Local adjoint source for Random Ray

docs/source/methods/random_ray.rst

@@ -1081,28 +1081,32 @@ lifetimes.
 
 In OpenMC, the random ray adjoint solver is implemented simply by transposing
 the scattering matrix, swapping :math:`\nu\Sigma_f` and :math:`\chi`, and then
-running a normal transport solve. When no external fixed source is present, no
-additional changes are needed in the transport process. However, if an external
-fixed forward source is present in the simulation problem, then an additional
-step is taken to compute the accompanying fixed adjoint source. In OpenMC, the
-adjoint flux does *not* represent a response function for a particular detector
-region. Rather, the adjoint flux is the global response, making it appropriate
-for use with weight window generation schemes for global variance reduction.
-Thus, if using a fixed source, the external source for the adjoint mode is
-simply computed as being :math:`1 / \phi`, where :math:`\phi` is the forward
-scalar flux that results from a normal forward solve (which OpenMC will run
-first automatically when in adjoint mode). The adjoint external source will be
-computed for each source region in the simulation mesh, independent of any
-tallies. The adjoint external source is always flat, even when a linear
-scattering and fission source shape is used. When in adjoint mode, all reported
-results (e.g., tallies, eigenvalues, etc.) are derived from the adjoint flux,
-even when the physical meaning is not necessarily obvious. These values are
-still reported, though we emphasize that the primary use case for adjoint mode
-is for producing adjoint flux tallies to support subsequent perturbation studies
-and weight window generation.
-
-Note that the adjoint :math:`k_{eff}` is statistically the same as the forward
-:math:`k_{eff}`, despite the flux distributions taking different shapes.
+running a normal transport solve. When no external fixed forward source is 
+present, or if an adjoint fixed source is specifically provided, no additional 
+changes are needed in the transport process. This adjoint source can 
+correspond, for example, to a detector response function in a particular 
+region. However, if an external fixed forward source is present in the 
+simulation problem without an adjoint fixed source, an additional step is taken 
+to compute the accompanying forward-weighted adjoint source. In this case, the
+adjoint flux does *not* represent the importance of locations in phase space to 
+detector response; rather, the "response" in question is a uniform distribution 
+of Monte Carlo particle density, making the importance provided by the adjoint 
+flux appropriate for use with weight window generation schemes for global 
+variance reduction. Thus, if using a fixed source, the forward-weighted 
+external source for adjoint mode is simply computed as being :math:`1 / \phi`, 
+where :math:`\phi` is the forward scalar flux that results from a normal 
+forward solve (which OpenMC will run first automatically when in adjoint mode). 
+The adjoint external source will be computed for each source region in the 
+simulation mesh, independent of any tallies. The adjoint external source is 
+always flat, even when a linear scattering and fission source shape is used. 
+
+When in adjoint mode, all reported results (e.g., tallies, eigenvalues, etc.) 
+are derived from the adjoint flux, even when the physical meaning is not 
+necessarily obvious. These values are still reported, though we emphasize that 
+the primary use case for adjoint mode is for producing adjoint flux tallies to 
+support subsequent perturbation studies and weight window generation. Note 
+however that the adjoint :math:`k_{eff}` is statistically the same as the 
+forward :math:`k_{eff}`, despite the flux distributions taking different shapes.
 
 ---------------------------
 Fundamental Sources of Bias

docs/source/methods/variance_reduction.rst

@@ -82,8 +82,8 @@ where it was born from.
 
 The Forward-Weighted Consistent Adjoint Driven Importance Sampling method, or
 `FW-CADIS method <https://doi.org/10.13182/NSE12-33>`_, produces weight windows
-for global variance reduction given adjoint flux information throughout the
-entire domain. The weight window lower bound is defined in Equation
+for global or local variance reduction given adjoint flux information throughout 
+the entire domain. The weight window lower bound is defined in Equation
 :eq:`fw_cadis`, and also involves a normalization step not shown here.
 
 .. math::
@@ -135,6 +135,18 @@ aware of this.
 
     \text{FOM} = \frac{1}{\text{Time} \times \sigma^2}
 
+Finally, one unique capability of the FW-CADIS weight window generator is to 
+produce weight windows for local variance reduction, given a list of the 
+responses of interest. This is controlled by optionally specifying target 
+tallies from the :class:`openmc.model.Model` to the 
+:class:`openmc.WeightWindowGenerator`, as illustrated in the
+:ref:`user guide<variance_reduction>`. If target tallies for local variance 
+reduction are supplied, then the adjoint sources are only populated after the 
+initial forward simulation in the source regions associated with those tallies. 
+In other regions, the adjoint source term is instead set to zero. The Random  
+Ray solver then determines the adjoint flux map used to generate FW-CADIS 
+weight windows following the usual technique.
+
 .. _methods_source_biasing:
 
 --------------

docs/source/usersguide/random_ray.rst

@@ -944,6 +944,8 @@ as::
 which will greatly improve the quality of the linear source term in 2D
 simulations.
 
