Adjoint Capabilities

Discrete adjoint capabilities are now available in MULTIF, in addition to finite differences (for the mass and thrust QoI only).

Important note : This requires the latest custom version of SU2 (see Installation#su2) configured with Automatic Differentiation (AD) support (see Information about AD build).

The only additional options in the MULTIF configuration file related to gradients computation are the following.

% ---- GRADIENTS COMPUTATIONS ----

% Output a gradient file? (only mass and thrust for now)
OUTPUT_GRADIENTS= YES
OUTPUT_GRADIENTS_FILENAME= grad.dat

% Method used for gradients computation (ADJOINT or FINITE_DIFF)
GRADIENTS_COMPUTATION_METHOD= ADJOINT

• OUTPUT_GRADIENTS : If set to YES, MULTIF will output a file (different from the MULTIF output file, and whose name is set by OUTPUT_GRADIENTS_FILENAME) containing thrust gradients with respect to the design variables. Note that even if this option is set to NO, gradients might be computed if they are required by a dakota input file (params.in).
• GRADIENTS_COMPUTATION_METHOD : use the discrete adjoint capabilities of SU2, or finite differences.

Example

An example is available in {MULTIF_ROOT}/example/gradients.

Note: The latest custom version of SU2 is required to run this example.

The input files are the following:

• general_gradients.cfg : configuration file for MULTIF.
• params.in : a Dakota input file containing the values of the design variables and requesting the value and gradients computations for mass and thrust.

This example can be run using this command line:

python /path/to/MULTIF/runModel.py -f general_gradients.cfg -l 0 -n 7

Three fidelity levels are defined in general_gradients.cfg, for which only the mesh complexities change. The flag -n specifies the number of cores to use when running parallel calculation in SU2.

The gradients will be outputted in two different files:

• in grad.dat : as required by the OUTPUT_GRADIENTS options,
• and in results.out, the dakota interface file.