Skip to content

Solver Details

Gerasimos Chourdakis edited this page Jun 13, 2019 · 13 revisions

Since deal.II is a library and you are free to implement your own stuff, this section provides details about the implemented solver analogous to the commented tutorial programs in deal.II. In case you want to modify this solver or use it for your own project, the mathematical aspects should help you start. In case you want to know more about capability of the implemented model, have a look in the section below.

Mathematical Aspects

This program was built on the step-8 tutorial program of the deal.II library, which deals with linear elastostatics. A lot of aspects are already explained there. However, this program deals with linear elastodynamics. Therefore, we need to consider inertial effects. Our starting equation, which is basically the Navier-Cauchy equation, reads as follows:

Field equation Where u is the displacement field, rho the material density, b the body forces and sigma the stress tensor, which is related to the linear strain measure epsilon via the fourth order elasticity tensor C. Equation 1.1 needs to be satisfied in the whole domain Omega and we apply the following boundary conditions:

BC

Here, the first boundary condition is applied to the Dirichlet boundary Gamma_u and we prescribe a zero displacement. The second boundary condition is applied to the Neumann boundary Gamma_sigma and describes basically our coupling interface, since the traction vector is obtained from our flow solver. As last point, the initial condition are given in in equation 1.3:

IC

Both initial values are chosen to be zero, but you are free to choose them differently according to your problem. The material is assumed as isotropic and thus fully described by the Lamé coefficients: ECT Where 1 and I are the second and fourth order unit tensors respectively. Finally, the weak formulation of equation 1.1 is given as WKF

Discretization

Discretization in space is obviously done using Finie Elements. By default, linear shape functions are applied, but you are free to specify the polynomial degree in the parameters.prm file. However, since the boundary conditions (eq. 1.2) are assumed to be constant per cell face, linear shape functions are recommended. More details about the Finite Element discretization are available in the step-8 tutorial description (see link above). The following section focuses on the time discretization. Therefore, the governing second order differential equation is transformed, similar to a state space model, in two first order equations:

StateSpace Here, a block notation for the global vectors and matrices is used, where M denotes the mass matrix, K, the stiffness matrix, D the displacement vector, V the velocity vector and F the load vector, which includes body loads and the prescribed traction. Note, that the load vetor F is due to the coupling time dependent. Time derivatives are approximated by using a one-step theta method

Theta

where theta is a parameter [0,1], which allows to modify the time stepping properties. Theta = 0 results in a forward Euler method and theta = 1 results in a backward Euler method, which are both first order accurate. Solely theta = 0.5 results in a second order accurate Crank-Nicolson scheme, which additionally provides energy conservation in the system.

Performing some equation massaging finally leads to the following system, which is actually implemented:

Solver

Hence, we solve in each time step for the unknown velocity and update later on for the unknown displacement. Before the time loop is entered, time invariant global matrices are assembled in the assemble_system() function, namely the stiffness matrix K, the mass matrix M, and the constant body loads (gravity). Since the composition of the linear system on the left-hand side, which consists of M and K, is also constant, we store it in a stepping matrix, in order to save the rebuilding in each time step. Note: in the tutorial cases, no gravity is needed and therefore, the term is zero in the example program.

The time dependent terms are assembled in the assemble_rhs() function in each time step. This is straightforward and comments have been added in the source code for more information. Part two of equation 2.3 is finally solved in the solve() function, before the displacement vector is updated in the update_displacement() function (part one of eq. 2.3).

Capability

This program is designed for single core and single thread computations. If you like to change the source code for parallel computations, have a look at the step-17 tutorial program, which shows how this can be done by using PETSc.

Furthermore, this section should point out that the underlying physical description of structural mechanics is not suitable for large deformations and large rotations. The reason is simply the linear measurement of strains, which are only valid for small deviations. Rigid body rotations lead already to an indicated artificial strain. Due to this, the structure usually gets bigger and a small rotation assumption is violated. Hence, a linear strain measure is typically used for rotations smaller than 6°. If you like to extend the solver to a geometrical nonlinear model, have a look in the open issue on this (TODO: Add link).

Clone this wiki locally