- adjoint_systemName of the system representing the adjoint problem.
C++ Type:NonlinearSystemName
Unit:(no unit assumed)
Controllable:No
Description:Name of the system representing the adjoint problem.
- forward_systemName of the system representing the forward problem.
C++ Type:NonlinearSystemName
Unit:(no unit assumed)
Controllable:No
Description:Name of the system representing the forward problem.
SteadyAndAdjoint
Executioner for evaluating steady-state simulations and their adjoint.
Overview
This executioner can be used to solve a steady-state forward problem with its adjoint. The forward solve is performed the same way as in Steady, but with the "solve_type" set to NEWTON
. This performs the nonlinear iteration in the form:
where is the number of iterations it took to converge the problem. The adjoint problem is then defined as:
where is the adjoint solution, is the residual of the adjoint system with , and is the transpose of the forward system's Jacobian evaluated with the converged forward solution. The adjoint system is basically a linearized and homogenized version of the forward problem with it's own definition of sources.
In order to accurately define the adjoint system, the fully consistent Jacobian must be evaluated. As such, the computeQpJacobian
routines in the forward problem kernels must be accurately defined or Automatic Differentiation must be used. Consider using the Jacobian debugger to ensure the Jacobian is computed accurately.
Example Input File Syntax
Solving Adjoint Problems
The first step is to add an adjoint nonlinear system using the "nl_sys_names" parameter in the Problem input block. It is convenient to define the forward system as nl0
.
[Problem]
nl_sys_names = 'nl0 adjoint'
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/self_adjoint.i)Next we need to add an adjoint variable for each forward variable, which is associated with the adjoint
system:
Single adjoint variable
[Variables]
[u]
[]
[u_adjoint]
solver_sys = adjoint
[]
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/self_adjoint.i)Multiple adjoint variables
[Variables]
[u]
[]
[v]
[]
[u_adjoint]
solver_sys = adjoint
[]
[v_adjoint]
solver_sys = adjoint
[]
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/multi_variable.i)Array adjoint variables
[Variables]
[u]
components = 2
[]
[u_adjoint]
components = 2
solver_sys = adjoint
[]
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/array_variable.i)Next we add kernels and BCs associated with the forward and adjoint variables. Only source-like kernels should be added to the adjoint variables like BodyForce, ConstantPointSource, or NeumannBC.
[Kernels]
[diff]
type = Diffusion
variable = u
[]
[src]
type = BodyForce
variable = u
value = 1
[]
[src_adjoint]
type = BodyForce
variable = u_adjoint
value = 10
[]
[]
[BCs]
[dirichlet]
type = DirichletBC
variable = u
boundary = 'top right'
value = 1
[]
[neumann]
type = NeumannBC
variable = u
boundary = 'left bottom'
value = 1
[]
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/nonhomogeneous_bc.i)For nonlinear, problems one should use AD
Kernels, BCs, and Materials.
[Kernels]
[diff]
type = ADMatDiffusion
variable = u
diffusivity = D
[]
[src]
type = ADBodyForce
variable = u
value = 1
[]
[src_adjoint]
type = ADBodyForce
variable = u_adjoint
value = 1
[]
[]
[BCs]
[dirichlet]
type = ADDirichletBC
variable = u
boundary = 'top right'
value = 0
[]
[]
[Materials]
[diffc]
type = ADParsedMaterial
property_name = D
expression = '0.1 + 5 * u'
coupled_variables = 'u'
[]
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/nonlinear_diffusion.i)Finally, we will add this executioner and set the forward/adjoint system tags. Note that the tolerance for the adjoint system solve is set solely by linear solver parameters like "l_tol", "l_abs_tol", and "l_max_its".
[Executioner]
type = SteadyAndAdjoint
forward_system = nl0
adjoint_system = adjoint
nl_rel_tol = 1e-12
l_tol = 1e-12
[]
(modules/optimization/test/tests/executioners/steady_and_adjoint/nonhomogeneous_bc.i)Utilization in Gradient-Based Optimization
Utilizing this executioner for gradient-based optimization is quite powerful since the adjoint used to compute the gradient is automatically assembled with this executioner, given that the forward problem Jacobian can be fully constructed.
To include the source for the adjoint problem, the ReporterPointSource can be used to add the simulation misfit from the forward solve, which is calculated in OptimizationData.
