A set of libraries and executables creating a workflow for performing
gradient-based optimisation loops. The main executable (adjointOptimisationFoam)
solves the flow (primal) equations, followed by the adjoint equations and,
eventually, the computation of sensitivity derivatives.
Current functionality supports the solution of the adjoint equations for
incompressible turbulent flows, including the adjoint to the Spalart-Allmaras
turbulence model and the adjoint to the nutUSpaldingWallFunction, [1], [2].
Sensitivity derivatives are computed with respect to the normal displacement of
boundary wall nodes/faces (the so-called sensitivity maps) following the
Enhanced Surface Integrals (E-SI) formulation, [3].
The software was developed by PCOpt/NTUA and FOSS GP, with contributions from
Dr. Evangelos Papoutsis-Kiachagias,
Konstantinos Gkaragounis,
Professor Kyriakos Giannakoglou,
Andy Heather
and contributions in earlier version from
Dr. Ioannis Kavvadias,
Dr. Alexandros Zymaris,
Dr. Dimitrios Papadimitriou
[1] A.S. Zymaris, D.I. Papadimitriou, K.C. Giannakoglou, and C. Othmer.
Continuous adjoint approach to the Spalart-Allmaras turbulence model for
incompressible flows. Computers & Fluids, 38(8):1528–1538, 2009.
[2] E.M. Papoutsis-Kiachagias and K.C. Giannakoglou. Continuous adjoint methods
for turbulent flows, applied to shape and topology optimization: Industrial
applications. 23(2):255–299, 2016.
[3] I.S. Kavvadias, E.M. Papoutsis-Kiachagias, and K.C. Giannakoglou. On the
proper treatment of grid sensitivities in continuous adjoint methods for shape
optimization. Journal of Computational Physics, 301:1–18, 2015.
Integration into the official OpenFOAM release by OpenCFD
Integration of VOF MULES new interfaces. Update of VOF solvers and all instances
of MULES in the code.
Integration of reactingTwoPhaseEuler and reactingMultiphaseEuler solvers and sub-models
Updating reactingEuler tutorials accordingly (most of them tested)
New eRefConst thermo used in tutorials. Some modifications at thermo specie level
affecting mostly eThermo. hThermo mostly unaffected
New chtMultiRegionTwoPhaseEulerFoam solver for quenching and tutorial.
Phases sub-models for reactingTwoPhaseEuler and reactingMultiphaseEuler were moved
to src/phaseSystemModels/reactingEulerFoam in order to be used by BC for
chtMultiRegionTwoPhaseEulerFoam.
Update of interCondensatingEvaporatingFoam solver.
- Extended runTimePostProcessing to include access to "live"
simulation objects such a geometry patches and sampled surfaces
stored on the "functionObjectObjects" registry.
- Add 'live' runTimePostProcessing of cloud data.
Extracts position and fields from the cloud via its objectRegistry writer
- For the "live" simulation objects, there are two new volume filters
that work directly with the OpenFOAM volume fields:
* iso-surface
* cutting planes
Both use the VTK algorithms directly and support multiple values.
Eg, can make multiple iso-levels or multiple planes parallel to each
other.
- When VTK has been compiled with MPI-support, parallel rendering will
be used.
- Additional title text properties (shadow, italic etc)
- Simplified handling of scalar-bar and visibility switches
- Support multiple text positions. Eg, for adding watermark text.
- fits better into the general sampling framework, improves flexibilty
and allows code reduction.
ENH: include surface fields on sampledSurfaces that support it
- changed the sectorCoeffs keyword to 'point' from 'axisPt'
for more similarity with other dictionaries.
Continue to accept 'axisPt' for compatibility.
- the utility had automatic triangulation removed some time ago, but
never changed its name.
- catch old uses with a surfaceMeshTriangulate deprecated script
- makes the intent clearer and avoids the need for additional
constructor casting. Eg,
labelList(10, Zero) vs. labelList(10, 0)
scalarField(10, Zero) vs. scalarField(10, scalar(0))
vectorField(10, Zero) vs. vectorField(10, vector::zero)
- this corresponds to 'never match', which may be useful in combination
with -constant selection.
Eg,
surfaceMeshTriangulate -constant -time none
selects only the constant entry and suppresses any automatic time loop
STYLE: adjust help for the standard -times option
- indicate that times can be comma or space separated, since this is
otherwise not apparent. Don't mention semicolon separators in the help
since that just adds even more clutter.
- support .vtp format for geometry, surface, line, cloud.
- use native reader for handling vtk, vtp, obj, stl surface files.
For other formats, use the MeshedSurface (the surfMesh lib) to
handle reading and Foam::vtk::Tools::Patch to handle the
conversion to vtkPolyData. This combination is more memory efficient.