+.. _usersguide_random_ray_run_modes:
+
 ---------------------------------
 Fixed Source and Eigenvalue Modes
 ---------------------------------
@@ -1073,22 +1075,47 @@ The adjoint flux random ray solver mode can be enabled as::
 
     settings.random_ray['adjoint'] = True
 
-When enabled, OpenMC will first run a forward transport simulation followed by
-an adjoint transport simulation. The purpose of the forward solve is to compute
-the adjoint external source when an external source is present in the
-simulation. Simulation settings (e.g., number of rays, batches, etc.) will be
-identical for both simulations. At the conclusion of the run, all results (e.g.,
-tallies, plots, etc.) will be derived from the adjoint flux rather than the
-forward flux but are not labeled any differently. The initial forward flux
-solution will not be stored or available in the final statepoint file. Those
-wishing to do analysis requiring both the forward and adjoint solutions will
-need to run two separate simulations and load both statepoint files.
+When enabled, OpenMC will first run a forward transport simulation if there are 
+no user-specified adjoint sources present, followed by an adjoint transport 
+simulation. Fixed adjoint sources can be specified on the 
+:attr:`openmc.Settings.random_ray` dictionary as follows::
+
+    # Geometry definition
+    ...
+    detector_cell = openmc.Cell(fill=detector_mat, name='cell where detector will be')
+    ...
+    # Define fixed adjoint neutron source
+    strengths = [1.0]
+    midpoints = [1.0e-4]
+    energy_distribution = openmc.stats.Discrete(x=midpoints, p=strengths)
+
+    adj_source = openmc.IndependentSource(
+        energy=energy_distribution, 
+        constraints={'domains': [detector_cell]}
+    )
+
+    # Add to random_ray dict
+    settings.random_ray['adjoint_source'] = adj_source
+
+The same constraints apply to the user-defined adjoint source as to the forward 
+source, described in the :ref:`Fixed Source and Eigenvalue section 
+<usersguide_random_ray_run_modes>`. If this source is not provided, a forward 
+solve must take place to compute the adjoint external source when a forward 
+external source is present in the problem. Simulation settings (e.g., number of 
+rays, batches, etc.) will be identical for both calculations. At the 
+conclusion of the run, all results (e.g., tallies, plots, etc.) will be 
+derived from the adjoint flux rather than the forward flux but are not labeled 
+any differently. The initial forward flux solution will not be stored or 
+available in the final statepoint file. Those wishing to do analysis requiring 
+both the forward and adjoint solutions will need to run two separate 
+simulations and load both statepoint files.
 
 .. note::
-    When adjoint mode is selected, OpenMC will always perform a full forward
-    solve and then run a full adjoint solve immediately afterwards. Statepoint
-    and tally results will be derived from the adjoint flux, but will not be
-    labeled any differently.
+    Use of the automated 
+    :ref:`FW-CADIS weight window generator<usersguide_fw_cadis>` is not 
+    currently compatible with user-defined adjoint sources. Instead, the 
+    initial forward calculation is used to assign "forward-weighted" adjoint 
+    sources to the tally regions of interest.
 
 ---------------------------------------
 Putting it All Together: Example Inputs

docs/source/usersguide/variance_reduction.rst

@@ -4,26 +4,27 @@
 Variance Reduction
 ==================
 
-Global variance reduction in OpenMC is accomplished by weight windowing
-or source biasing techniques, the latter of which additionally provides a
-local variance reduction capability. OpenMC is capable of generating weight
-windows using either the MAGIC or FW-CADIS methods. Both techniques will
-produce a ``weight_windows.h5`` file that can be loaded and used later on. In
+Global and local variance reduction are possible in OpenMC through both weight 
+windowing and source biasing techniques. OpenMC is capable of generating weight
+windows using either the MAGIC or FW-CADIS methods, the latter with an optional
+capability for local variance reduction. Both techniques will produce a
+``weight_windows.h5`` file that can be loaded and used later on. In
 this section, we first break down the steps required to generate and apply
 weight windows, then describe how source biasing may be applied.
 
