Dissertation

This repository contains slightly modified copies of several R scripts I wrote between 2017 and 2021, which appear in my dissertation. This is a long term work-in-progress, which morphed into a CRAN package:

https://www.github.com/willzywiec/criticality

These scripts build a deep neural network metamodel of Monte Carlo radiation transport simulations, which are subsequently used to estimate process criticality accident risk [P(keff) $\geq$ 1 🤯]. The default configuration of these scripts generate approximately 85 GB of data, which are saved to the test folder. The hardware and software I used are provided below.

GWU dissertations don't allow an epigraph. If they did, I would've selected this one:

"Risk comes from not knowing what you're doing." - Warren Buffett

I initially started out with 16 GB (2 x 8 GB) DDR4 3000 MHz RAM and ended up running into issues loading some of the larger datasets. Upgrading to 32 GB solved these issues.

Deep Learning Desktop Computer

AMD Ryzen 7 1700 3.0 GHz 8-core CPU
Asus ROG Strix B350-i Motherboard
Corsair Vengeance LPX 32 GB (2 x 16 GB) DDR4 3200 MHz RAM
DAN A4-SFGX v2 Mini ITX Case
Noctua NH-L9a + AM4 Kit Cooler (I also bought a Cryorig C7, but it sat a bit higher than I liked)
Noctua NT-H1 Thermal Compound
Nvidia GeForce GTX 1080 8 GB GDDR5X Founders Edition GPU
Samsung 960 EVO 500 GB M.2-2280 SSD

Prerequisites

Python 3.7+
TensorFlow 2.1+
CUDA 10+
R 4+
Rtools
MCNP6.2 (optional)
ENDF/B-VII.1 nuclear data (optional)
32+ GB RAM

Running the Code

install.R installs all necessary R packages.

MCNP input decks were built using build/grid.R and run using linux/copy.py and linux/volley.py, which are configured to run on Red Hat Linux using Slurm. The data I used in my dissertation consisted of a 130 MB mcnp-output.csv file, which contains output from 1,542,397 MCNP simulations that were run on the Quartz supercomputer at LLNL. I haven't posted the .csv file or simulation output, but an example of the required .csv file format is shown below. With the exception of the MCNP simulations, everything was run on a desktop computer I built for this project.

dist/dist.R needs to be run for the truncated probability distribution fits called in source/notebook.Rmd (e.g., 'gamma'). Once that's done, everything else can be run from source/notebook.Rmd.

All scripts were tested on 4/24/2021 using R 4.0.2, Python 3.8.5, TensorFlow 2.4, and CUDA 11.3.

R Packages Needed

BiocManager
BiocManager::install('graph')
BiocManager::install('RBGL')
BiocManager::install('Rgraphviz')
bnlearn
caret
dplyr
EnvStats
evd
fitdistrplus
forcats (optional)
ggplot2
gRain
igraph
keras
knitr
magrittr
metR (optional)
parallel
ParetoPosStable (optional)
reshape (optional)
reticulate
rmarkdown
scales
snow
xfun

Closing Remarks

I recently compiled an updated PDF of my dissertation using MiKTeX and the latexmk dissertation.tex command. I didn't include the LaTeX source code here, but if I forget how it's done or someone else stumbles upon this thread, it's generally easier to install LaTeXTools and compile PDFs directly in Sublime Text using package control.