- update tutorial case to include vtp surface geometry
- this allows more use of the runTimePostProcessing functionObject
that will fail more gracefully if the proper version could not be
built.
The dummy functionObject simply emits a message that it is not available.
Reports the min|max|average AMI weights to text file and optionally
writes VTK surfaces of the sum of the weights, and mask field for
ACMI patches.
Example usage:
AMIWeights
{
type AMIWeights;
libs ("libfieldFunctionObjects.so");
writeControl writeTime;
writeFields yes;
}
Reference:
Comte-Bellot, G., and Corrsin, S., "Simple Eulerian Time Correlation of
Full- and Narrow-Band Velocity Signals in Grid-Generated, 'Isotropic'
Turbulence," Journal of Fluid Mechanics, Vol. 48, No. 2, 1971,
pp. 273–337.
- tutorials based on squareBend used Default_Boundary_Region explicitly
defined since they predated the defaultPatch renaming (2008).
The name 'Default_Boundary_Region' was for convenience as the default
name when converting to PROSTAR or CCM formation, but can now be
changed to something more generic.
- define wall boundary conditions for squareBend using a general regex
to allow future splitting of wall types by name.
- parallel output.
The output is now postProcessing/<name> for similar reasoning as
mentioned in #866 - better alignment with other function objects, no
collision with foamToVTK output.
- align the input parameters with those of vtkCloud so that we can
specify the ASCII precision and the padding width for the output
file names as well.
- emit TimeValue field, support file series generation
- support internal or boundary meshes, combining the result into a vtm
file.
- can restrict conversion based on zone names, enclosing volumes,
bounding box
- deprecate dimensionedType constructors using an Istream in favour of
versions accepting a keyword and a dictionary.
Dictionary entries are almost the exclusive means of read
constructing a dimensionedType. By construct from the dictionary
entry instead of doing a lookup() first, we can detect possible
input errors such as too many tokens as a result of a input syntax
error.
Constructing a dimensionedType from a dictionary entry now has
two forms.
1. dimensionedType(key, dims, dict);
This is the constructor that will normally be used.
It accepts entries with optional leading names and/or
dimensions. If the entry contains dimensions, they are
verified against the expected dimensions and an IOError is
raised if they do not correspond. On conclusion, checks the
token stream for any trailing rubbish.
2. dimensionedType(key, dict);
This constructor is used less frequently.
Similar to the previous description, except that it is initially
dimensionless. If entry contains dimensions, they are used
without further verification. The constructor also includes a
token stream check.
This constructor is useful when the dimensions are entirely
defined from the dictionary input, but also when handling
transition code where the input dimensions are not obvious from
the source.
This constructor can also be handy when obtaining values from
a dictionary without needing to worry about the input dimensions.
For example,
Info<< "rho: " << dimensionedScalar("rho", dict).value() << nl;
This will accept a large range of inputs without hassle.
ENH: consistent handling of dimensionedType for inputs (#1083)
BUG: incorrect Omega dimensions (fixes#2084)
- helps reduce clutter in the topoSetDict files.
Caveats when using this.
The older specification styles using "name" will conflict with the
set name. Eg,
{
name f0
type faceSet;
action add;
source patchToFace;
sourceInfo
{
name inlet;
}
}
would flattened to the following
{
name f0
type faceSet;
action add;
source patchToFace;
name inlet;
}
which overwrites the "name" used for the faceSet.
The solution is to use the updated syntax:
{
name f0
type faceSet;
action add;
source patchToFace;
patch inlet;
}
- old 'DELETE' enum was easily confused with 'REMOVE', which removes
the set, not the elements from the set.
- provide corresponding subtractSet() method
STYLE: HashSet set/unset instead of insert/erase methods in topoSetSource
- simplifies switching to/from bitSet storage
Previously the coordinate system functionality was split between
coordinateSystem and coordinateRotation. The coordinateRotation stored
the rotation tensor and handled all tensor transformations.
The functionality has now been revised and consolidated into the
coordinateSystem classes. The sole purpose of coordinateRotation
is now just to provide a selectable mechanism of how to define the
rotation tensor (eg, axis-angle, euler angles, local axes) for user
input, but after providing the appropriate rotation tensor it has
no further influence on the transformations.
--
The coordinateSystem class now contains an origin and a base rotation
tensor directly and various transformation methods.
- The origin represents the "shift" for a local coordinate system.
- The base rotation tensor represents the "tilt" or orientation
of the local coordinate system in general (eg, for mapping
positions), but may require position-dependent tensors when
transforming vectors and tensors.
For some coordinate systems (currently the cylindrical coordinate system),
the rotation tensor required for rotating a vector or tensor is
position-dependent.