 .. _ww_generator:
 
-------------------------------------
-Generating Weight Windows with MAGIC
-------------------------------------
+-------------------------------------------
+Generating Global Weight Windows with MAGIC
+-------------------------------------------
 
 As discussed in the :ref:`methods section <methods_variance_reduction>`, MAGIC
 is an iterative method that uses flux tally information from a Monte Carlo
-simulation to produce weight windows for a user-defined mesh. While generating
-the weight windows, OpenMC is capable of applying the weight windows generated
-from a previous batch while processing the next batch, allowing for progressive
-improvement in the weight window quality across iterations.
+simulation to produce weight windows for a user-defined mesh with the objective 
+of global variance reduction. While generating the weight windows, OpenMC is 
+capable of applying the weight windows generated from a previous batch while 
+processing the next batch, allowing for progressive improvement in the weight 
+window quality across iterations.
 
 The typical way of generating weight windows is to define a mesh and then add an
 :class:`openmc.WeightWindowGenerator` object to an :attr:`openmc.Settings`
@@ -71,15 +72,20 @@ At the end of the simulation, a ``weight_windows.h5`` file will be saved to disk
 for later use. Loading it in another subsequent simulation will be discussed in
 the "Using Weight Windows" section below.
 
-------------------------------------------------------
-Generating Weight Windows with FW-CADIS and Random Ray
-------------------------------------------------------
+.. _usersguide_fw_cadis:
+
+----------------------------------------------------------------------
+Generating Global or Local Weight Windows with FW-CADIS and Random Ray
+----------------------------------------------------------------------
 
 Weight window generation with FW-CADIS and random ray in OpenMC uses the same
-exact strategy as with MAGIC. An :class:`openmc.WeightWindowGenerator` object is
-added to the :attr:`openmc.Settings` object, and a ``weight_windows.h5`` will be
-generated at the end of the simulation. The only difference is that the code
-must be run in random ray mode. A full description of how to enable and setup
+exact strategy as with MAGIC. Using FW-CADIS, however, also enables 
+local variance reduction in fixed source problems through the :attr:`targets` 
+attribute, which is described later in this section. To enable FW-CADIS, an 
+:class:`openmc.WeightWindowGenerator` object is added to the 
+:attr:`openmc.Settings` object, and a ``weight_windows.h5`` will be generated 
+at the end of the simulation. The only procedural difference is that the code 
+must be run in random ray mode. A full description of how to enable and setup 
 random ray mode can be found in the :ref:`Random Ray User Guide <random_ray>`.
 
 .. note::
@@ -90,7 +96,7 @@ random ray mode can be found in the :ref:`Random Ray User Guide <random_ray>`.
     ray solver. A high level overview of the current workflow for generation of
     weight windows with FW-CADIS using random ray is given below.
 
-1. Begin by making a deepy copy of your continuous energy Python model and then
+1. Begin by making a deep copy of your continuous energy Python model and then
    convert the copy to be multigroup and use the random ray transport solver.
    The conversion process can largely be automated as described in more detail
    in the :ref:`random ray quick start guide <quick_start>`, summarized below::
@@ -148,7 +154,53 @@ random ray mode can be found in the :ref:`Random Ray User Guide <random_ray>`.
     assigning to ``model.settings.random_ray['source_region_meshes']``) and for
     weight window generation.
 
-3. When running your multigroup random ray input deck, OpenMC will automatically
+3. (Optional) If local variance reduction is desired in a fixed-source problem,  
+   populate the :attr:`targets` attribute with an :class:`openmc.Tallies` 
+   instance or an iterable of tally IDs indicating the tallies of interest for 
+   variance reduction::
+
+    # Build a new example and WWG for local variance reduction
+    from openmc.examples import random_ray_three_region_cube_with_detectors
+    new_model = random_ray_three_region_cube_with_detectors()
+
+    ww_mesh = openmc.RegularMesh()
+    n = 7
+    width = 35.0
+    ww_mesh.dimension = (n, n, n)
+    ww_mesh.lower_left = (0.0, 0.0, 0.0)
+    ww_mesh.upper_right = (width, width, width)
+
+    wwg = openmc.WeightWindowGenerator(
+        method="fw_cadis", 
+        mesh=ww_mesh, 
+        max_realizations=new_model.settings.batches
+    )
+    new_model.settings.weight_window_generators = wwg
+    new_model.settings.random_ray['volume_estimator'] = 'naive'
+
+    # Get the tallies of interest
+    target_tallies = openmc.Tallies()
+
+    for tally in list(new_model.tallies):
+        if tally.name in {"Detector 1 Tally", "Detector 2 Tally"}:
+            target_tallies.append(tally)
+
+    # Add to WeightWindowGenerator
+    wwg.targets = target_tallies
+
+.. warning::
+    The tallies designated as FW-CADIS targets to the 
+    :class:`~openmc.WeightWindowGenerator` must be present under the
+    :class:`~openmc.model.Model.tallies` attribute of the 
+    :class:`~openmc.model.Model` as well in order to be recognized as valid 
+    local variance reduction targets. This check is performed when the 
+    :func:`openmc.model.Model.export_to_model_xml` or 
+    :func:`openmc.model.Model.export_to_xml` functions are called, meaning 
+    that the standalone :func:`openmc.Settings.export_to_xml` and 
+    :func:`openmc.Tallies.export_to_xml` methods should not be used with 
+    FW-CADIS local variance reduction.
+
+4. When running your multigroup random ray input deck, OpenMC will automatically
    run a forward solve followed by an adjoint solve, with a
    ``weight_windows.h5`` file generated at the end. The ``weight_windows.h5``
    file will contain FW-CADIS generated weight windows. This file can be used in

