
at the spatial coordinate x
"""
# map x to xi
xx=np.array(x)
xi=(xx-self.__endpts[0])/(self.__endpts[1]-self.__endpts[0])
# evaluate
return self.__sfns.eval(n,xi)*(xi >= 0.)*(xi <= 1.)

def ddx(self,n,x):
"""
evaluate the n-th shape function on this element
at the spatial coordinate x
"""
# map x to xi
xi=(x-self.__endpts[0])/(self.__endpts[1]-self.__endpts[0])
# evaluate
return self.__sfns.ddx(n,xi)*(xi>=0.)*(xi <= 1.0)

21 / 45
Gau integration on the reference element

P
I Q= i f (i )wi
I Integration points (Gau points) tabulated
I Weights wi tabulated
I Integral on the true element requires affine transformation,
x = L + x0
I Assume functions f are given at DOF points.

22 / 45
FiniteElement integration
def integral(self, f1=None, f2=None, derivative=False):
"""
Integrate either phi[i](xi)*f1(xi)*f2(xi) or dphi[i]*f1*f2
over this element, depending on if derivative is False or True
Returns a vector of 3 results, one for
phi[0], one for phi[1], and one for phi[2].
f1 and f2 are assumed to have been mapped to this element
as arrays
if derivative is True, phi is replaced with dphi
"""
L = self.__endpts[1]-self.__endpts[0] # length of element
t = self.__gausswts.copy()
gp = self.__gausspts

23 / 45
FiniteElement integration
def integral(self, f1=None, f2=None, derivative=False):
"""
Integrate either phi[i](xi)*f1(xi)*f2(xi) or dphi[i]*f1*f2
over this element, depending on if derivative is False or True
Returns a vector of 3 results, one for
phi[0], one for phi[1], and one for phi[2].
f1 and f2 are assumed to have been mapped to this element
as arrays
if derivative is True, phi is replaced with dphi
"""
L = self.__endpts[1]-self.__endpts[0] # length of element
t = self.__gausswts.copy()
gp = self.__gausspts

23 / 45
FiniteElement integration
def integral(self, f1=None, f2=None, derivative=False):
"""
Integrate either phi[i](xi)*f1(xi)*f2(xi) or dphi[i]*f1*f2
over this element, depending on if derivative is False or True
Returns a vector of 3 results, one for
phi[0], one for phi[1], and one for phi[2].
f1 and f2 are assumed to have been mapped to this element
as arrays
if derivative is True, phi is replaced with dphi
"""
L = self.__endpts[1]-self.__endpts[0] # length of element
t = self.__gausswts.copy()
gp = self.__gausspts

if f1 != None:
assert(len(f1) == self.__numDofs)
fvals = np.zeros([self.__gausspts.size])
for n in range(self.__numDofs):
fvals += f1[n]*self.__gaussvals[n,:]
t*=fvals

23 / 45
FiniteElement integration
def integral(self, f1=None, f2=None, derivative=False):
"""
Integrate either phi[i](xi)*f1(xi)*f2(xi) or dphi[i]*f1*f2
over this element, depending on if derivative is False or True
Returns a vector of 3 results, one for
phi[0], one for phi[1], and one for phi[2].
f1 and f2 are assumed to have been mapped to this element
as arrays
if derivative is True, phi is replaced with dphi
"""
L = self.__endpts[1]-self.__endpts[0] # length of element
t = self.__gausswts.copy()
gp = self.__gausspts

if f1 != None:
assert(len(f1) == self.__numDofs)
fvals = np.zeros([self.__gausspts.size])
for n in range(self.__numDofs):
fvals += f1[n]*self.__gaussvals[n,:]
t*=fvals

if f2 != None:
assert(len(f2) == self.__numDofs)
fvals = np.zeros([self.__gausspts.size])
for n in range(self.__numDofs):
fvals += f2[n]*self.__gaussvals[n,:]
t *= fvals

23 / 45
FiniteElement integration contd

if derivative:
# really: t *= L*(1/L)
q = np.dot(np.array([self.__sfns.ddx(0,gp), \
self.__sfns.ddx(1,gp), \
self.__sfns.ddx(2,gp)]),t)
else:
t *= L # correct for affine map x->xi
q = np.dot(np.array([self.__sfns.eval(0,gp), \
self.__sfns.eval(1,gp), \
self.__sfns.eval(2,gp)]),t)

return q

24 / 45
Test it!

Test early! Test often! Dont discard your test code!