The new coordinateSystem and its derivates (cartesian, cylindrical,
indirect) now provide a uniform() method to define if the rotation
tensor is position dependent/independent.
The coordinateSystem transform and invTransform methods are now
available in two-parameter forms for obtaining position-dependent
rotation tensors. Eg,
... = cs.transform(globalPt, someVector);
In some cases it can be useful to use query uniform() to avoid
storage of redundant values.
if (cs.uniform())
{
vector xx = cs.transform(someVector);
}
else
{
List<vector> xx = cs.transform(manyPoints, someVector);
}
Support transform/invTransform for common data types:
(scalar, vector, sphericalTensor, symmTensor, tensor).
====================
Breaking Changes
====================
- These changes to coordinate systems and rotations may represent
a breaking change for existing user coding.
- Relocating the rotation tensor into coordinateSystem itself means
that the coordinate system 'R()' method now returns the rotation
directly instead of the coordinateRotation. The method name 'R()'
was chosen for consistency with other low-level entities (eg,
quaternion).
The following changes will be needed in coding:
Old: tensor rot = cs.R().R();
New: tensor rot = cs.R();
Old: cs.R().transform(...);
New: cs.transform(...);
Accessing the runTime selectable coordinateRotation
has moved to the rotation() method:
Old: Info<< "Rotation input: " << cs.R() << nl;
New: Info<< "Rotation input: " << cs.rotation() << nl;
- Naming consistency changes may also cause code to break.
Old: transformVector()
New: transformPrincipal()
The old method name transformTensor() now simply becomes transform().
====================
New methods
====================
For operations requiring caching of the coordinate rotations, the
'R()' method can be used with multiple input points:
tensorField rots(cs.R(somePoints));
and later
Foam::transformList(rots, someVectors);
The rotation() method can also be used to change the rotation tensor
via a new coordinateRotation definition (issue #879).
The new methods transformPoint/invTransformPoint provide
transformations with an origin offset using Cartesian for both local
and global points. These can be used to determine the local position
based on the origin/rotation without interpreting it as a r-theta-z
value, for example.
================
Input format
================
- Streamline dictionary input requirements
* The default type is cartesian.
* The default rotation type is the commonly used axes rotation
specification (with e1/e2/3), which is assumed if the 'rotation'
sub-dictionary does not exist.
Example,
Compact specification:
coordinateSystem
{
origin (0 0 0);
e2 (0 1 0);
e3 (0.5 0 0.866025);
}
Full specification (also accepts the longer 'coordinateRotation'
sub-dictionary name):
coordinateSystem
{
type cartesian;
origin (0 0 0);
rotation
{
type axes;
e2 (0 1 0);
e3 (0.5 0 0.866025);
}
}
This simplifies the input for many cases.
- Additional rotation specification 'none' (an identity rotation):
coordinateSystem
{
origin (0 0 0);
rotation { type none; }
}
- Additional rotation specification 'axisAngle', which is similar
to the -rotate-angle option for transforming points (issue #660).
For some cases this can be more intuitive.
For example,
rotation
{
type axisAngle;
axis (0 1 0);
angle 30;
}
vs.
rotation
{
type axes;
e2 (0 1 0);
e3 (0.5 0 0.866025);
}
- shorter names (or older longer names) for the coordinate rotation
specification.
euler EulerRotation
starcd STARCDRotation
axes axesRotation
================
Coding Style
================
- use Foam::coordSystem namespace for categories of coordinate systems
(cartesian, cylindrical, indirect). This reduces potential name
clashes and makes a clearer declaration. Eg,
coordSystem::cartesian csys_;
The older names (eg, cartesianCS, etc) remain available via typedefs.
- added coordinateRotations namespace for better organization and
reduce potential name clashes.
- functionObjectLibs -> libs
- redirectType -> name
- change deprecated writeCompression flags types to Switch.
- cleanup some trailing ';;' from some dictionaries
- for larger problems with a smaller region of interest, can apply a
bounding to limit the size of the ensight geometry and fields created.
Since the implementation uses a fvMeshSubset, there is an additional
per-process memory overhead.
A high output frequency should be avoided with moving meshes, since
this indirectly forces a frequent update of the submesh.
- signedDistance() method is like distance() but retains
the positive/negative sign for the side of the plane.
- the sign() method returns the sign as -1,0,+1 integer for
classification purposes where it is important to distinguish between
a zero value and a positive value (eg, for cutting). Optional
tolerance can be supplied to round for zero.
- refactor and inlined simple and frequently used methods.
- add boundBox faceCentre() method, which can be useful for creating
clipping planes from a bounding box.
Relocated treeBoundBox faceNormals to boundBox since they apply
equally there - the meaning of the faces (x-min, x-max, etc)
is the same, even if the point addressing for the faces differs.
