Solver Settings

Analysis > Options > Solution Settings > Solver Settings

The Solver Settings allow the user to set the Flux solvers options which will be taken by Flux (in batch) to solve the current solution.



Initialization

This option concerns the initialization of the state variables at the start of a step.

It is used for the following solutions:
  • Transient solutions
  • Static and AC solutions with Motion load and constraint
By default, the state variables are initialized with the previous step solution, except if there is any hysteretic material, the state variable will be initialized with a predicted solution. A solving option allows the user to modify the initialization to one of the following options:
  • With zero solution

    Robust for non-linear convergence

  • With the previous step solution (by default)

    Faster, but can impact the non-linear convergence

  • With a predicted solution

    Recommended for hysteretic materials.

Linear Solver Options

A linear system of n equations and n unknowns can be written under the following matrix form:

AX=B

where:
  • A is a square matrix of n x n size (with n lines and n columns)
  • X is a column vector of size n, representing the system unknowns
  • B is a column vector of size n, with known components

To find the solution, i.e. the vector X, a linear solver must be employed. Two types of solvers can be distinguished: the Direct and Iterative solvers.

The Direct solvers follow a process that leads to a very accurate solution but at the cost of several time and memory-expensive operations for large problems.

The Iterative solvers follow an iterative process that approximates a solution at each iteration. Usually, the process tends to provide a solution closer to the real solution of the linear system from iteration to the next. The process comes with a stopping criterion, depending on the accuracy threshold set for the approximated solution and the maximal number of iterations. Thus, the Iterative solver can lead to faster solutions and lower memory consumption for large problems but can be less accurate.

The linear solver is used in all EM solutions.

The automatic solver bases its decision to use the Direct or Iterative solver on the solution type and the number of unknowns, as summarized in the following Table:
Dimension 2D 3D
Solution Any Others Magneto static Integral method Steady-state AC
Number of nodes Any < 500,000 > 500,000 Any Any
Solver Direct Direct Iterative Iterative Direct
When a user wants to manually select the solver considering its computational resources, we can advise with the following guidelines:
  • For HPC (High Performance Computing) hardware or computers with a great amount of RAM, the Direct solver will be the most suitable choice thanks to its scalability and parallel potential.
  • For computers with a low amount of RAM, it would be more advisable to use the Iterative solver for its low memory consumption.
  • For 2D problems, the number of unknowns is relatively low, thus Direct solver should deliver better results in terms of solving time.

This section targets advanced users and wants to help understand what solver parameters do. Here are the solver options for Direct and Iterative solvers:



The Direct solver option available consists in using the disk for the solving in addition to the RAM. This option is very useful since it allows extending the amount of usable memory. However, using the disk implies slower operations since disks are significantly slower to compute i/o operations. If the disk memory usage is checked and no directory is selected, the default scratch directory is taken.

On the right-hand side figure, two options are displayed for the Iterative solver: the precision level (or threshold) and the maximal number of iterations. The automatic accuracy level is based on the problem type. It is possible to select a High accuracy level to ensure the result quality but it will lead to slower computation. The maximal number of iterations field allows to set a limitation on the solving time. Increasing the value can lead to more accurate results but with slower computation. Reducing the value can lead to faster computation but with reduced accuracy.