Elmer Release Notes for version 8.3
===================================
Previous release: 8.2
Period covered: 12 March 2016 - 18 May 2017
Number of commits: ~620
ElmerSolver library functionality
--------------------------------
This is functionality that can be controlled by the user by choose of keywords.
- IDR(s): New linear solver "IDR(s)" implemented. This is a contribution
from Martin van Gijzen. It is called by setting > Linear System Iterative Method = "idrs".
The s-parameter is set by "Idrs Parameter = 4", for example. The method is quite robust and
may often provide the best available strategy for nonsymmetric matrices.
The convergence may be made monotonic by setting > Idrs Smoothing = True <
- Hcurl conforming elements
o Second-order pyramidic (31-DOF) edge element added: curl-conforming
approximation of degree 2 is now possible with all element shapes
o General Dirichlet BCs can now be specified for all available curl-conforming
(edge) elements (second order, optimal family, and second kind basis options
supported); see Appendix F of Elmer Solver Manual
o Mortar BCs of Galerkin type for second-order edge elements over the lowest-
order background mesh
- Linear solver control: These are routines that should make the use of linear solvers more robust.
o Upper limit for detecting divergence in the internal Krylov methods of Elmer.
Activated by keyword > Linear System Divergence Limit <, default is 1e20.
o Changing linear system on-the-fly: There is a mapping between nonlinear (or coupled system)
iteration and iteration strategy. Different iteration levels may have different linear solvers.
See test case LinearSolverNamespace for example.
o The linear system may be solved for the change of solution only i.e. system Ax=b is
replaced by system Adx=b-Ax0. This is activated by setting > Linear System Residual Mode = True <.
o Added strategies that ensure that even in case of divergence the almost best solution so far is
used. This is applicable to BiCGStabl mainly. Set > Linear System Robust = True <.
- Scanning Loops: each solver can have "Scanning Loops" that call the same solver repeatedly.
This can be used to scan over parameter space, for example frequency. Often the same iteration
could be achieved with steady state or time level iterations, but this gives additional flexibility.
The scanning loops introduce an additional variable "scan" that may be used similarly to "time"
in parametric dependencies. Look for test cases *Scan for examples.
- SlaveSolvers: each solve can have a number of slave solvers i.e. other solvers that are called
within the primary solver. They may be called at various slots associated with specific
DefaultRoutines in the solver. Currently the possible slave solver slots are:
Pre Solvers(n) = Integer
Nonlinear Pre Solvers(n) = Integer
Nonlinear Post Solvers(n) = Integer
Post Solvers(n) = Integer
Look for test cases SlaveSolverSlots* for example.
- Allow for computation of ConstraintModes i.e. modes resulting from unit nodal load.
These can be used as a basis in model reduction, possibly alongside with eigenmodes.
- Mortar FEs: the mortar projector may involve coordinate transformation from 3D to 2D
assuming either cylindrical/rotational, axial or radial interfaces.
- Sif parser: Optionally the sif parser can renumber all entities in the order they appear.
The idea is that it may be easier to copy-paste BCs, for example, between sif files.
This has not been taken into use but may be attractive to some users.
- Transient Restart: Enable restart in higher order time-stepping schemes with
> Transient Restart = Logical True < keyword in Simulation section. In effect the code will
save all the previous values needed also for higher order schemes. Note that this can result
to bloated output files. If this is not used the restart will always mean that the 1st timestep
uses 1st order time stepping scheme.
- Added a Predictor-Corrector time scheme for adaptive time stepping
method. It automatically adjusts the time steps based on stability
criteria and the requirement of the accuracy. An example is given in
"PredictorCorrector" test case.
- The user can choose to continue after all numerical/convergence errors by setting
"Global Abort Not Converged = Logical False". Alternatively, the user can specify
"Linear System Abort Not Converged = Logical False" in an individual solver section."
- Rotational normals: for rotating machines the vector fields should be continuous over the
rotating interface. This asks for optimal normal-tangential coordinate system. This is
achieved by setting > Rotational Normals = True < at the BC.
- Structured operations: Allow for midlayer in addition to top and bottom layers.
ElmerSolver library routines:
----------------------------
These are features that only affect those writing new code.
- Metainformation: GetVersion(), GetRevision() & GetCompilationDate() functions added
that can be used in writing meta information in the output files.
