The distributions have been extended in various ways to permit usage in
a greater variety of situations...
The distributions now have write methods which allow a distribution to
be written into a field file for restart, therby permitting their usage
in boundary conditions and similar. Their read methods now also support
dimension-checked unit conversions for all their parameters.
An additional selector has been added that allows a distribution to be
re-constructed with a different sample size exponent.
The distributions now own their random generator, thereby simplifying
their usage and preventing the need for a (potentially dangling)
reference member. This makes sense now as the random generators do not
rely on global state; individual sub-models can and should own their own
random generator and manage its initialisation and restart. This
principle should be extended to other parts of the code in future.
This new class hierarchy replaces the distributions previously provided
by the Lagrangian library.
All distributions (except fixedValue) now require a "size exponent", Q,
to be specified along with their other coefficients. If a distribution's
CDF(x) (cumulative distribution function) represents what proportion of
the distribution takes a value below x, then Q determines what is meant
by "proportion":
- If Q=0, then "proportion" means the number of sampled values expected
to be below x divided by the total number of sampled values.
- If Q=3, then "proportion" means the expected sum of sampled values
cubed for values below x divided by the total sum of values cubed. If
x is a length, then this can be interpreted as a proportion of the
total volume of sampled objects.
- If Q=2, and x is a length, then the distribution might represent the
proportion of surface area, and so on...
In addition to the user-specification of Q defining what size the given
distribution relates to, an implementation that uses a distribution can
also programmatically define a samplingQ to determine what sort of
sample is being constructed; whether the samples should have an equal
number (sampleQ=0), volume (sampleQ=3), area (sampleQ=2), etc...
A number of fixes to the distributions have been made, including fixing
some fundamental bugs in the returned distribution of samples, incorrect
calculation of the distribution means, renaming misleadingly named
parameters, and correcting some inconsistencies in the way in which
tabulated PDF and CDF data was processed. Distributions no longer
require their parameters to be defined in a sub-dictionary, but a
sub-dictionary is still supported for backwards compatibility.
The distributions can now generate their PDF-s as well as samples, and a
test application has been added (replacing two previous applications),
which thoroughly checks consistency between the PDF and the samples for
a variety of combinations of values of Q and sampleQ.
Backwards incompatible changes are as follows:
- The standard deviation keyword for the normal (and multi-normal)
distribution is now called 'sigma'. Previously this was 'variance',
which was misleading, as the value is a standard deviation.
- The 'massRosinRammler' distribution has been removed. This
functionality is now provided by the standard 'RosinRammler'
distributon with a Q equal to 0, and a sampleQ of 3.
- The 'general' distribution has been split into separate distributions
based on whether PDF or CDF data is provided. These distributions are
called 'tabulatedDensity' and 'tabulatedCumulative', respectively.
The timeName() function simply returns the dimensionedScalar::name() which holds
the user-time name of the current time and now that timeName() is no longer
virtual the dimensionedScalar::name() can be called directly. The timeName()
function implementation is maintained for backward-compatibility.
Sampled sets and streamlines now write all their fields to the same
file. This prevents excessive duplication of the geometry and makes
post-processing tasks more convenient.
"axis" entries are now optional in sampled sets and streamlines. When
omitted, a default entry will be used, which is chosen appropriately for
the coordinate set and the write format. Some combinations are not
supported. For example, a scalar ("x", "y", "z" or "distance") axis
cannot be used to write in the vtk format, as vtk requires 3D locations
with which to associate data. Similarly, a point ("xyz") axis cannot be
used with the gnuplot format, as gnuplot needs a single scalar to
associate with the x-axis.
Streamlines can now write out fields of any type, not just scalars and
vectors, and there is no longer a strict requirement for velocity to be
one of the fields.
Streamlines now output to postProcessing/<functionName>/time/<file> in
the same way as other functions. The additional "sets" subdirectory has
been removed.
The raw set writer now aligns columns correctly.
The handling of segments in coordSet and sampledSet has been
fixed/completed. Segments mean that a coordinate set can represent a
number of contiguous lines, disconnected points, or some combination
thereof. This works in parallel; segments remain contiguous across
processor boundaries. Set writers now only need one write method, as the
previous "writeTracks" functionality is now handled by streamlines
providing the writer with the appropriate segment structure.
Coordinate sets and set writers now have a convenient programmatic
interface. To write a graph of A and B against some coordinate X, in
gnuplot format, we can call the following:
setWriter::New("gnuplot")->write
(
directoryName,
graphName,
coordSet(true, "X", X), // <-- "true" indicates a contiguous
"A", // line, "false" would mean
A, // disconnected points
"B",
B
);
This write function is variadic. It supports any number of
field-name-field pairs, and they can be of any primitive type.
Support for Jplot and Xmgrace formats has been removed. Raw, CSV,
Gnuplot, VTK and Ensight formats are all still available.
The old "graph" functionality has been removed from the code, with the
exception of the randomProcesses library and associated applications
(noise, DNSFoam and boxTurb). The intention is that these should also
eventually be converted to use the setWriters. For now, so that it is
clear that the "graph" functionality is not to be used elsewhere, it has
been moved into a subdirectory of the randomProcesses library.
This is faster than the library functionality that it replaces, as it
allows the compiler to do inlining. It also does not utilise any static
state so generators do not interfere with each other. It is also faster
than the the array lookup in cachedRandom. The cachedRandom class
therefore offers no advantage over Random and has been removed.