[Reporters]
[measurement_locations]
type = OptimizationData
objective_name = objective_value
variable = forwardT
[]
[params]
type = ConstantReporter
real_vector_names = 'heat_source'
real_vector_values = '0' # Dummy
[]
[]
[DiracKernels]
[pt]
type = ReporterPointSource
variable = adjointT
x_coord_name = measurement_locations/measurement_xcoord
y_coord_name = measurement_locations/measurement_ycoord
z_coord_name = measurement_locations/measurement_zcoord
value_name = measurement_locations/misfit_values
[]
[]
(modules/optimization/test/tests/optimizationreporter/nonlinear_material/forward_and_adjoint.i)The gradient can then be computed using an inner-product vector-postprocessor like ElementOptimizationSourceFunctionInnerProduct. Note that these vector-postprocessors must be executed on ADJOINT_TIMESTEP_END
which occurs after the adjoint system is solved.
[VectorPostprocessors]
[gradient_vpp]
type = ElementOptimizationSourceFunctionInnerProduct
function = volumetric_heat_func
variable = adjointT
execute_on = ADJOINT_TIMESTEP_END
[]
[]
(modules/optimization/test/tests/optimizationreporter/nonlinear_material/forward_and_adjoint.i)The driving optimize app will thus have only have a single FullSolveMultiApp which then transfers both the simulation data (for the objective evaluation) and the inner products (for the gradient evaluation).
[MultiApps]
[forward]
type = FullSolveMultiApp
input_files = forward_and_adjoint.i
execute_on = FORWARD
[]
[]
[Transfers]
[to_forward]
type = MultiAppReporterTransfer
to_multi_app = forward
from_reporters = 'main/measurement_xcoord
main/measurement_ycoord
main/measurement_zcoord
main/measurement_time
main/measurement_values
OptimizationReporter/heat_source'
to_reporters = 'measurement_locations/measurement_xcoord
measurement_locations/measurement_ycoord
measurement_locations/measurement_zcoord
measurement_locations/measurement_time
measurement_locations/measurement_values
params/heat_source'
[]
[from_forward]
type = MultiAppReporterTransfer
from_multi_app = forward
from_reporters = 'measurement_locations/objective_value
gradient_vpp/inner_product'
to_reporters = 'OptimizationReporter/objective_value
OptimizationReporter/grad_heat_source'
[]
[]
(modules/optimization/test/tests/optimizationreporter/nonlinear_material/main.i)Input Parameters
- time0System time
Default:0
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:System time
- verboseFalseSet to true to print additional information
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Set to true to print additional information
Optional Parameters
- accept_on_max_fixed_point_iterationFalseTrue to treat reaching the maximum number of fixed point iterations as converged.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:True to treat reaching the maximum number of fixed point iterations as converged.
- auto_advanceFalseWhether to automatically advance sub-applications regardless of whether their solve converges, for transient executioners only.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Whether to automatically advance sub-applications regardless of whether their solve converges, for transient executioners only.
- custom_abs_tol1e-50The absolute nonlinear residual to shoot for during fixed point iterations. This check is performed based on postprocessor defined by the custom_pp residual.
Default:1e-50
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:The absolute nonlinear residual to shoot for during fixed point iterations. This check is performed based on postprocessor defined by the custom_pp residual.
- custom_ppPostprocessor for custom fixed point convergence check.
C++ Type:PostprocessorName
Unit:(no unit assumed)
Controllable:No
Description:Postprocessor for custom fixed point convergence check.
- custom_rel_tol1e-08The relative nonlinear residual drop to shoot for during fixed point iterations. This check is performed based on the postprocessor defined by custom_pp residual.
Default:1e-08
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:The relative nonlinear residual drop to shoot for during fixed point iterations. This check is performed based on the postprocessor defined by custom_pp residual.
- direct_pp_valueFalseTrue to use direct postprocessor value (scaled by value on first iteration). False (default) to use difference in postprocessor value between fixed point iterations.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:True to use direct postprocessor value (scaled by value on first iteration). False (default) to use difference in postprocessor value between fixed point iterations.
- disable_fixed_point_residual_norm_checkFalseDisable the residual norm evaluation thus the three parameters fixed_point_rel_tol, fixed_point_abs_tol and fixed_point_force_norms.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Disable the residual norm evaluation thus the three parameters fixed_point_rel_tol, fixed_point_abs_tol and fixed_point_force_norms.
