SUMMARY

The development and spread of new nuclear technologies create a demand for new methods of safeguarding the fuel to prevent diversion for proliferation of nuclear weapons. One proposed method of controlled monitoring is the use of antineutrino detectors, which can be used to make assessments of isotopic inventory inside of the reactor based on the counts it detects externally. When continuous monitoring with such a detector is not available, the burnup of the reactor can be determined by placing the detector near the reactor, taking a measurement of antineutrino count rates across the energy spectrum, and then making burnup predictions from that given information. This thesis is a survey of machine learning methods that can be used to make such a prediction. The performance of each are compared to each other along with various feature engineering methods and hyperparameter selection to determine which model would be best for application in the field. Based on the studies performed in this survey, a simple ordinary least-squares polynomial regression of degree 3 and standard scaling is the fastest and most accurate method for predicting reactor burnup from antineutrino yield spectra when trained on similar labeled data. These results demonstrate that for small data set applications of this method, simple methods can outperform complex machine learning methods and should not be neglected in favor of something more complex.