- The name of the MPI communicator in ElmerSolver was changed from MPI_COMM_WORLD
to ELMER_COMM_WORLD. The idea is that Elmer may be coupled with other codes (e.g. OpenFOAM)
using MPI such that both codes have their own internal communicator.
- Multithreading and vectorization
o Added OpenMP SIMD implementation of high-order H^1 elements for Line,
Triangle, Quadrilateral, Tetrahedral, Prismatic and Hexahedral elements.
o Added MultiColour Consistent (Logical) keyword for creating consistent mesh colorings
when requested (serial computation).
o GetCurrentElement and SetCurrentElement modified to give correct return values when
called from within OpenMP parallel regions.
o Added GetRealValues routine having the same functionality as GetReal but without having
to use an additional intermediate buffer.
- Keyword Optimization: the excessive retrieving of some keywords was reduced by
replacaing some commonly fetched keywords with hard-coded variables (e.g. DG and
Bubble flags). Also some operations where masked behind ListCheckPresentAnyBC
type of keywords.
- New more effective ListGet operations that are intended for element loops.
The ideas is that each keyword has a handle that helps to retrieve its value
with minimal effort. The speedup of the new routines is in the range from 1 to
20 depending on the type of evaluation. The new routines include:
o ListInitElementKeyword - initialize the keyword before element loop
o ListGetElementReal - retrieve real valued keyword in gaussian integration point
o ListGetElementRealVec - retrieve real valued keyword in all gaussian points
o ListGetElementInteger - retrieve integer valued keyword
o ListGetElementLogical - retrieve logical valued keyword
o ListGetElementString - retrieve string valued keyword
o ListCompareElementString - compare to given string and return result of comparison
- ListAddNew*: The keywords may now be conditionally added more economically by using
ListAddNewLogical etc. type of keywords that check whether the keyword is already defined.
The keyword value is only added if it does not exist. This can be used in setting default
values in _Init sections of the solvers.
- ChildSolver: A primary solver may create child solvers that share the same
active domain but may have different dofs. The intended use of this strategy is for
people working with various block solution techniques, for example.
The child solver may also inherit some keywords with an appropriate namespace.
- "Assert takes one LOGICAL mandatory argument, and stops the simulation if this argument
is false. E.g. CALL Assert(Mesh % NumberOfNodes > 0). The second and third arguments are
optional strings which specify the calling routine and the error message."
ElmerSolver Solver modules
---------------------------
These are improvements related to different solver modules, or completely new ones.
The Chapter for the documentaion of the solver in Elmer Models Manual is given if available.
Note that the documentation might not even include all the new features.
- MagnetoDynamics, Ch. 15
o Penalized Lagrangian formulation to remove null-space of the curl operator in statics.
Also in transient, but that is not theoretically justified. Useful sometimes with nonlinearities
and direct solver.
o Calculate nodal forces in harmonic case similarly airgap torque is calculated
o Enable calculation of harmonic average power P = E.sigma E/2. Calculate peak power by setting
> Calculate peak harmonic power = True <
o Real tensorial reluctivity. Enable with > Relative Reluctivity(3,3) < in material block.
Overrides related scalar linear keywords. Behaves additively .withHB curve.
o Enable saving of Joule losses also as nodal losses (in terms of Watts in SI units).
- Circuits and Dynamics (new), Ch. 16
o Model documentation
- Basic circuit equation functionality
- Massive, stranded and foil winding formulations
o Explicit coil resistance
o Output AC Power, DC resistance and AC resistance for circuit components
o Large number of new tests: massive, stranded and foil windings in 2D and 3D +
basic homogenization for stranded windings
- MagnetoDynamics2D, Ch. 18
o Proximity effect due to in plane magnetic field in foil windings
o Complex power computation for circuit components
o Lorentz force computation
- FourierLossSolver, Ch. 32
o A heavy facelift for the FourierLoss solver. It is now possible to use any number of loss components.
The visualization and export features have been improved etc.