include/openmc/random_ray/flat_source_domain.h

@@ -40,9 +40,10 @@ class FlatSourceDomain {
   void random_ray_tally();
   virtual void accumulate_iteration_flux();
   void output_to_vtk() const;
-  void convert_external_sources();
+  void convert_external_sources(bool use_adjoint_sources);
   void count_external_source_regions();
-  void set_adjoint_sources();
+  void set_fw_adjoint_sources();
+  void set_local_adjoint_sources();
   void flux_swap();
   virtual double evaluate_flux_at_point(Position r, int64_t sr, int g) const;
   double compute_fixed_source_normalization_factor() const;
@@ -76,6 +77,7 @@ class FlatSourceDomain {
   // Static Data members
   static bool volume_normalized_flux_tallies_;
   static bool adjoint_; // If the user wants outputs based on the adjoint flux
+  static bool fw_cadis_local_;
   static double
     diagonal_stabilization_rho_; // Adjusts strength of diagonal stabilization
                                  // for transport corrected MGXS data
@@ -84,6 +86,8 @@ class FlatSourceDomain {
   static std::unordered_map<int, vector<std::pair<Source::DomainType, int>>>
     mesh_domain_map_;
 
+  static std::vector<size_t> fw_cadis_local_targets_;
+
   //----------------------------------------------------------------------------
   // Static data members
   static RandomRayVolumeEstimator volume_estimator_;

include/openmc/random_ray/random_ray_simulation.h

@@ -20,8 +20,9 @@ class RandomRaySimulation {
   //----------------------------------------------------------------------------
   // Methods
   void apply_fixed_sources_and_mesh_domains();
-  void prepare_fixed_sources_adjoint();
-  void prepare_adjoint_simulation();
+  void prepare_fw_fixed_sources_adjoint();
+  void prepare_local_fixed_sources_adjoint();
+  void prepare_adjoint_simulation(bool fw_adjoint);
   void simulate();
   void output_simulation_results() const;
   void instability_check(
@@ -34,15 +35,9 @@ class RandomRaySimulation {
   // Accessors
   FlatSourceDomain* domain() const { return domain_.get(); }
 
-  //----------------------------------------------------------------------------
-  // Public data members
-
-  // Flag for adjoint simulation;
-  bool adjoint_needed_;
-
 private:
   //----------------------------------------------------------------------------
-  // Private data members
+  // Data members
 
   // Contains all flat source region data
   unique_ptr<FlatSourceDomain> domain_;
@@ -57,17 +52,13 @@ class RandomRaySimulation {
   // Number of energy groups
   int negroups_;
 
-  // Toggle for first simulation
-  bool is_first_simulation_;
-
 }; // class RandomRaySimulation
 
 //============================================================================
 //! Non-member functions
 //============================================================================
 
 void validate_random_ray_inputs();
-void print_adjoint_header();
 void openmc_finalize_random_ray();
 
 } // namespace openmc

include/openmc/source.h

@@ -43,6 +43,7 @@ class Source;
 namespace model {
 
 extern vector<unique_ptr<Source>> external_sources;
+extern vector<unique_ptr<Source>> adjoint_sources;
 
 // Probability distribution for selecting external sources
 extern DiscreteIndex external_sources_probability;

include/openmc/weight_windows.h

@@ -223,6 +223,9 @@ class WeightWindowsGenerator {
   double threshold_ {1.0}; //<! Relative error threshold for values used to
                            // update weight windows
   double ratio_ {5.0};     //<! ratio of lower to upper weight window bounds
+
+  // Local FW-CADIS target tallies
+  std::vector<size_t> targets_;
 };
 
 //==============================================================================