- fixed_point_abs_tol1e-50The absolute nonlinear residual to shoot for during fixed point iterations. This check is performed based on the main app's nonlinear residual.
Default:1e-50
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:The absolute nonlinear residual to shoot for during fixed point iterations. This check is performed based on the main app's nonlinear residual.
- fixed_point_algorithmpicardThe fixed point algorithm to converge the sequence of problems.
Default:picard
C++ Type:MooseEnum
Unit:(no unit assumed)
Controllable:No
Description:The fixed point algorithm to converge the sequence of problems.
- fixed_point_force_normsFalseForce the evaluation of both the TIMESTEP_BEGIN and TIMESTEP_END norms regardless of the existence of active MultiApps with those execute_on flags, default: false.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Force the evaluation of both the TIMESTEP_BEGIN and TIMESTEP_END norms regardless of the existence of active MultiApps with those execute_on flags, default: false.
- fixed_point_max_its1Specifies the maximum number of fixed point iterations.
Default:1
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Specifies the maximum number of fixed point iterations.
- fixed_point_min_its1Specifies the minimum number of fixed point iterations.
Default:1
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Specifies the minimum number of fixed point iterations.
- fixed_point_rel_tol1e-08The relative nonlinear residual drop to shoot for during fixed point iterations. This check is performed based on the main app's nonlinear residual.
Default:1e-08
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:The relative nonlinear residual drop to shoot for during fixed point iterations. This check is performed based on the main app's nonlinear residual.
- relaxation_factor1Fraction of newly computed value to keep.Set between 0 and 2.
Default:1
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Fraction of newly computed value to keep.Set between 0 and 2.
- transformed_postprocessorsList of main app postprocessors to transform during fixed point iterations
C++ Type:std::vector<PostprocessorName>
Unit:(no unit assumed)
Controllable:No
Description:List of main app postprocessors to transform during fixed point iterations
- transformed_variablesList of main app variables to transform during fixed point iterations
C++ Type:std::vector<std::string>
Unit:(no unit assumed)
Controllable:No
Description:List of main app variables to transform during fixed point iterations
Fixed Point Iterations Parameters
- automatic_scalingFalseWhether to use automatic scaling for the variables.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Whether to use automatic scaling for the variables.
- compute_scaling_onceTrueWhether the scaling factors should only be computed once at the beginning of the simulation through an extra Jacobian evaluation. If this is set to false, then the scaling factors will be computed during an extra Jacobian evaluation at the beginning of every time step.
Default:True
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Whether the scaling factors should only be computed once at the beginning of the simulation through an extra Jacobian evaluation. If this is set to false, then the scaling factors will be computed during an extra Jacobian evaluation at the beginning of every time step.
- ignore_variables_for_autoscalingList of variables that do not participate in autoscaling.
C++ Type:std::vector<std::string>
Unit:(no unit assumed)
Controllable:No
Description:List of variables that do not participate in autoscaling.
- off_diagonals_in_auto_scalingFalseWhether to consider off-diagonals when determining automatic scaling factors.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Whether to consider off-diagonals when determining automatic scaling factors.
- resid_vs_jac_scaling_param0A parameter that indicates the weighting of the residual vs the Jacobian in determining variable scaling parameters. A value of 1 indicates pure residual-based scaling. A value of 0 indicates pure Jacobian-based scaling
Default:0
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:A parameter that indicates the weighting of the residual vs the Jacobian in determining variable scaling parameters. A value of 1 indicates pure residual-based scaling. A value of 0 indicates pure Jacobian-based scaling
- scaling_group_variablesName of variables that are grouped together for determining scale factors. (Multiple groups can be provided, separated by semicolon)
C++ Type:std::vector<std::vector<std::string>>
Unit:(no unit assumed)
Controllable:No
Description:Name of variables that are grouped together for determining scale factors. (Multiple groups can be provided, separated by semicolon)
Solver Variable Scaling Parameters
- contact_line_search_allowed_lambda_cuts2The number of times lambda is allowed to be cut in half in the contact line search. We recommend this number be roughly bounded by 0 <= allowed_lambda_cuts <= 3
Default:2
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:The number of times lambda is allowed to be cut in half in the contact line search. We recommend this number be roughly bounded by 0 <= allowed_lambda_cuts <= 3
- contact_line_search_ltolThe linear relative tolerance to be used while the contact state is changing between non-linear iterations. We recommend that this tolerance be looser than the standard linear tolerance
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:The linear relative tolerance to be used while the contact state is changing between non-linear iterations. We recommend that this tolerance be looser than the standard linear tolerance
- line_searchdefaultSpecifies the line search type (Note: none = basic)
Default:default
C++ Type:MooseEnum
Unit:(no unit assumed)
Controllable:No
Description:Specifies the line search type (Note: none = basic)
- line_search_packagepetscThe solver package to use to conduct the line-search
Default:petsc
C++ Type:MooseEnum
Unit:(no unit assumed)
Controllable:No
Description:The solver package to use to conduct the line-search
Solver Line Search Parameters
- control_tagsAdds user-defined labels for accessing object parameters via control logic.
