An introduction to Finite Element Method


Krishna Kumar, kks32@cam.ac.uk
University of Cambridge




5R7 Lent 2018

Finite Element Analysis

FEA

Strong form

Strong form for a 1D bar

Strong form for a 1D bar

The forces $h$ acting at the ends of the bar are related to the normal force in the bar $N$ via $h = N n_x$, where $n_x$ is the ‘outward unit normal’ at the ends of the bar.

To satisfy equilibrium, forces acting on the bar must sum to zero.

$$0 = h(L) + h(0) + \int_{0}^{L}{f \mathrm{dx}} => N(L) - N(0) + \int_{0}^{L}{f \mathrm{dx}}$$ (since $h = Nn_x$) $$0 = \int_{0}^{L}{\frac{dN}{dx} \mathrm{dx}} + \int_{0}^{L}{f \mathrm{dx}}$$ All ‘smaller’ parts of the bar must also be in equilibrium, therefore we can remove the integrals, $-\frac{dN}{dx} = f$.

Strong form for a 1D bar

  • Equilibrium equation: $ -\frac{dN}{dx} = f $
  • Linear elasticity: $ N = A \sigma = EA \frac{du}{dx} = EA\epsilon $
  • Inserting the constitutive model into the equilibrium equation: $$ -\frac{d}{dx}\left(EA\frac{du}{dx}\right) = f$$
  • Boundary conditions:
    $u = 0$ at $x = 0$ (displacement or "Dirichlet" boundary condition)
    $EA\epsilon = h$ at $ x = L $ (force or "Neumann" boundary condition)

Converting Strong form to Weak form

The weak form is solved approximately using the finite element method

  • Multiply the strong form equation by a weight function $v$ which is equal to zero where Dirichlet (displacement) boundary conditions are applied, but is otherwise arbitrary (must be sufficiently continous)
  • Use integration by parts to `shift` derivatives to the weight fucntion
  • Insert the Neumann (force) boundary conditions

Weak form for a 1D bar

  • Multiply by weight function: $$ -\int_{0}^{L}v\frac{dN}{dx}dx = \int_{0}^{L}vf dx$$
  • Integration by parts: $$ \int_{0}^{L}\frac{dv}{dx} N dx = \int_{0}^{L}vf dx + vN|_{x=0}^{x=L}$$
  • Since $v(0) = 0$, inserting the constitutive relation and considering the force boundary at $x = L$: $$ \int_{0}^{L}\frac{dv}{dx} EA \frac{du}{dx} dx = \int_{0}^{L}vf dx + v(L)h$$

Finite Element formulation in 1D

  • FE is a Galerkin method , which approximates a PDE by replacing the unknown function (the displacement $u$ in the case of the elastic bar) and the weight function $v$ by approximate fields ($u_h$ and $v_h$). Inserting these fields: $$ \int_{0}^{L}EA \frac{dv_h}{dx} \frac{du_h}{dx} dx = \int_{0}^{L}v_h f dx + v_h(L)h$$
  • The approximate displacement field $u_h$ is defined using a set of basis functions $N_i(x)$ and values $a_i$ at a discrete number of points $x_i$ (known as nodes). For $n$ number of nodes, the approximate displacement field: $$ u_h(x) = \sum_{i}^{n} N_i (x) a_i$$
  • In FE, the basis functions $N_i$ are "Shape functions" . The approximate strain field follows: $$ \epsilon_h = \frac{du_h}{dx} = \sum_{i}^{n}\frac{dN_i(x)}{dx} a_i$$

Basis functions

Finite Element formulation in 1D

  • FE involves finding the coefficients of $a_i$, which gives approximate solution $u_h$ and $\epsilon_h$.
  • The simplest FE basis function for 1D are hat-like continous piecewise linear polynomials. Each node $i$ has its own shape function and own dof. A shape function is equal to 1 at its own node and 0 at all others