Compare vs. M ATLAB symbolic toolbox

elt = FiniteElement(mesh,sfns,0,[0,1,2])

if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:

# test some integrals
print "elt integral() err=",max(abs(elt.integral()-[1./6,2./3,1./6]))
print "integral(deriv) err=",\
max(abs(elt.integral(derivative=True)-[-1,0,1]))

25 / 45
Test it!

Test early! Test often! Dont discard your test code!

Compare vs. M ATLAB symbolic toolbox

elt = FiniteElement(mesh,sfns,0,[0,1,2])

if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:

# test some integrals
print "elt integral() err=",max(abs(elt.integral()-[1./6,2./3,1./6]))
print "integral(deriv) err=",\
max(abs(elt.integral(derivative=True)-[-1,0,1]))
# pick the function f(x)=x, find its expansion coefs
ex = np.empty([sfns.size()])
ex[0] = elt.endpts()[0]
ex[2] = elt.endpts()[1]
ex[1] = .5*(ex[0]+ex[2])
ex2 = ex**2
print "integral(x) err=",max(abs(elt.integral(ex)-[0,1./3,1./6]))

25 / 45
Test it!

Test early! Test often! Dont discard your test code!

Compare vs. M ATLAB symbolic toolbox

elt = FiniteElement(mesh,sfns,0,[0,1,2])

if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:

# test some integrals
print "elt integral() err=",max(abs(elt.integral()-[1./6,2./3,1./6]))
print "integral(deriv) err=",\
max(abs(elt.integral(derivative=True)-[-1,0,1]))
# pick the function f(x)=x, find its expansion coefs
ex = np.empty([sfns.size()])
ex[0] = elt.endpts()[0]
ex[2] = elt.endpts()[1]
ex[1] = .5*(ex[0]+ex[2])
ex2 = ex**2
print "integral(x) err=",max(abs(elt.integral(ex)-[0,1./3,1./6]))
print "integral(x**2) err=",max(abs(elt.integral(ex2)-[-1./60,1./5,3./20]))
print "integral(x**2) err=",max(abs(elt.integral(ex,ex)-[-1./60,1./5,3./20])

25 / 45
Test it!

Test early! Test often! Dont discard your test code!

Compare vs. M ATLAB symbolic toolbox

elt = FiniteElement(mesh,sfns,0,[0,1,2])

if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:

# test some integrals
print "elt integral() err=",max(abs(elt.integral()-[1./6,2./3,1./6]))
print "integral(deriv) err=",\
max(abs(elt.integral(derivative=True)-[-1,0,1]))
# pick the function f(x)=x, find its expansion coefs
ex = np.empty([sfns.size()])
ex[0] = elt.endpts()[0]
ex[2] = elt.endpts()[1]
ex[1] = .5*(ex[0]+ex[2])
ex2 = ex**2
print "integral(x) err=",max(abs(elt.integral(ex)-[0,1./3,1./6]))
print "integral(x**2) err=",max(abs(elt.integral(ex2)-[-1./60,1./5,3./20]))
print "integral(x**2) err=",max(abs(elt.integral(ex,ex)-[-1./60,1./5,3./20])
print "integral(x,phi) err=",\
max(abs(elt.integral(ex,derivative=True)-[-1./6,-2./3,5./6]))
print "integral(x**2,phi) err=",\
max(abs(elt.integral(ex2,derivative=True)-[0,-2./3,2./3]))

25 / 45
Test results

elt integral() err= 1.11022302463e-16

integral(deriv) err= 0.0
integral(x) err= 5.55111512313e-17
integral(x**2) err= 2.77555756156e-17
integral(x**2) err= 2.77555756156e-17
integral(x,phi) err= 1.11022302463e-16
integral(x**2,phi) err= 2.22044604925e-16

26 / 45
FunctionSpace class

I A FunctionSpace has a list of elements

I Elements are numbered as in mesh
I Constructed from mesh and Shapefns
I Methods:
size(): number of elements
ndofs(): number of all dofs
dofpts(): coordinates of dof[n] orRall dofs
int_phi_phi(c,derivative): c(x)i (x)j (x) dx
I c is optional
I derivative is a pair of booleans determining whether to use or 0
I returns 2D array of values
R
int_phi(f,derivative): f (x)i (x) dx
I f is optional
I derivative is a boolean determining or 0
I returns vector of values

27 / 45
FunctionSpace constructor

def __init__(self,mesh,sfns):
"""
mesh is the mesh
sfns is the Shapefuns
"""