C++ Type:std::vector<std::string>
Unit:(no unit assumed)
Controllable:No
Description:Adds user-defined labels for accessing object parameters via control logic.
- enableTrueSet the enabled status of the MooseObject.
Default:True
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Set the enabled status of the MooseObject.
- outputsVector of output names where you would like to restrict the output of variables(s) associated with this object
C++ Type:std::vector<OutputName>
Unit:(no unit assumed)
Controllable:No
Description:Vector of output names where you would like to restrict the output of variables(s) associated with this object
- skip_exception_checkFalseSpecifies whether or not to skip exception check
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Specifies whether or not to skip exception check
Advanced Parameters
- l_abs_tol1e-50Linear Absolute Tolerance
Default:1e-50
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Linear Absolute Tolerance
- l_max_its10000Max Linear Iterations
Default:10000
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Max Linear Iterations
- l_tol1e-05Linear Relative Tolerance
Default:1e-05
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Linear Relative Tolerance
- reuse_preconditionerFalseIf true reuse the previously calculated preconditioner for the linearized system across multiple solves spanning nonlinear iterations and time steps. The preconditioner resets as controlled by reuse_preconditioner_max_linear_its
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:If true reuse the previously calculated preconditioner for the linearized system across multiple solves spanning nonlinear iterations and time steps. The preconditioner resets as controlled by reuse_preconditioner_max_linear_its
- reuse_preconditioner_max_linear_its25Reuse the previously calculated preconditioner for the linear system until the number of linear iterations exceeds this number
Default:25
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Reuse the previously calculated preconditioner for the linear system until the number of linear iterations exceeds this number
Linear Solver Parameters
- max_xfem_update4294967295Maximum number of times to update XFEM crack topology in a step due to evolving cracks
Default:4294967295
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Maximum number of times to update XFEM crack topology in a step due to evolving cracks
- update_xfem_at_timestep_beginFalseShould XFEM update the mesh at the beginning of the timestep
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Should XFEM update the mesh at the beginning of the timestep
Xfem Fixed Point Iterations Parameters
- mffd_typewpSpecifies the finite differencing type for Jacobian-free solve types. Note that the default is wp (for Walker and Pernice).
Default:wp
C++ Type:MooseEnum
Unit:(no unit assumed)
Controllable:No
Description:Specifies the finite differencing type for Jacobian-free solve types. Note that the default is wp (for Walker and Pernice).
- petsc_optionsSingleton PETSc options
C++ Type:MultiMooseEnum
Unit:(no unit assumed)
Controllable:No
Description:Singleton PETSc options
- petsc_options_inameNames of PETSc name/value pairs
C++ Type:MultiMooseEnum
Unit:(no unit assumed)
Controllable:No
Description:Names of PETSc name/value pairs
- petsc_options_valueValues of PETSc name/value pairs (must correspond with "petsc_options_iname"
C++ Type:std::vector<std::string>
Unit:(no unit assumed)
Controllable:No
Description:Values of PETSc name/value pairs (must correspond with "petsc_options_iname"
Petsc Parameters
- n_max_nonlinear_pingpong100The maximum number of times the nonlinear residual can ping pong before requesting halting the current evaluation and requesting timestep cut
Default:100
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:The maximum number of times the nonlinear residual can ping pong before requesting halting the current evaluation and requesting timestep cut
- nl_abs_div_tol1e+50Nonlinear Absolute Divergence Tolerance. A negative value disables this check.
Default:1e+50
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Nonlinear Absolute Divergence Tolerance. A negative value disables this check.
