GitLab

(old keywords are still valid, throwing a compatibility warning)

- using (U,V,W) instead of (X1,X2,X3)
- using confine instead of bound

for all objective functions.

- The normalization is useful for practically all update methods dealing
with constraints (e.g. SQP, MMA). The normalization factor can be either
given explicitly or, if not given, will be the value of the objective
function in the first optimisation cycle.
- The target value is useful when using the objective as a constraint in
constrained optimisation problems (e.g. drag - dragTarget). It should
only be used with update methods that understand the value of the
constraint (e.g. SQP, MMA) but not when the objective in hand is the
only objective of the optimisation problem. In such a case, a squared
objective should be used (e.g. sqr(drag - dragTarget))

- Objective now inherits from localIOdictionary and writes the mean
objective value under the uniform folder, each time mesh.write() is
called. This is crucial for getting the correct old merit function value
if the simulation is continued from a previous state and lineSearch is
used.
- Objectives are now computed and written even if the corresponding
adjoint solver is inactive. This, among others, is also essential for
getting the correct old merit function value in case of continuation.
- Writing of the objective function (and its mean, if present) history
has now moved to updatePrimalBasedQuantities, instead of the preLoop
part of the adjoint solvers. This was decided to get the objective
values to files, even if the adjoint solver is inactive. Arguably, an
even better place to write the objective functions would be the postLoop
part of the primal solvers, however this might cause multiple writes of
the objective value for the inner iterations of lineSearch, if one is
used.

to be used in separating instantaneous objective values of different
optimisation cycles.

which qualitatively quantifies noise through a volume integral of the squared
turbulent viscosity.

Moved part common to all derived classes (e.g. update) to the base
class to avoid code duplication. Practically, only the protected
updateDesignVariables has to be overwritten in each derived class now.
steadyOptimisation was also affected in a minor way.

holding the sensitivity maps

given a global control point ID

without recomputing them. Use with caution!

- Added function returning the underlaying surface sensitivities
- Added boolean to control whether to write the underlaying sensitivity
  map (defaults to false)

Encapsulates all terms that are common in both E(SI) and FI
formulations, like direct sensitivities and sensitivities due to
primal boundary conditions. Added the latter to all derived sensitivity
types, except for sensitivity maps.

Same as adjointWallVelocity but also returns the contribution
of the differentiation of the rotatingWallVelocity BC wrt the
face centres, to be added to the sensitivity derivatives.

Main reason was the insertion of a templated virtual function
returning the contribution of the differentiation of the primal
boundary condition, in the case the latter directly depends on a
a geometric quantity (e.g. rotatingWallVelocity).

- previously introduced `getOrDefault` as a dictionary _get_ method,
  now complete the transition and use it everywhere instead of
  `lookupOrDefault`. This avoids mixed usage of the two methods that
  are identical in behaviour, makes for shorter names, and promotes
  the distinction between "lookup" access (ie, return a token stream,
  locate and return an entry) and "get" access (ie, the above with
  conversion to concrete types such as scalar, label etc).

- makes format of ExecutionTime = ... output configurable (#788)
  and reduces code clutter.

- use dictionary::get<..> instead of lookup in a few more places

- reduces the number of parameters that are being passed around
  and allows future additions into the IOstreamOption with mininal
  effort.

When more than one volumetric B-Splines control boxes are present, the
sensitivity constituents corresponding to the non-active design
variables were not bounded(zeroed) correctly. The resultant
sensitivities, used in the optimization, were bounded correctly, so this
was more a bug pertaining to the output file of the sensitivities rather
than a functional one.

When activeDesignVariables are not set explicitly, all design variables
are treated as active. These were allocated properly when starting from
0 but not when starting from an intermediate optimisation cycle
(e.g. running 5 optimisation cycles, stopping and restarting).

TUT: added a new tutorial including the restart of an optimisation run
to help identify future regression

- The core of the FatalIOError message was not printed due to exiting
  with FatalError instead of FatalIOError
- Changed the TypeName in all derived classes of displacementMethod so
  that the toc printed by the FatalIOError corresponds to what the user
  should add in dynamicMeshDict

to allow for continuation with averaging, line search and multiple
objectives

returning number of design variables per direction when symmetry is
applied

to parameterize regions not centered in (0,0,0)

findPointsInBox was receiving points in the Cartesian system,
not local coordinate system

- Failed due to double*Matrix<float> multiplication.

Style changes

- use SquareMatrix with Identity on construction

- use Zero in constructors

- remove trailing space and semi-colons

The adjoint library is enhanced with new functionality enabling
automated shape optimisation loops.  A parameterisation scheme based on
volumetric B-Splines is introduced, the control points of which act as
the design variables in the optimisation loop [1, 2].  The control
points of the volumetric B-Splines boxes can be defined in either
Cartesian or cylindrical coordinates.

The entire loop (solution of the flow and adjoint equations, computation
of sensitivity derivatives, update of the design variables and mesh) is
run within adjointOptimisationFoam. A number of methods to update the
design variables are implemented, including popular Quasi-Newton methods
like BFGS and methods capable of handling constraints like loop using
the SQP or constraint projection.

The software was developed by PCOpt/NTUA and FOSS GP, with contributions from

Dr. Evangelos Papoutsis-Kiachagias,
Konstantinos Gkaragounis,
Professor Kyriakos Giannakoglou,
Andy Heather

[1] E.M. Papoutsis-Kiachagias, N. Magoulas, J. Mueller, C. Othmer,
K.C.  Giannakoglou: 'Noise Reduction in Car Aerodynamics using a
Surrogate Objective Function and the Continuous  Adjoint Method with
Wall Functions', Computers & Fluids, 122:223-232, 2015

[2] E. M. Papoutsis-Kiachagias, V. G. Asouti, K. C. Giannakoglou,
K.  Gkagkas, S. Shimokawa, E. Itakura: ‘Multi-point aerodynamic shape
optimization of cars based on continuous adjoint’, Structural and
Multidisciplinary Optimization, 59(2):675–694, 2019

BUG: RASModelVariables::SpalartAllmaras cannot be combined with an fvMotionSolver diffusivity which depends on wall distances (fixes #1501)