28 / 45
FunctionSpace constructor

def __init__(self,mesh,sfns):
"""
mesh is the mesh
sfns is the Shapefuns
"""
self.__size=mesh.size()
# number the elements in same way as mesh
self.__elts = list([])
self.__dofpts = list([])
self.__nDOFs=0

28 / 45
FunctionSpace constructor

def __init__(self,mesh,sfns):
"""
mesh is the mesh
sfns is the Shapefuns
"""
self.__size=mesh.size()
# number the elements in same way as mesh
self.__elts = list([])
self.__dofpts = list([])
self.__nDOFs=0
for n in range(self.__size):
# ASSUMING only boundary points are number 0 and (self.__size)
if n == 0:
self.__nDOFs += 3
dofs = [2*n, 2*n+1, 2*n+2]
newdofs = range(3)
else:
self.__nDOFs += 2
dofs=[2*n, 2*n+1, 2*n+2]
newdofs = range(1,3)
fe = FiniteElement(mesh,sfns,n,dofs)
self.__elts.append(fe)
for i in newdofs:
self.__dofpts.append(fe.dofpts()[i])

28 / 45
FunctionSpace constructor

def __init__(self,mesh,sfns):
"""
mesh is the mesh
sfns is the Shapefuns
"""
self.__size=mesh.size()
# number the elements in same way as mesh
self.__elts = list([])
self.__dofpts = list([])
self.__nDOFs=0
for n in range(self.__size):
# ASSUMING only boundary points are number 0 and (self.__size)
if n == 0:
self.__nDOFs += 3
dofs = [2*n, 2*n+1, 2*n+2]
newdofs = range(3)
else:
self.__nDOFs += 2
dofs=[2*n, 2*n+1, 2*n+2]
newdofs = range(1,3)
fe = FiniteElement(mesh,sfns,n,dofs)
self.__elts.append(fe)
for i in newdofs:
self.__dofpts.append(fe.dofpts()[i])

28 / 45
FunctionSpace constructor

def __init__(self,mesh,sfns):
"""
mesh is the mesh
sfns is the Shapefuns
"""
self.__size=mesh.size()
# number the elements in same way as mesh
self.__elts = list([])
self.__dofpts = list([])
self.__nDOFs=0
for n in range(self.__size):
# ASSUMING only boundary points are number 0 and (self.__size)
if n == 0:
self.__nDOFs += 3
dofs = [2*n, 2*n+1, 2*n+2]
newdofs = range(3)
else:
self.__nDOFs += 2
dofs=[2*n, 2*n+1, 2*n+2]
newdofs = range(1,3)
fe = FiniteElement(mesh,sfns,n,dofs)
self.__elts.append(fe)
for i in newdofs:
self.__dofpts.append(fe.dofpts()[i])
self.__dofpts = np.array(self.__dofpts)

28 / 45
FunctionSpace int_phi_phi

I derivative is now a pair of booleans

I A is a large matrix nDOFs X nDOFs
I Elements have to be assembled into A in the right places
I Trick: treat first like f when call elemental integration

29 / 45
int_phi_phi code

def int_phi_phi(self, c=None, derivative=[False,False]):

"""
assemble $\int c(x)\phi(x)\phi(x) dx$ or with $d\phi/dx$
"""
A = np.zeros([self.__nDOFs,self.__nDOFs])
# loop over elements
for elt in self.__elts:
d0=elt.dofnos()

N = elt.numDofs()

for j in range(N):

phi = elt.eval(j,elt.dofpts())
A[d0,d0[j]] += elt.integral(phi,cc,derivative=derivative[0])
return A

30 / 45
int_phi_phi code

def int_phi_phi(self, c=None, derivative=[False,False]):

"""
assemble $\int c(x)\phi(x)\phi(x) dx$ or with $d\phi/dx$
"""
A = np.zeros([self.__nDOFs,self.__nDOFs])
# loop over elements
for elt in self.__elts:
d0=elt.dofnos()
if c != None:
cc = c[d0]
else:
cc = None
N = elt.numDofs()

for j in range(N):

phi = elt.eval(j,elt.dofpts())
A[d0,d0[j]] += elt.integral(phi,cc,derivative=derivative[0])
return A

30 / 45
int_phi_phi code

def int_phi_phi(self, c=None, derivative=[False,False]):

"""
assemble $\int c(x)\phi(x)\phi(x) dx$ or with $d\phi/dx$
"""
A = np.zeros([self.__nDOFs,self.__nDOFs])
# loop over elements
for elt in self.__elts:
d0=elt.dofnos()
if c != None:
cc = c[d0]
else:
cc = None
N = elt.numDofs()
endpts=elt.endpts()
L = endpts[1]-endpts[0] # length of elt
for j in range(N):
if derivative[1]:
# chain rule: d(xi)/d(x)=1/L
phi = elt.ddx(j,elt.dofpts())/L
else:
phi = elt.eval(j,elt.dofpts())
A[d0,d0[j]] += elt.integral(phi,cc,derivative=derivative[0])
return A

30 / 45
Exercise 12 (5 points)

Complete the code for FunctionSpace.int_phi Be sure to test

your code!