- CoilSolver, Ch. 33
o Let user to explicitly define coilsolver normal
Coil Normal(3,1) = Real nx ny nz
- ElasticitySolver, Ch. 6
o Nonlinear elasticity solver can now handle incompressible neo-Hookean
materials via a mixed FE formulation
- ShellSolver (new), Ch, 7
o First version of a new solver for elastic shell equations based on a kinematic assumption of
Reissner-Naghdi type, see the section "Shell Equations of Linear Elasticity" in
Elmer Models Manual
o The solver is still under development and strain reduction operators needed in practice
have been worked out only for the lowest-order elements (a 3-node triangle or 4-node quadrilateral)
- RigidMeshMapper, Ch. 40
o Enable source term >Mesh Relax Source< to relaxation field solver, and
enable normalization to unity of this field.
- Elmer2OpenFoamIO (new), Ch. 52
o First version of a file based coupler from Elmer to OpenFOAM coupler.
The idea is that Elmer computes some field (i.e. heat source) that is written in
OpenFOAM format for a OpenFOAM solver to read. Currently only one-directional
operation is supported.
o Rudimentary documentation in Models Manual.
- VtuOutputSolver, Ch. 49
o The simple way of calling VtuOutputSolver is to set the >Post File< to have suffix .vtu.
Now this operation inherits all keywords from Simulation section with the prefix "vtu:",
for example >vtu: Save Elemental Fields = Logical False<.
o Enable vtu output using linear basis for DG fields.
o Enable mapping of geometry IDs such that the indexes in VTU format may be defined
in the sif file for each body and boundary.
o Enable user to save results only in the current mesh by specifying
> Save This Mesh Only = Logical True <
- SaveLine, Ch. 47
o Enable user given resolution rather than using the resolution obtained from the mesh faces.
Set > Polyline Intervals(n) = m1 m2 ... < to give number of subelements for each segment.
- CraigBamptonSolver (new)
o Solver can use eigen modes and constrained modes to perform model reduction.
o Undocumented solver.
- ScannedFieldSolver (new)
o Outputs the scanned fields with a prefix "Scan i" and
sums up the fields in the original field name
o Undocumented solver. (test: circuits2D_scan_harmonics)
- PartitionMeshSolver (new)
o Firsts steps for an internal ElmerSolver routine to perform partitioning.
o The idea is that it is easier to consider the constraints of physics with access to the whole case.
o Various geometric strategies implemented.
o Partitioning may be done in several groups enabling minimal communication between constacts, for example.
o Undocumented solver.
- ElementStats (new), Ch. 61
o Solver gives information on the mesh including element size, skewness, edge ratio.
o The results are reported in a range but also as a distrubution.
- DirectionSolver
o Computes body wise direction potentials for coils
o Undocumented solver
- RotMSolver
o Computes Rotation matrix based on direction potentials
o Undocumented solver
- WPotentialSolver
o Computes Wire Direction potential for coils based on rotation matrix and conductivity
o Undocumented solver
- ShellMultiSolver (new):
o Contributed solver provided by prof. Gupta.
This is placed under "modules/contrib" (as for contributed modules) as is compiled optionally.
o Documentation provided in the source directory.
ElmerGrid
---------
- Enable hybrid partitioning using radial splitting.
- Enable reading of mesh.names also in native Elmer format.
- Number of small fixes.
ElmerGUI
--------
- Add colors to the coundary condition editor if 'stateBcColors'.
Build system
------------
- Allow inclusion of contributed solvers under fem/src/modules/contrib
- Cleanup of redundant code
- Allow ElmerGrid to be compiled as a standalone application
- Allow splitting of modules in subdirectories, see fem/modules/SaveData for example.
- Enable QT5 support in ElmerGUI (cmake option WITH_QT5=TRUE)
- Added travis integration that builds and runs tests with `quick` label.
Tests
-----
- There are currently 533 tests which is a significant increase from the
403 of the previous release. This is without the Elmer/ICE specific tests.
Acknowledgements
----------------
Apart from the core Elmer team at CSC git log shows contributions from from
Mikko Byckling, Eelis Takala, Pavel Ponomarev, Erik Edelmann, Joe Todd, Cheng Gong,
Fabien Gillet-Chaulet, Laure Tavard, Erik Edelmann, Rupert Gladstone, Juris Vencels,
and Ondrej Meca. Additionally there are many ongoing deveopments in several branches
that have not been merged to this release, and are therefore not covered heere. Also
sometimes the code has been passed on by the original author by means other than the
git. This is the case, for example, with IDR(s) solver provided by Martin van Gijzen.
