| A fuel load optimization process for the ATR was developed using machine learning. Cycle data was collected from engineering documents for ATR operating cycles 46A through 169A. Cycles 165A through 169A were then held back to be used to test the machine learning process. The total of the training/testing dataset was 10,400 inputs over 260 ATR cycles. KNN-imputation was used in the instance that a missing value was present in the dataset. Once the data was fully collected, exploratory data analysis was completed to understand any trends in the dataset. Three regression algorithms were considered for the fuel load optimization: linear regression, random forest regression, and neural networks. Linear regression performed the worst overall and could not account for fuel element position in the dataset causing all models to be underfit. Random forest performed best in terms of R2 value but contained severe spikes when incorrect. Neural networks were found to be the best fit due to predicting the closest to the as run burnup data. Both the feature selection and train-test split values were carefully considered with the best results coming from a 75%/25% train/test split with the important features being fuel element position, cycle MWd, total core power, and cycle length. Predicting fuel element burnup performed the best of all the machine learning algorithms and k-nearest neighbors was used to get a corresponding initial 235U loading. The predicted values were then scaled based on errors within the dataset. Ultimately, six of the seven cycles tested were able to use fewer fresh fuel elements and the cycle that used more fresh fuel elements was the result of a mismatch between the desired cycle and reality. Ultimately, a total of 108 fresh fuel elements was used over seven cycles compared to 124 fresh fuel elements used in the corresponding as runs, with a 13% decrease in fresh fuel element usage. |