31 / 45
Test it!
Test early! Test often! Dont discard your test code!
V=FunctionSpace(mesh,sfns)
print "V.Ndofs()-correct=",V.Ndofs()-(2*N+1)
print "V.size()-correct=",V.size()-N

x = V.dofpts()
f = x.copy()
print "error in integral x over [",x[0],",",x[-1],"]=",\
np.sum(V.int_phi(f))-x[-1]**2/2.
f = 0.0*x+1
print "error in integral 1 over [",x[0],",",x[-1],"]=",\
np.sum(V.int_phi(f))-x[-1]
f = x.copy()**2
print "error in integral x**2 over [",x[0],",",x[-1],"]=",\
np.sum(V.int_phi(f))-x[-1]**3/3.
f = x.copy()**3
print "error in integral x**3 over [",x[0],",",x[-1],"]=",\
np.sum(V.int_phi(f))-x[-1]**4/4.
f = x.copy()**4
print "error in integral x**4 over [",x[0],",",x[-1],"]=",\
np.sum(V.int_phi(f))-x[-1]**5/5.," should be nonzero."

print "norm(V.dofpts()-correct)=",\
la.norm(V.dofpts()-np.linspace(0,coords[-1],2*N+1))

P
Note that i i (x) = 1.
32 / 45
Test results

V.Ndofs()-correct= 0
V.size()-correct= 0
error in integral x over [ 0.0 , 5.0 ]= 0.0
error in integral 1 over [ 0.0 , 5.0 ]= 8.881784197e-16
error in integral x**2 over [ 0.0 , 5.0 ]= 0.0
error in integral x**3 over [ 0.0 , 5.0 ]= 2.84217094304e-14
error in integral x**4 over [ 0.0 , 5.0 ]= 0.0416666666667 should be nonzero.
norm(V.dofpts()-correct)= 0.0

33 / 45
Test FunctionSpace matrix integration
Z
Aij = i j

A=V.int_phi_phi()
if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:
print "error A00=",A[0,0]-2./15.
print "error A01=",A[0,1]-1./15.
print "error A02=",A[0,2]+1./30.
print "error A11=",A[1,1]-8./15.
print "error A12=",A[1,2]-1./15.
print
print "error A22=",A[2,2]-4./15.
print "error A23=",A[2,3]-1./15.
print "error A24=",A[2,4]+1./30.
print "error A33=",A[3,3]-8./15.
print "error A34=",A[3,4]-1./15.
print

# trivial check with coefficient

c = np.ones([Ndofs])
A1 = V.int_phi_phi(c)
print "Norm difference matrices=",la.norm(A-A1)

34 / 45
More test FunctionSpace matrix integration

Z
Bij = ci j

c = (1.0+x)
B = V.int_phi_phi(c)
if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:
print "error B00=",B[0,0]-3./20.
print "error B01=",B[0,1]-1./15.
print "error B02=",B[0,2]+1./20.
print "error B11=",B[1,1]-12./15.
print "error B12=",B[1,2]-2./15.
print
print "error B22=",B[2,2]-8./15.
print "error B23=",B[2,3]-2./15.
print "error B24=",B[2,4]+1./12.
print "error B33=",B[3,3]-4./3.
print "error B34=",B[3,4]-3./15.

35 / 45
Test FunctionSpace Laplace matrix integration

Z
Cij = 0i 0j

C = V.int_phi_phi(derivative=[True,True])
if N == 5 and np.abs(coords[-1]-5.0) < 1.e-10:
print "\n Laplace Matrix"
print "error C00*3=",C[0,0]-7./3.
print "error C01*3=",C[0,1]+8./3.
print "error C02*3=",C[0,2]-1./3.
print "error C11*3=",C[1,1]-16./3.
print "error C12*3=",C[1,2]+8./3.
print
print "error C22*3=",C[2,2]-14./3.
print "error C23*3=",C[2,3]+8./3.
print "error C24*3=",C[2,4]-1./3.
print "error C33*3=",C[3,3]-16./3.
print "error C34*3=",C[3,4]+8./3.
print

