Parts of the adjoint optimisation library were re-designed to generalise
the way sensitivity derivatives (SDs) are computed and to allow easier
extension to primal problems other than the ones governed by
incompressible flows. In specific:
- the adjoint solver now holds virtual functions returning the part of
SDs that depends only on the primal and the adjoint fields.
- a new class named designVariables was introduced which, apart from
defining the design variables of the optimisation problem and
providing hooks for updating them in an optimisation loop, provides
the part of the SDs that affects directly the flow residuals (e.g.
geometric variations in shape optimisation, derivatives of source
terms in topology optimisation, etc). The final assembly of the SDs
happens here, with the updated sensitivity class acting as an
intermediate.
With the new structure, when the primal problem changes (for instance,
passive scalars are included), the same design variables and sensitivity
classes can be re-used for all physics, with additional contributions to
the SDs being limited (and contained) to the new adjoint solver to be
implemented. The old code structure would require new SD classes for
each additional primal problem.
As a side-effect, setting up a case has arguably become a bit easier and
more intuitive.
Additional changes include:
---------------------------
- Changes in the formulation and computation of shape sensitivity derivatives
using the E-SI approach. The latter is now derived directly from the
FI approach, with proper discretization for the terms and boundary
conditions that emerge from applying the Gauss divergence theorem used
to transition from FI to E-SI. When E-SI and FI are based on the same
Laplace grid displacement model, they are now numerically equivalent
(the previous formulation proved the theoretical equivalence of the
two approaches but numerical results could differ, depending on the
case).
- Sensitivity maps at faces are now computed based (and are deriving
from) sensitivity maps at points, with a constistent point-to-face
interpolation (requires the differentiation of volPointInterpolation).
- The objective class now allocates only the member pointers that
correspond to the non-zero derivatives of the objective w.r.t. the
flow and geometric quantities, leading to a reduced memory footprint.
Additionally, contributions from volume-based objectives to the
adjoint equations have been re-worked, removing the need for
objectiveManager to be virtual.
- In constrained optimisation, an adjoint solver needs to be present for
each constraint function. For geometric constraints though, no adjoint
equations need to solved. This is now accounted for through the null
adjoint solver and the geometric objectives which do not allocate
adjoint fields for this kind of constraints, reducing memory
requirements and file clutter.
- Refactoring of the updateMethod to collaborate with the new
designVariables. Additionally, all updateMethods can now read and
write restart data in binary, facilitating exact continuation.
Furthermore, code shared by various quasi-Newton methods (BFGS, DBFGS,
LBFGS, SR1) has been organised in the namesake class. Over and above,
an SQP variant capable of tackling inequality constraints has been
added (ISQP, with I indicating that the QP problem in the presence of
inequality constraints is solved through an interior point method).
Inequality constraints can be one-sided (constraint < upper-value)
or double-sided (lower-value < constraint < upper-value).
- Bounds can now be defined for the design variables.
For volumetricBSplines in specific, these can be computed as the
mid-points of the control points and their neighbouring ones. This
usually leads to better-defined optimisation problems and reduces the
chances of an invalid mesh during optimisation.
- Convergence criteria can now be defined for the optimisation loop
which will stop if the relative objective function reduction over
the last objective value is lower than a given threshold and
constraints are satisfied within a give tolerance. If no criteria are
defined, the optimisation will run for the max. given number of cycles
provided in controlDict.
- Added a new grid displacement method based on the p-Laplacian
equation, which seems to outperform other PDE-based approaches.
TUT: updated the shape optimisation tutorials and added a new one
showcasing the use of double-sided constraints, ISQP, applying
no-overlapping constraints to volumetric B-Splines control points
and defining convergence criteria for the optimisation loop.
- use Foam::zero as a dispatch tag
FIX: return moleculeCloud::constProps() List by reference not copy
STYLE: range-for when iterating cloud parcels
STYLE: more consistent typedefs / declarations for Clouds
- in most cases a parallel-consistent order is required.
Even when the order is not important, it will generally require
fewer allocations to create a UPtrList of entries instead of a
HashTable or even a wordList.
- update TimeState access methods
- use writeTime() instead of old method name outputTime()
- use deltaTValue() instead of deltaT().value()
to avoids pointless construct of intermediate
- provides a more succinct way of writing
{fa,fv}PatchField<Type>::patchInternalField(*this)
as well as a consistent naming that can be used for patches derived
from valuePointPatchField
ENH: readGradientEntry helper method for fixedGradient conditions
- simplifies coding and logic.
- support different read construct modes for fixedGradient
- construct from components, or use word::null to ensure
consistent avoid naming between IOobject vs dimensioned type.
- support construct with parameter ordering as per DimensionedField
ENH: instantiate a uniformDimensionedLabelField
- eg, for registering standalone integer counters
- attributes such as assignable(), coupled() etc
- common patchField types: calculatedType(), zeroGradientType() etc.
This simplifies reference to these types without actually needing a
typed patchField version.
ENH: add some basic patchField types to fieldTypes namespace
- allows more general use of the names
ENH: set extrapolated/calculated from patchInternalField directly
- avoids intermediate tmp
- newer naming allows for less confusing code.
Eg,
max(lower) -> clamp_min(lower)
min(upper) -> clamp_max(upper)
- prefer combined method, for few operations.
Eg,
max(lower) + min(upper) -> clamp_range(lower, upper)
The updated naming also helps avoid some obvious coding errors.
Eg,
Re.min(1200.0);
Re.max(18800.0);
instead of
Re.clamp_range(1200.0, 18800.0);
- can also use implicit conversion of zero_one to MinMax<Type> for
this type of code:
lambda_.clamp_range(zero_one{});