- nl_abs_step_tol0Nonlinear Absolute step Tolerance
Default:0
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Nonlinear Absolute step Tolerance
- nl_abs_tol1e-50Nonlinear Absolute Tolerance
Default:1e-50
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Nonlinear Absolute Tolerance
- nl_div_tol1e+10Nonlinear Relative Divergence Tolerance. A negative value disables this check.
Default:1e+10
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Nonlinear Relative Divergence Tolerance. A negative value disables this check.
- nl_forced_its0The Number of Forced Nonlinear Iterations
Default:0
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:The Number of Forced Nonlinear Iterations
- nl_max_funcs10000Max Nonlinear solver function evaluations
Default:10000
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Max Nonlinear solver function evaluations
- nl_max_its50Max Nonlinear Iterations
Default:50
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:Max Nonlinear Iterations
- nl_rel_step_tol0Nonlinear Relative step Tolerance
Default:0
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Nonlinear Relative step Tolerance
- nl_rel_tol1e-08Nonlinear Relative Tolerance
Default:1e-08
C++ Type:double
Unit:(no unit assumed)
Controllable:No
Description:Nonlinear Relative Tolerance
- num_grids1The number of grids to use for a grid sequencing algorithm. This includes the final grid, so num_grids = 1 indicates just one solve in a time-step
Default:1
C++ Type:unsigned int
Unit:(no unit assumed)
Controllable:No
Description:The number of grids to use for a grid sequencing algorithm. This includes the final grid, so num_grids = 1 indicates just one solve in a time-step
- residual_and_jacobian_togetherFalseWhether to compute the residual and Jacobian together.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Whether to compute the residual and Jacobian together.
- snesmf_reuse_baseTrueSpecifies whether or not to reuse the base vector for matrix-free calculation
Default:True
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Specifies whether or not to reuse the base vector for matrix-free calculation
- splittingTop-level splitting defining a hierarchical decomposition into subsystems to help the solver.
C++ Type:std::vector<std::string>
Unit:(no unit assumed)
Controllable:No
Description:Top-level splitting defining a hierarchical decomposition into subsystems to help the solver.
- use_pre_SMO_residualFalseCompute the pre-SMO residual norm and use it in the relative convergence check. The pre-SMO residual is computed at the begining of the time step before solution-modifying objects are executed. Solution-modifying objects include preset BCs, constraints, predictors, etc.
Default:False
C++ Type:bool
Unit:(no unit assumed)
Controllable:No
Description:Compute the pre-SMO residual norm and use it in the relative convergence check. The pre-SMO residual is computed at the begining of the time step before solution-modifying objects are executed. Solution-modifying objects include preset BCs, constraints, predictors, etc.
Nonlinear Solver Parameters
Restart Parameters
Input Files
- (modules/optimization/test/tests/outputs/exodus_optimization_steady/forward_and_adjoint_iteration_output.i)
- (modules/optimization/test/tests/executioners/steady_and_adjoint/multi_variable.i)
- (modules/optimization/test/tests/executioners/steady_and_adjoint/self_adjoint.i)
- (modules/combined/test/tests/optimization/invOpt_mechanics/forward_and_adjoint.i)
- (modules/optimization/test/tests/misc/scaling_test/scaling_test.i)
- (modules/optimization/test/tests/executioners/steady_and_adjoint/nonhomogeneous_bc.i)
- (modules/optimization/test/tests/executioners/steady_and_adjoint/nonlinear_diffusion.i)
- (modules/optimization/test/tests/optimizationreporter/bimaterial/model_and_adjoint.i)
- (modules/optimization/test/tests/optimizationreporter/general_opt/point_loads_gen_opt/forward_and_adjoint.i)
- (modules/optimization/test/tests/optimizationreporter/mesh_source/forward_and_adjoint.i)
- (modules/optimization/test/tests/executioners/steady_and_adjoint/array_variable.i)
- (modules/optimization/test/tests/optimizationreporter/general_opt/point_loads_gen_opt/forward_and_adjoint_transfer_data.i)
- (modules/optimization/test/tests/executioners/constrained/inequality/forward_and_adjoint.i)
- (modules/optimization/test/tests/optimizationreporter/nonlinear_material/forward_and_adjoint.i)
- (modules/optimization/examples/materialFrequency/wave1D/model_grad.i)
- (modules/optimization/test/tests/optimizationreporter/point_loads/forward_and_adjoint.i)
- (modules/optimization/test/tests/optimizationreporter/bc_load_linearFunction/forward_and_adjoint.i)