openmc_UQ

Workflow for running nuclear data Monte Carlo UQ with OpenMC. The workflow wraps the SANDY and OpenMC in UncertaintyQuantification.jl. The workflow generates nuclear data samples and runs openmc in the same slurm submission. Random seeds for SANDY are interpolated into an input file by UncertaintyQuantification.jl.

Features:

• OpenMC configuration agnostic (can use DAGMC + libmesh)
• Convenient HPC resource manager (can use multiple nodes for OpenMC simulations). Enabled through UQ.jl HPC simulation manager
• Sample multiple nuclides simultaneously (from available ENDF covariance files)
• Exposure to more advanced sampling algorithms from UncertaintyQuantification.jl
• Dimension reduction of Nuclear Data covariance matrices using SVDs

Contributions and issues are welcome.

Installation

1. Install OpenMC
2. Install NJOY
3. Install SANDY
   • pip install sandy
   • Or from source
4. Install Julia. Recommended to use the Juliaup installation manager
5. Install UncertaintyQuantification.jl and other julia packages
   • julia -e 'using Pkg; Pkg.add("UncertaintyQuantification")'
   • julia -e 'using Pkg; Pkg.add("DelimitedFiles")'
   • julia -e 'using Pkg; Pkg.add("HDF5")'
   • julia -e 'using Pkg; Pkg.add("StatsBase")'
   • julia -e 'using Pkg; Pkg.add("Plots")'
6. Download nuclear data (ENDF and HDF5)
7. Install openmc_uq python package:
   • pip install .

To run a basic Monte Carlo OpenMC wrap of SANDY

See example/simple_tokamak for example

1. Configure the uq_inputs.json input script

2. Run it! julia MonteCarlo.jl uq_inputs.json

• Or submit julia MonteCarlo.jl uq_inputs.json in a slurm script with a few cpus

Things to consider while configuring the input file:

• Modify solver_exe
• Modify your endf_dir
• Modify your openmc cross_sections.xml (ideally the same library as whatever ENDF library you've specified)
• Modify the openmc_xml_dir, should point to the directory containing your OpenMC input files
• Modify the HPC credentials
  • You can optionally change the throttle (max samples run simultaneous)
  • You can optionally change ntasks (5 cores gives a simulation time ~2 mins per simulation)
  • You can optionally change nodes for larger simulations
  • You can optionally set a batchsize (max samples submitted simultaneous) see docs
  • You can optionally change the number of samples. Currently set to 100 samples
• Modify the command list, to what needs to be loaded to run your OpenMC model
• Modify the work_dir, to the directory where you would like simulations to be run

This should produce the similar figures to example/simple_tokamak/figures. If you're not interested in computing a failure probability, the limitstate function can be ignored. You make also simulate samples without using probability_of_failure function. This can be done by

samples = sample(inputs)
evaluate!(ext, samples)

Where inputs and ext are the inputs and external model as defined in MonteCarlo.jl.

To run a more advanced sampling strategy (variance reduction)

See example/simple_tokamak_advanced for example

1. Configure the uq_inputs.json input script

2. Setup the simulation python3 setup.py uq_inputs.json

3. Run it! julia LatinHypercube.jl uq_inputs.json

• Or submit julia LatinHypercube.jl uq_inputs.json in a slurm script with a few cpus

Other variance reduction, such as Subset Simulation are available, for evaluating design reliability, such as P(TBR < 1.05).

What does setup.py do?

It performs an SVD of all the covariances available within your specified ENDF library, stores the matrices, and computes the number of components to capture 99% of the variance of each nuclide.

For most nuclides it is quite expensive to compute SVDs, so parallelism is also provided. In a slurm script:

SLURM_NTASKS=10 #(e.g.)
python3 setup.py uq_inputs.json

Note: currently only MF33 (cross section) covariances are captured by this.