36 / 45
Test FunctionSpace Laplace matrix multiply

Z
Cij = 0i 0j

soln2 = np.ones([Ndofs])
b2 = np.dot(C,soln2)
print "const soln Laplace, norm check=",la.norm(b2)

soln = x
b0 = np.dot(C,soln)
rhs0 = V.int_phi(np.zeros([Ndofs]))
# natural b.c. not satisfied, dont check them
rhs0[0] = -b0[0]
rhs0[-1] = -b0[-1]
print "soln=x Laplace, norm check=",la.norm(rhs0+b0)

soln = x**2
b1 = np.dot(C,soln)
rhs1 = V.int_phi(2.0*np.ones([Ndofs]))
# natural b.c. not satisfied on right, dont check it
rhs1[-1] = -b1[-1]
print "soln=x**2 Laplace, norm check=",la.norm(rhs1+b1)

37 / 45
Test FunctionSpace derivative matrix multiply

Z
Dij = i 0j

D = V.int_phi_phi(derivative=[False,True])
soln = np.ones([V.Ndofs()])
b2 = np.dot(D,soln)
print "norm check (rhs d/dx+Neumann, const soln)=",la.norm(b2)

D[0,0] = 1.0
D[0,1:] = 0.0
D[-1,-1] = 1.0
D[-1,0:-1] = 0.0
soln = x
b3 = np.dot(D,soln)
rhs3 = V.int_phi(np.ones([Ndofs]))
rhs3[0] = soln[0]
rhs3[-1] = soln[-1]
print "norm check (d/dx+Dirichlet soln=x)=",la.norm(rhs3-b3)

38 / 45
Topics

Introduction

Code

Verify and run

39 / 45
Verification

I Solve an easy problem like yours!

I Verification
I Is your mesh fine enough?
I Is the program capable of solving your problem?
I Exact solutions are excellent choices

u 00 + 2u 0 + u = f x [0, 1]

1. u = 1, f = 1, Dirichlet b.c.
2. u = x, f = 1 + x, Dirichlet b.c.
3. u = x 2 , f = 2 + 2x + x 2 , Dirichlet b.c.

40 / 45
Verification function

def verification(title,N,rhsfn,exactfn):
"""
generic verification runs for PDE u+2*u+u=rhs on [0,1]
Dirichlet b.c.
title= descriptive title
N=number of elements
rhsfn=function for rhs as function of x
exactfn=function for exact solution as function of x
MMS 4/16/14
"""
# N elements in [0,1]
mesh = Mesh(N,0.0,1.0)

# shape functions, function space

sfns = Shapefns()
V = FunctionSpace(mesh,sfns)

# rhs and exact

x = V.dofpts()
exact = exactfn(x)

rhs = rhsfn(x)
b = V.int_phi(rhs)

41 / 45
Verification function contd

# assemble stiffness matrix: u+2*u+u

# integration by parts on first term introduces minus sign
A= -V.int_phi_phi(derivative=[True,True]) \
+2*V.int_phi_phi(derivative=[False,True]) \
+V.int_phi_phi()

# insert boundary conditions

# left bndry u=1 (from exact solution)
A[0,0] = 1.0
A[0,1:] = 0.0
b[0] = exact[0]
# right bndry u=1 (from exact solution)
A[-1,-1] = 1.0
A[-1,0:-1] = 0.0
b[-1] = exact[-1]

# solve
u = la.solve(A,b)

# check
print title," relative error=",la.norm(u-exact)/la.norm(exact)

42 / 45
Run verification

verification("Case 1", 5, rhsfn=lambda(x):0.0*x+1.0, \

exactfn=lambda(x):0.0*x+1.0)
verification("Case 2", 5, rhsfn=lambda(x):x+2.0, \
exactfn=lambda(x):x)
verification("Case 3", 5, rhsfn=lambda(x):x**2+4*x+2.0, \
exactfn=lambda(x):x**2)

43 / 45
Solve the given problem

Homework, Exercise 13, 10 points

44 / 45
Convergence estimate

N Relative Error Error Ratio

5 5.09313786541e-06 15.85
10 3.21248851405e-07 15.91
20 2.01935531431e-08 15.94
40 1.26705299734e-09 15.86
80 7.98849978221e-11 7.18
160 1.11333740702e-11

I Appears to be O(h4 )
I Only expect O(h3 )
I Superconvergent at the nodes because mesh is uniform?

45 / 45