Finite Element formulation in 1D

  • Weakform: $$ \int_{0}^{L}EA \frac{dv_h}{dx} \frac{du_h}{dx} dx = \int_{0}^{L}v_h f dx + v_h(L)h$$
  • Inserting expressions for $u_h$ and $v_h$: $$ \int_{0}^{L}EA \left(\sum_i^n\frac{dN_i}{dx}{a_i^*}\right)\left(\sum_j^n\frac{dN_j}{dx}{a_j}\right) = \int_{0}^{L}\left(\sum_i^nN_ia_i^*\right) f dx + \left(\sum_i^nN_i(L)a_i^*\right)h$$
  • Since $a_i^*$ and $a_j$ are not functions of $x$: $$ \sum_i^n {a_i^*} \left(\sum_j^n {a_j} \int_{0}^{L}EA \frac{dN_i}{dx}\frac{dN_j}{dx}\right) = \sum_i^n a_i^*\int_{0}^{L}N_i f dx + \sum_i^nN_i(L)a_i^*h$$

Finite Element formulation in 1D

  • Since $a_i^*$ is arbitrary for each $i$ we set $a_{k=i}^*=1$ and $a_{k \ne i}^*=1$. Then for each $i$ we have an equation with $n$ unknowns (the values of $a_j$): $$ i = 1: \sum_j^n {a_j} \int_{0}^{L}EA \frac{dN_1}{dx}\frac{dN_j}{dx} dx = \int_{0}^{L}N_1 f dx + N_i(L)h,$$ $$ i = 2: \sum_j^n {a_j} \int_{0}^{L}EA \frac{dN_2}{dx}\frac{dN_j}{dx} dx = \int_{0}^{L}N_2 f dx + N_i(L)h,$$ $$\dots$$ $$ i = n: \sum_j^n {a_j} \int_{0}^{L}EA \frac{dN_n}{dx}\frac{dN_j}{dx} dx = \int_{0}^{L}N_n f dx + N_i(L)h,$$
  • A system of linear equationos is most conveniently expressed as a matrix: $$ \mathbf{K a} = \mathbf{b} $$
  • Stiffness matrix: $K_{ij} = \int_{0}^{L}EA \frac{dN_i}{dx}\frac{dN_j}{dx} dx$
    right-hand side vector: $b_i = \int_{0}^{L}N_i f dx + N_i(L)h$.

1D Shape functions

  • Linear shape function between 2 nodes. Shape function is equal to 1 at its own node, and 0 at other nodes of the element:
  • The displacement field: $u_h(x) = N_1(x) a_1 + N_2 (x) a_2 = (-\frac{x}{l} +1) a_1 + \frac{x}{l}a_2$
  • The strain field: $\epsilon_h(x) = \frac{dN_1(x)}{dx} a_1 + \frac{dN_2 (x)}{dx} a_2 = -\frac{1}{l} a_1 + \frac{1}{l}a_2$
  • In matrix format: $ u_h = \mathbf{Na_e} $ and $\epsilon_h = \mathbf{Ba_e}$.
  • Shape function: $\mathbf{N} =\left[N_1(x) \quad N_2(x)\right]$ and derivatives: $\mathbf{B} = \left[ \frac{dN_1(x)}{dx} \quad \frac{dN_2(x)}{dx}\right]$

FE formulation: 1D element

  • Stiffness matrix for 1D element: $K_{ij} = \int_{0}^{L} \frac{dN_i}{dx} EA \frac{dN_j}{dx} dx$
  • FE matrix notation: $$K_{e} = \int_{0}^{L} \mathbf{B}^T EA \mathbf{B} dx$$
    $$K_{e} = \int_0^L \left[ \begin{matrix} \frac{dN_1}{dx} \\ \frac{dN_2}{dx} \end{matrix} \right] EA \left[ \begin{matrix} \frac{dN_1}{dx} & \frac{dN_2}{dx} \end{matrix} \right] dx$$
    $$K_{e} = \frac{EA}{l} \left[\begin{matrix} 1 & -1 \\ -1 & 1 \end{matrix} \right]$$

Finite Element Analysis of a Cantilever beam

Sleipner A offshore platform sprang leak

sank on 23 August 1991 during the controlled sinking operation in the open sea.

Using the finite element program NASTRAN, the shear stresses in the tricells were under- estimated by 47%. First, the chosen finite element mesh was exceedingly coarse so that the finite element shear stress was significantly too small.

Sleipner A offshore platform sprang leak

Second, the shear stresses at the boundary have been quadratically extrapolated using the shear forces at points A, B and C. We know however from beam theory that the shear force distribution is linear so that the shear force at the boundary is underestimated by ≈ 40%

Sleipner A offshore platform sprang leak

To make matters worse, the necessary reinforcement was automatically dimensioned based on the finite element results without any checking by an engineer.